A user game data analysis management method and system

Abstract

This application discloses a user game data analysis and management method and system, belonging to the fields of data analysis, big data, and big data technology. By performing online semantic parsing on player behavior data units, it dynamically obtains content-sensitive and privacy-sensitive attributes, and jointly maps them into desensitization depth parameters and processing priority labels. This enables differentiated desensitization and load-aware simplification for different data units, and dynamically adjusts the allocation of computing resources based on data semantics. This method effectively solves the problems of uniform privacy protection granularity, imprecise load control, and static task scheduling in traditional cross-source game data analysis. Thus, it optimizes the utilization of computing and communication resources while ensuring player privacy, and ensures the timeliness and accuracy of key game operation collaborative analysis tasks.

Description

Technical Field

[0001] This application relates to the fields of data analysis, big data, and big data technology, and in particular to a user game data analysis and management method and system. Background Technology

[0002] In cross-source game data analysis, collaborative analysis of multi-source player behavior data is required under privacy protection to produce conclusions such as churn prediction and anomaly detection. Related technologies typically configure privacy desensitization, load simplification, and task scheduling independently when processing such data: privacy protection uses a uniform desensitization granularity, performing feature extraction and truncation at the same depth on all data, without distinguishing between differences in data content importance and privacy sensitivity; edge load control adjusts the sampling rate only based on node resource occupancy, without considering the expected contribution of data to the analysis task as a basis for simplification decisions; and task scheduling relies on preset static priorities, failing to dynamically adjust computational resource allocation based on data semantics.

[0003] The aforementioned independent configuration results in critical analysis tasks being unable to guarantee timely processing under sudden loads due to low-value data consuming computing and communication resources. At the same time, the anonymization of highly privacy-sensitive data and low privacy-sensitive data at the same granularity leads to insufficient protection for the former or waste of communication bandwidth for the latter.

[0004] The relevant technologies do not provide a comprehensive solution for jointly making decisions on data semantic parsing, differential desensitization, load-aware simplification, and scheduling priority to solve the problem of multi-dimensional resource trade-offs in cross-source game data collaborative analysis. Summary of the Invention

[0005] This application provides a user game data analysis and management method and system, the technical solution of which is as follows: On the one hand, a method for analyzing and managing user game data is provided, the method comprising: Online semantic analysis is performed on the incoming player behavior data units to obtain content-sensitive attributes and privacy-sensitive attributes. The content-sensitive attributes characterize the expected impact of the player behavior data unit on the conclusion of the current collaborative analysis task, and the privacy-sensitive attributes characterize the degree to which the information carried by the player behavior data unit can identify an individual. The content-sensitive attribute and the privacy-sensitive attribute are jointly mapped to obtain the desensitization depth parameter and the processing priority label. The desensitization depth parameter determines the truncation layer position of the local feature extraction network, and the processing priority label distinguishes the computational resource contention weight. Based on the desensitization depth parameter, hierarchical feature extraction and truncation are performed in the local feature extraction network to obtain the desensitized intermediate representation. Based on the content sensitivity attribute and the real-time load status of the local computing node, load-aware simplification is performed on the player behavior data unit to obtain the retained data unit. The desensitized intermediate representation and the cross-source association identifier carried by the retained data unit are fused and analyzed to generate game operation collaborative analysis conclusions. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 This is a flowchart of a user game data analysis and management method provided in an embodiment of this application; Figure 2 This is a flowchart of another user game data analysis and management method provided in the embodiments of this application; Figure 3 This is a flowchart of another user game data analysis and management method provided in the embodiments of this application; Figure 4 This is a flowchart of another user game data analysis and management method provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of a user game data analysis and management system provided in an embodiment of this application. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0009] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0010] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0011] Before describing the technical solution of this application in detail, the system architecture applicable to this application is explained. The user game data analysis and management method provided in this application runs on a distributed computing architecture, which includes multiple local computing nodes and at least one collaborative computing node. The local computing node is deployed on the game user game data analysis and management system or edge computing device, and is responsible for collecting player behavior data units and performing online semantic parsing, local feature extraction, and load-aware simplification. The collaborative computing node is deployed on the central user game data analysis and management system or cloud platform, and is responsible for receiving the de-identified intermediate representation from each local computing node, performing forward computation of the remaining layer, cross-source feature aggregation, and classification reasoning, and generating game operation collaborative analysis conclusions. The local computing nodes and the collaborative computing node communicate with each other through a secure network channel, and the de-identified intermediate representation and the retained data units are transmitted through the secure network channel after encryption. The collaborative computing node also runs a feedback loop module, which is used to extract feedback signals from the game operation collaborative analysis conclusions and issue strategy update instructions to each local computing node to realize adaptive collaborative optimization of system strategy parameters. Under the above architecture, this application achieves differentiated privacy protection and intelligent resource scheduling for multi-source player behavior data through closed-loop collaboration of online semantic parsing, joint mapping, hierarchical feature extraction and load-aware simplification.

[0012] Traditional cross-source game data analysis methods typically configure privacy anonymization, load simplification, and task scheduling independently when processing multi-source player behavior data units. This results in a uniform anonymization granularity for privacy protection, failing to differentiate between the importance and privacy sensitivity of data content. Edge load control adjusts the sampling rate solely based on node resource occupancy, without considering the expected contribution of data to the analysis task. Task scheduling relies on preset static priorities, making it impossible to dynamically adjust computing resource allocation. These issues lead to unreliable processing time for critical analysis tasks under sudden loads, insufficient protection of highly privacy-sensitive data, and wasted communication bandwidth.

[0013] In response, this application proposes a user game data analysis and management method, see [link to relevant documentation]. Figure 1 The method includes: 101. Perform online semantic analysis on the incoming player behavior data units to obtain content-sensitive attributes and privacy-sensitive attributes. The content-sensitive attribute represents the expected impact of the player behavior data unit on the conclusion of the current collaborative analysis task, and the privacy-sensitive attribute represents the degree to which the information carried by the player behavior data unit can identify an individual.

[0014] 102. Jointly map the content-sensitive attribute and the privacy-sensitive attribute to obtain the desensitization depth parameter and the processing priority label. The desensitization depth parameter determines the cutoff layer position of the local feature extraction network, and the processing priority label distinguishes the computational resource contention weights.

[0015] 103. Based on the desensitization depth parameter, perform hierarchical feature extraction and truncation in the local feature extraction network to obtain the desensitized intermediate representation. Based on the content sensitivity attribute and the real-time load status of the local computing node, perform load-aware simplification on the player behavior data unit to obtain the retained data unit.

[0016] 104. Perform fusion analysis on the desensitized intermediate representation and the cross-source association identifier carried by the retained data unit to generate game operation collaborative analysis conclusions.

[0017] For ease of understanding, the following explains some key terms in this embodiment: A player behavior data unit refers to the smallest data record in a game system generated by player actions or system events, such as player clicks, purchases, chat logs, or task completion status.

[0018] Online semantic parsing refers to the process by which a system analyzes incoming player behavior data units in real time to understand their inherent meaning and contextual information.

[0019] Content sensitivity is a quantitative indicator used to characterize the expected impact of player behavior data units on the conclusions of the currently being performed collaborative analysis task.

[0020] Privacy sensitivity is a quantitative indicator used to characterize the degree to which the personal identity of a player's behavioral data unit is identifiable.

[0021] The desensitization depth parameter is a control parameter whose value determines at which level the local feature extraction network should truncate the feature extraction when processing player behavior data units.

[0022] A local feature extraction network (NMR) is a neural network model deployed on a local computing node. Its function is to extract meaningful features from raw player behavior data units. In a specific embodiment, the NMR adopts a deep convolutional neural network or Transformer encoder architecture, including an input embedding layer, multiple feature extraction layers, and an output projection layer, with a total number of layers denoted as L (e.g., L = 12 to 48 layers). The input embedding layer maps the raw player behavior data units (e.g., behavior sequences after one-hot encoding or word embedding) into d-dimensional feature vectors (e.g., d = 256 to 1024). Each feature extraction layer consists of stacked self-attention sub-layers and feedforward neural network sub-layers, with residual connections and layer normalization between layers. The output projection layer maps the output of a feature extraction layer to a feature space of a preset dimension. The truncation layer K ranges from 1 to L-1. When K = 1, truncation occurs in the shallowest layer (only passing through the input embedding layer), and when K = L-1, truncation occurs in the deepest feature extraction layer. After completing the forward computation of the first K layers on the local computing node, the intermediate activation features of the Kth layer are used as the desensitized intermediate representation, with a dimension of d.

[0023] Desensitized intermediate representation refers to the intermediate activation features output after player behavior data units are processed by the local feature extraction network to the truncated layer. This representation has been partially desensitized and no longer contains all the details of the original data.

[0024] Load-aware simplification refers to the process by which the system dynamically filters, downsamples, or discards data units based on the content sensitivity attributes of player behavior data units and the real-time load status of local computing nodes in order to optimize resource utilization.

[0025] Game operation collaborative analysis conclusions refer to the insights and predictions about the game's operational status, player behavior patterns, or potential risks generated by fusing and analyzing anonymized intermediate representations and retained data units from different sources.

[0026] In this embodiment, the user game data analysis and management method performs online semantic parsing on incoming player behavior data units to obtain content-sensitive attributes and privacy-sensitive attributes. Specifically, the acquisition of content-sensitive attributes can adopt preset static rules, such as classifying player behavior data units into different content sensitivity levels based on the event type. For example, all "payment" events can be set to high content sensitivity, while "login" events are set to medium content sensitivity. The acquisition of privacy-sensitive attributes can be based on the field types contained in the player behavior data unit. For example, if the data unit contains a "user nickname" field, its privacy-sensitive attribute is set to a higher value. If it only contains a "game duration" field, it is set to a lower value. It should be noted that the simplified implementation based on static rules described below is only used to help understand the basic principles of this application, and the scope of protection of this application is not limited thereto. In the preferred embodiment of this application, the determination of content-sensitive attributes and privacy-sensitive attributes adopts the dynamic multi-dimensional fusion method described in subsequent embodiments to achieve more refined data value assessment and privacy protection.

[0027] This method performs a joint mapping on the acquired content-sensitive attributes and privacy-sensitive attributes to obtain an anonymization depth parameter and a processing priority label. The anonymization depth parameter can be generated using a linear mapping mechanism, for example, by weighted summing of the values ​​of the content-sensitive and privacy-sensitive attributes, and then linearly mapping the sum to a preset cutoff layer depth range. The processing priority label can be determined based on a single threshold of the content-sensitive attribute; for example, when the content-sensitive attribute exceeds a certain fixed threshold, the processing priority label is set to high priority; otherwise, it is set to low priority.

[0028] Based on the desensitization depth parameter, hierarchical feature extraction and truncation are performed in the local feature extraction network to obtain a desensitized intermediate representation. For example, the local feature extraction network can be a multi-layer neural network, and the desensitization depth parameter directly specifies a certain layer in the network as the truncation layer. For example, if the desensitization depth parameter is K, the network performs forward computation on the first K layers and uses the output of the Kth layer as the desensitized intermediate representation. This method performs load-aware simplification of player behavior data units based on content sensitivity attributes and the real-time load status of the local computing node to obtain retained data units. For example, the local computing node can continuously monitor its CPU utilization. When the CPU utilization exceeds a preset threshold, the system will randomly discard or downsample player behavior data units with content sensitivity attributes below a certain fixed threshold based on the content sensitivity attributes of the player behavior data units.

[0029] This method fuses and analyzes the cross-source association identifiers carried by anonymized intermediate representations and retained data units to generate collaborative analysis conclusions for game operations. Specifically, the system can collect anonymized intermediate representations and retained data units from different local computing nodes and align them using cross-source association identifiers such as player IDs, match IDs, or timestamps. The aligned multi-source data is then input into a pre-defined collaborative analysis model, which performs feature concatenation and aggregation operations and performs classification or regression inference based on the aggregated features, thereby outputting collaborative analysis conclusions for game operations such as player churn prediction, abnormal behavior detection, or payment tendency rating.

[0030] The method provided in this embodiment dynamically obtains content-sensitive and privacy-sensitive attributes by performing online semantic parsing on player behavior data units, and jointly maps them into desensitization depth parameters and processing priority labels. This enables differentiated desensitization and load-aware simplification for different data units, and dynamically adjusts computing resource allocation based on data semantics. This method effectively solves the problems of uniform privacy protection granularity, imprecise load control, and static task scheduling in traditional cross-source game data analysis. It optimizes the utilization of computing and communication resources while protecting player privacy, ensuring the timeliness and accuracy of critical game operation collaborative analysis tasks.

[0031] In some of the embodiments described above in this application, online semantic parsing of incoming player behavior data units is proposed to obtain content-sensitive attributes. However, in the implementation process, how to comprehensively and accurately assess the expected impact of player behavior data units on the conclusions of collaborative analysis tasks and quantify it as content-sensitive attributes is a technical problem that requires meticulous handling. Simple parsing may not be able to fully capture the intrinsic value, timeliness, and contribution to specific behavioral patterns of the data units, thereby affecting the effectiveness of subsequent de-identification and resource allocation.

[0032] In response, this application further proposes a specific method for performing online semantic analysis on the incoming player behavior data units to obtain content-sensitive attributes, see [link to relevant documentation]. Figure 2 The method includes: 201. Determine the event rarity based on the frequency of occurrence of the player behavior data unit in the historical distribution of the same group.

[0033] The event rarity reflects the scarcity of the event type corresponding to the player behavior data unit. For example, event rarity quantifies how rare a specific player behavior event is within the historical context of similar player groups. A rare event may carry more information or indicate a unique situation. This can be achieved by maintaining a historical player behavior event log or database categorized by event type and player group. For an incoming player behavior data unit, its event type is identified, and then the frequency of that event type within a preset historical time window and within relevant player groups (e.g., players with similar demographics, playstyles, or game progress) is calculated. Event rarity can be negatively correlated with this frequency; that is, the lower the frequency, the higher the rarity.

[0034] 202. Based on the distilled simplified version of the master model corresponding to the current collaborative analysis task, perform forward inference on the player behavior data unit to estimate the degree of behavioral influence of the player behavior data unit on the conclusion of the current collaborative analysis task, and obtain the behavioral influence degree.

[0035] The Behavioral Impact measure measures the potential impact of a specific player behavior data unit on the current collaborative analysis task outcome or prediction. It is an estimate of how that data point will change the task conclusion. The distilled simplified version of the main model is a smaller, faster model trained to mimic the behavior of a large, complex main model, which can be deployed locally for rapid inference. When a player behavior data unit arrives, it is fed into this simplified model. The model outputs a score or probability indicating the relevance or influence of that data unit on the task objective (e.g., predicting player churn, identifying anomalous behavior); this score is the Behavioral Impact.

[0036] 203. Determine the real-time requirement component based on the upper limit of the latency of the current collaborative analysis task. The shorter the upper limit of the latency, the larger the real-time requirement component.

[0037] The real-time requirement component quantifies the urgency with which player behavior data units need to be processed for the current collaborative analytics task. Tasks with tighter deadlines require a higher real-time requirement component. Each collaborative analytics task (e.g., real-time fraud detection versus weekly player churn prediction) has an associated upper bound on latency, which defines the maximum acceptable latency for processing data related to that task. A shorter upper bound on latency indicates higher urgency. The real-time requirement component can be derived from this upper bound, for example, by using an inverse relation or mapping function, where a smaller upper bound produces a larger component value.

[0038] 204. Determine the degree of pattern completion of the player's behavior data unit in the detected behavior patterns to obtain the pattern completion contribution.

[0039] The pattern completion contribution indicates the degree to which a specific player behavior data unit contributes to completing or reinforcing a known or emerging behavioral pattern. Player data units that complete key parts of a pattern are more valuable. This involves maintaining a set of predefined or dynamically detected behavioral patterns (e.g., a series of actions leading to a purchase, a series of suspicious activities). When a player behavior data unit arrives, it is compared to these patterns. If the data unit represents a missing or next step in an incomplete pattern, its contribution to completing that pattern is evaluated. This contribution can be a score based on the importance of the completed step in the pattern or the overall progress of pattern completion.

[0040] 205. Multiplicatively modulate the impact of the behavior based on the rarity of the event, and then weight and fuse the modulated impact of the behavior, the real-time requirement component, and the contribution of the mode completion to obtain the content sensitivity attribute.

[0041] This step involves integrating the content-sensitive attributes of player behavior data units. It combines various aspects of the data unit's importance (rarity, impact, urgency, and pattern contribution) into a single, comprehensive metric. Behavioral impact is adjusted by multiplying the event's rarity, meaning that rare events with high behavioral impact are amplified. This modulated behavioral impact, the real-time requirement component, and the pattern completion contribution are then combined using a weighted sum. The weights assigned to each component can be predefined or dynamically adjusted based on the specific requirements of the collaborative analysis task. The result of this weighted fusion is the content-sensitive attribute, a scalar value representing the overall importance and urgency of the player behavior data unit.

[0042] The above technical solution enables a comprehensive and accurate assessment of the expected impact of player behavior data units on the conclusions of collaborative analysis tasks. By comprehensively considering multiple dimensions such as event rarity, behavioral impact, real-time requirement components, and pattern completion contribution, and employing a multiplicative modulation and weighted fusion mechanism, the resulting content sensitivity attribute becomes a highly discriminative indicator. This allows the system to prioritize data units that are not only highly influential but also rare, urgent, or crucial for pattern recognition. Therefore, subsequent processing such as determining the desensitization depth parameter and load-aware simplification (as described in the above implementation) can be executed more effectively, ensuring that valuable and urgent data are appropriately processed, while less important data can undergo deeper desensitization or even be discarded, thereby optimizing computational resource utilization and maintaining the accuracy of the analysis. Utilizing a distillation-simplified model for behavioral impact assessment ensures that this complex online semantic parsing process can be completed with low latency.

[0043] In some embodiments described above, this application proposes obtaining content-sensitive attributes by performing online semantic parsing on player behavior data units. The determination of these attributes relies on a weighted fusion of event rarity, behavioral impact, real-time requirement components, and pattern completion contribution. However, in practical applications, how to accurately and quantitatively determine event rarity and real-time requirement components, and ensure that these components effectively participate in the calculation of content-sensitive attributes to accurately reflect the value and timeliness of the data units, remains a technical problem that requires further refinement and resolution.

[0044] To address this, this application further proposes online semantic parsing of incoming player behavior data units to obtain content-sensitive attributes, including: determining event rarity based on the frequency of occurrence of the player behavior data unit in the historical distribution of the same group, where the event rarity reflects the scarcity of the event type corresponding to the player behavior data unit; performing forward inference on the player behavior data unit based on a distilled simplified version of the master model corresponding to the current collaborative analysis task to estimate the behavioral influence of the player behavior data unit on the conclusion of the current collaborative analysis task, obtaining the behavioral influence degree; determining the real-time requirement component based on the upper bound of the latency of the current collaborative analysis task, where a shorter upper bound results in a larger real-time requirement component; determining the pattern completion degree of the player behavior data unit in the detected behavioral patterns, obtaining the pattern completion contribution degree; and performing multiplicative modulation on the behavioral influence degree using the event rarity, and weighted fusing the modulated behavioral influence degree, the real-time requirement component, and the pattern completion contribution degree to obtain the content-sensitive attribute.

[0045] For example, determining the event rarity based on the frequency of occurrence of the player behavior data unit in the historical distribution of the same group, and determining the real-time requirement component based on the latency upper bound of the current collaborative analysis task, includes: counting the total number of times the event type corresponding to the player behavior data unit appears in the same group of players within a preset historical time window, determining the total number of active events of the same group of players within the preset historical time window, using the relative size of the total number of occurrences to the total number of active events as the occurrence frequency, and determining the event rarity in a way that is negatively monotonically correlated with the occurrence frequency. Based on the type identifier of the current collaborative analysis task, querying a preset latency upper bound mapping table, determining the latency upper bound of the current collaborative analysis task, determining the basic real-time metric in a way that is negatively monotonically correlated with the latency upper bound, normalizing the basic real-time metric to a preset numerical range, and obtaining the real-time requirement component. The event rarity and the real-time requirement component are respectively used as input values ​​for the corresponding dimensions in the weighted fusion of content sensitivity attributes, wherein the event rarity is also used as a modulation factor for the multiplicative modulation of the behavior influence.

[0046] When determining event rarity, it's necessary to count the total number of times the event type corresponding to a player behavior data unit appears within a preset historical time window among the same group of players. For example, a sliding time window (such as the past 24 hours or 7 days) can be set, and "same group of players" can be defined as a user group with similar game behaviors or characteristics. The total number of active events for this same group of players within the preset historical time window is determined, representing the total amount of behavior of this group within a specific time period. The relative magnitude of this total occurrence frequency to the total number of active events is used as the occurrence frequency, for example, calculated as (total occurrence frequency of event type / total number of active events). The event rarity is determined by a negative monotonically correlated relationship with this occurrence frequency. This means that the higher the occurrence frequency, the more common the event, and the lower its rarity. Conversely, the lower the occurrence frequency, the rarer the event. For example, the rarity of an event can be quantified using functions such as rarity = 1 - occurrence frequency or rarity = 1 / (occurrence frequency + ε) (where ε is a small constant to avoid division by zero).

[0047] When determining the real-time requirement component, a pre-defined latency upper bound mapping table is consulted based on the type identifier of the current collaborative analysis task to determine the latency upper bound of the current collaborative analysis task. Different collaborative analysis tasks have different real-time requirements. For example, the latency upper bound of a real-time anti-fraud task may be very short, while the latency upper bound of a monthly operations report generation task may be relatively long. This mapping table pre-stores the relationship between different task types and their corresponding latency upper bounds. A basic real-time metric is determined by a negative monotonically correlated relationship with this latency upper bound. The shorter the latency upper bound, the higher the real-time requirement of the task, and therefore the larger the basic real-time metric should be. For example, the basic real-time metric can be calculated as 1 / (latency upper bound + δ) (where δ is a small constant) to reflect the urgency of real-time performance. This basic real-time metric is normalized to a pre-defined numerical range, such as between 0 and 1, to obtain the real-time requirement component. Normalization ensures that this component can be compared and weighted with other attributes on a uniform scale.

[0048] In the weighted fusion process of content-sensitive attributes, the event rarity and the real-time requirement component serve as input values ​​for their respective dimensions, directly participating in the calculation of the content-sensitive attributes. This means that the rarer an event is and the higher the real-time requirement of the task, the higher its content-sensitive attribute value is likely to be. The event rarity also serves as a modulation factor in the multiplicative modulation of the behavior's impact. This mechanism ensures that even if a player's behavior data unit has a high impact, its contribution to the impact will be reduced by the rarity factor if the event type is very common (low rarity). Conversely, if the event type is very rare, its contribution to the impact will be amplified.

[0049] Through the above technical solution, this application provides a quantitative and refined method for calculating content-sensitive attributes. By statistically analyzing the frequency of player behavior data units within the same group to accurately determine event rarity, and combining this with a task type identifier query latency upper bound mapping table to quantify real-time requirement components, the calculation of content-sensitive attributes can more accurately reflect the actual value and timeliness of data units. In particular, using event rarity as a multiplicative modulation factor for behavioral impact effectively highlights key events that possess both high behavioral impact and uniqueness, avoiding the dilution effect of common events on the analysis results, thereby improving the accuracy and discriminative power of content-sensitive attributes. Using these quantified components as inputs for weighted fusion ensures the comprehensiveness and rationality of content-sensitive attributes, providing a more reliable basis for subsequent desensitization processing and resource allocation, thereby improving the efficiency and insight of game operation collaborative analysis.

[0050] In some embodiments described above in this application, online semantic parsing is proposed to obtain content-sensitive attributes from incoming player behavior data units, where behavioral impact is an important component of these attributes. However, directly using a complex master model to perform forward inference on each incoming player behavior data unit to estimate its behavioral impact can lead to significant computational burden and latency, especially in online analysis scenarios requiring real-time responses. This can result in slow system response and affect the timeliness and efficiency of collaborative analysis tasks.

[0051] To address this, this application further proposes a method based on a distilled simplified version of the main model corresponding to the current collaborative analysis task. This method performs forward inference on the player behavior data unit to estimate the behavioral influence of the player behavior data unit on the conclusion of the current collaborative analysis task, obtaining the behavioral influence degree. This includes: inputting the player behavior data unit into the distilled simplified version, which then performs feature encoding and forward calculation of the classification head on the player behavior data unit, outputting the conditional influence probability of the player behavior data unit on at least one preset analysis target. Normalization and truncation are then applied to this conditional influence probability to obtain the behavioral influence degree. This behavioral influence degree is used as the input value for the behavioral influence degree dimension in the content-sensitive attribute weighted fusion, and the behavioral influence degree is multiplicatively modulated by the rarity of the event in this weighted fusion. A specific example illustrates this: assuming the current collaborative analysis task is player churn prediction, the main model is a 24-layer Transformer network, and the student model obtained after distillation is a 6-layer Transformer network. When a player behavior data unit (e.g., an event record indicating that a player has completed a recharge operation) is received, the 6-layer student model performs the following calculations on that data unit: The data unit is converted into a 768-dimensional feature vector through an embedding layer, and then sequentially passed through a 6-layer feature encoding layer of a self-attention plus feedforward network. The classification head outputs a conditional influence probability of 0.72 relative to the churn prediction target (indicating that the recharge event reduces the churn probability). After applying Sigmoid normalization to truncate this conditional influence probability, the behavior influence score is obtained as 0.85 (on a scale of 0 to 1).

[0052] For example, player behavior data units are raw player actions, events, or state records received by the system, such as clicks, purchases, and movements within the game. To efficiently evaluate the impact of these data units, this application employs a distillation-simplified version. This distillation-simplified version is a smaller and lighter model trained using model distillation techniques compared to the original main model (used for collaborative analysis tasks). It approximates the behavior of the main model by learning its output or intermediate representations, thereby significantly reducing computational complexity and inference time while maintaining core predictive capabilities. When player behavior data units are input into this distillation-simplified version, feature encoding is performed, converting the raw, unstructured data into a numerical vector representation that the model can process. This can be achieved, for example, by mapping discrete features to a continuous space through embedding layers or by extracting sequence features through convolutional layers. The forward computation of the classification head refers to calculating the conditional probability of the player behavior data unit for one or more preset analysis objectives (e.g., predicting player churn, identifying abnormal behavior, assessing payment propensity, etc.) based on the encoded features through the output layer at the end of the model (usually a fully connected layer combined with an activation function, such as Softmax). These conditional influence probabilities quantify the potential contribution or impact of the behavioral data unit on a specific analytical conclusion. The training process for this distillation-simplified version is as follows: Using the original master model (i.e., the complete collaborative analysis model deployed on the collaborative computing node) as the teacher model, a student model with fewer layers and fewer parameters is constructed (e.g., the teacher model is a 24-layer Transformer, and the student model is a 6-layer Transformer). Historical player behavior data is used as training samples, and the soft labels output by the teacher model to the training samples (i.e., the category probability distribution softened by the temperature parameter T) are used as the training objective. The student model is trained by minimizing the KL divergence loss between the student model output and the teacher model soft labels. After training, this student model is deployed as a distillation-simplified version on a local computing node for rapid estimation of behavioral influence in online semantic parsing. The typical value of the temperature parameter T ranges from 2 to 20; a higher T value makes the soft label distribution smoother, which is beneficial for the student model to learn the teacher model's hidden knowledge.

[0053] After obtaining the conditional influence probabilities, this application performs normalization and truncation on them to obtain the behavioral influence degree. The conditional influence probability is the raw probability value output by the model, and its numerical range and distribution may vary depending on the model and the task. The normalization process aims to map these probability values ​​to a uniform, comparable numerical range, such as [0,1] or [0,X], to facilitate subsequent weighted fusion. For example, linear normalization or the Sigmoid function can be used for normalization. The truncation process, after normalization, restricts probability values ​​that exceed a specific threshold range, such as truncating excessively high or low probability values ​​to a preset maximum or minimum value, to prevent extreme values ​​from having an excessive impact on the calculation of subsequent content-sensitive attributes, thereby improving the robustness and stability of the calculation. Through this process, the original conditional influence probability is transformed into a more stable and comparable behavioral influence degree, which can more accurately reflect the actual impact of player behavior data units on the conclusions of the collaborative analysis task.

[0054] The behavioral impact is used as the input value for the behavioral impact dimension in the weighted fusion of content sensitivity attributes. The calculation of content sensitivity attributes is a weighted fusion process of multi-dimensional information, with behavioral impact being a key dimension that directly participates in the calculation. This behavioral impact is multiplicatively modulated by the event's rarity in this weighted fusion. Event rarity (defined by the above implementation method, reflecting the scarcity of the event type) multiplicatively modulates the behavioral impact, meaning that the magnitude of the behavioral impact is affected by the event's rarity. For example, an event with high behavioral impact, if its event rarity is also high (i.e., the event is very rare), may have its contribution to the content sensitivity attribute further amplified after multiplicative modulation, indicating that rare and highly influential behavioral data units should be assigned higher sensitivity. Conversely, if an event with high behavioral impact but low event rarity (i.e., the event is very common), its contribution to the content sensitivity attribute may not be excessively amplified. This multiplicative modulation mechanism makes the calculation of content-sensitive attributes more refined and reasonable, and can better balance the impact of behavior and the scarcity of events, thereby more accurately assessing the overall content sensitivity of player behavior data units.

[0055] By inputting player behavior data units into a simplified distillation version for feature encoding and forward computation of the classification head, and outputting conditional influence probabilities, this application can initially assess the potential impact of player behavior data units on the conclusions of collaborative analysis tasks with lower computational overhead and faster speed. Compared to directly using a complex master model for inference, the simplified distillation version reduces the consumption of local computing resources and shortens the response time of online semantic parsing, thereby improving the real-time performance of the entire data analysis and management method. Performing normalization truncation on the conditional influence probabilities can convert the original probability values ​​output by the model into stable and comparable behavior influence values, effectively avoiding the interference of extreme probability values ​​on the subsequent calculation of content-sensitive attributes, and enhancing the robustness of the calculation. Using this behavior influence value as the key input for weighted fusion of content-sensitive attributes and subjecting it to multiplicative modulation of event rarity ensures that the evaluation of content-sensitive attributes considers not only the influence of the behavior itself but also the scarcity of the event. This refined evaluation mechanism ensures that player behavior data units that are both highly influential and relatively rare receive higher content sensitivity attribute values, thus receiving more appropriate attention and processing in subsequent desensitization and load-aware simplification, thereby optimizing the balance between resource allocation and privacy protection.

[0056] The overall approach involves online semantic parsing of player behavior data units to obtain content-sensitive attributes, which are crucial for determining subsequent de-identification depth parameters and processing priority labels. Among these, "pattern completion contribution" is a key component of the content-sensitive attributes. However, in practical applications, accurately and dynamically evaluating the value of a player behavior data unit in identifying or completing predefined behavior patterns is a challenge. Relying solely on static rules or rough estimates may fail to fully reflect the actual role of the data unit in behavior pattern recognition, thus affecting the accuracy of the content-sensitive attributes. This, in turn, leads to suboptimal allocation of de-identification depth or processing priority, impacting the efficiency and accuracy of collaborative analysis.

[0057] To address this, this application further proposes a step for determining the pattern completion degree of the player behavior data unit within the detected behavior patterns, and obtaining the pattern completion contribution. This step includes: maintaining a set of candidate behavior patterns that are currently partially matched, where each candidate pattern records a sequence of matched key nodes; determining whether the event type corresponding to the player behavior data unit belongs to any unmatched key node in any candidate pattern; if so, determining the pattern completion contribution based on the weight ratio of the completed node in the corresponding candidate pattern; if not, setting the pattern completion contribution to a preset minimum contribution value.

[0058] For example, this method maintains a set of candidate behavioral patterns that are currently partially matched. This feature aims to dynamically track and manage predefined or learned behavioral patterns within the system. These behavioral patterns can be sequences of specific actions a player performs in the game, such as "completion of tutorial," "payment conversion path," or "abnormal cheating behavior." The system continuously monitors incoming player behavioral data units and attempts to match them with these predefined behavioral patterns. When a part of a behavioral pattern (i.e., one or more key nodes) is matched, the pattern is added to or updated in the candidate behavioral pattern set. Each candidate pattern in this set records the sequence of matched key nodes to help determine whether new player behavioral data units can further complete these patterns. For example, a "payment conversion" pattern might include key nodes such as "browsing the store," "adding to cart," and "clicking to pay." Once the player completes "browsing the store" and "adding to cart," the pattern is maintained as a partially matched candidate pattern, and the matched nodes are recorded. The behavioral patterns in this candidate behavioral pattern set can be obtained in the following ways: (a) Game operation experts pre-set key behavioral sequence templates based on domain knowledge, such as defining high-paying conversion patterns including key node sequences such as first-time recharge, three consecutive days of login, purchase of monthly card, and purchase of limited skins. (b) Frequent behavioral sequences are automatically discovered by performing sequence pattern mining algorithms (such as PrefixSpan and GSP algorithms) on historical player behavior data, and sequences that meet the minimum support threshold are selected as candidate behavioral patterns. Each key node in each candidate behavioral pattern has a pre-set weight, which reflects the relative importance of the node in completing the entire behavioral pattern recognition. The sum of the weights of all key nodes is normalized to 1.

[0059] The core logic of pattern completion is to determine whether the event type corresponding to the player behavior data unit belongs to any unmatched key node in any candidate pattern. When a new player behavior data unit is received, the system extracts its event type (e.g., "click to pay", "complete task X", etc.) and then iterates through the currently maintained set of candidate behavior patterns. For each candidate pattern in the set, the system checks all unmatched key nodes in that pattern. If the event type of the current player behavior data unit matches one or more unmatched key nodes in a candidate pattern, it indicates that the data unit may contribute to the completion of that pattern. For example, if the event type of a player behavior data unit is "click to pay", and there is a "paid conversion" pattern in the candidate behavior pattern set that has matched "browse the store" and "add to cart", and "click to pay" is the next unmatched key node in that pattern, then it is determined to belong to the candidate pattern.

[0060] If it falls under the category of a candidate mode, the contribution to mode completion is determined based on the weight percentage of the completed node within that mode. When the event type of a player behavior data unit is determined to belong to a key node in a candidate mode that has not yet been matched, that data unit plays a positive role in completing that mode. To quantify this contribution, the system determines the mode completion contribution based on the preset weight percentage of the completed node in the corresponding candidate mode. Different key nodes may have different importance in different behavior modes; for example, "click to pay" may have a higher weight in the "paid conversion" mode than "browsing the store." This weight percentage can be set and optimized based on expert experience, historical data analysis, or machine learning models. In this way, the mode completion contribution accurately reflects the actual value of the data unit for the completion of a specific behavior mode; the higher the weight percentage, the greater the contribution.

[0061] If a player's behavior data unit's event type fails to match any unmatched key node in any candidate behavior pattern, it means that the data unit currently makes no direct contribution to any known and ongoing pattern completion. In this case, to avoid its pattern completion contribution being completely ignored due to zero, while still reflecting its low pattern completion value, the system sets its pattern completion contribution to a preset minimum value. This minimum contribution value can be a positive number close to zero, ensuring that the data unit retains some weight in the weighted fusion of content-sensitive attributes, but its contribution to pattern completion is far lower than that of behavior data units that can complete key nodes.

[0062] The aforementioned technical solution enables dynamic and precise evaluation of the actual value of each player behavior data unit in behavior pattern recognition and completion. By maintaining and judging candidate behavior patterns and quantifying contribution based on the weight ratio of completion nodes, it avoids rough estimation of pattern completion contribution, making the calculation of content sensitivity attributes more refined and accurate. When a player behavior data unit can complete a key behavior pattern, its pattern completion contribution is reasonably increased, resulting in a higher content sensitivity attribute value. This can lead to deeper desensitization and higher processing priority, ensuring priority processing of high-value, highly relevant data and more meticulous privacy protection. Conversely, data unrelated to the current pattern completion is assigned a lower contribution, avoiding unnecessary resource waste. This mechanism allows the system to more intelligently identify and respond to key player behaviors, optimizes data processing workflows, and improves the efficiency and insight of game operation collaborative analysis.

[0063] In some of the embodiments described above in this application, it is necessary to accurately assess the privacy-sensitive attributes of player behavior data units to characterize the degree to which the information they carry can identify an individual. However, in practical applications, assessing the overall privacy risk solely based on the sensitivity of a single data field may have limitations. In some cases, a single field may not be highly sensitive, but when multiple fields are combined, they may uniquely identify a specific player, leading to an underestimation of the potential privacy leakage risk and affecting the effectiveness of subsequent de-identification processing.

[0064] In response, this application proposes a method for determining privacy-sensitive attributes, see [link to relevant documentation]. Figure 3 This method is implemented in the following way: 301. Perform field decomposition on the player behavior data unit to obtain the set of field types that make up the player behavior data unit.

[0065] 302. Based on the preset sensitivity mapping relationship, determine the sensitivity weight corresponding to each field type in the field type set.

[0066] 303. When there are multiple combinations of field types in the set of field types that can uniquely identify a specific player, perform joint enhancement processing on the sensitivity weights corresponding to each field type in the combination, and take the maximum value of the enhanced sensitivity weights as the privacy sensitive attribute.

[0067] 304. When the combination does not exist in the set of field types, take the maximum value of the sensitivity weight corresponding to each field type in the set of field types as the privacy sensitive attribute.

[0068] For example, field decomposition of player behavior data units refers to parsing the received raw player behavior data units into their constituent elements and identifying the type of each element. Player behavior data units typically exist in structured or semi-structured forms, such as logs or database records in JSON or XML formats, containing multiple fields such as player identifier, device identifier, IP address, geographical location, in-game event type, timestamp, and transaction amount. The process of field decomposition is to identify and extract these independent fields and their corresponding data types. This can be achieved through a predefined data schema parser, rule-based text analysis, or machine learning models (such as named entity recognition), ensuring that each meaningful data fragment is identified as an independent field type.

[0069] After obtaining the set of field types, the sensitivity weight of each field type in the set is determined according to a pre-defined sensitivity mapping relationship. This sensitivity mapping relationship is a pre-established set of rules or a lookup table that associates specific data field types with quantified sensitivity weights. The weight is typically a value between 0 and 1, representing the likelihood that the field type directly or indirectly identifies personal information. For example, fields such as player account IDs and device unique identifiers may be assigned higher sensitivity weights, while fields such as in-game item names and event occurrence times may be assigned lower sensitivity weights. This mapping relationship can be developed and maintained by data privacy experts and compliance teams based on data types, relevant laws and regulations (such as GDPR and CCPA), and experience from historical data breaches.

[0070] This application innovatively considers the privacy risks of field combinations. When multiple field types in the set can uniquely identify a specific player—for example, while a single IP address or device model may not be sufficient to uniquely identify a player, a combination of "IP address + device model + game behavior sequence within a specific time period" may have high uniqueness—the system performs joint boosting processing on the sensitivity weights corresponding to each field type in the combination. This joint boosting processing aims to reflect this combined identification effect; even if the original sensitivity weight of a single field in the combination is not high, its weight will be amplified due to its crucial role in combined identification. Boosting methods can include, but are not limited to, multiplicative factors, additive increments, or nonlinear functions. For example, if the combination of field A and field B has uniqueness, and their original weights are WA and WB respectively, then after boosting, they may become WA and WB, where WA > WA and WB > WB. The maximum value among the boosted sensitivity weights is taken as the privacy-sensitive attribute of the player's behavior data unit. This strategy of maximizing the value ensures that even if the original weight of a certain field in the combination is low, its increased weight can still reflect the overall high privacy risk due to its key role in the combination recognition. This allows privacy-sensitive attributes to accurately capture the privacy risks brought about by the combination recognition.

[0071] When a combination of fields does not exist in the set of field types (i.e., no combination uniquely identifies a specific player), the maximum sensitivity weight of each field type in the set is taken as the privacy-sensitive attribute. This means that in this case, the privacy risk of the player behavior data unit is primarily determined by its most sensitive single field. The system iterates through all decomposed field types, finds their corresponding weights in the sensitivity mapping relationship, and selects the maximum value among these weights as the privacy-sensitive attribute of the data unit.

[0072] The aforementioned technical solution enables a more refined assessment of the privacy-sensitive attributes of player behavior data units. For example, by decomposing data units into fields and determining the weights of each field based on a preset sensitivity mapping relationship, a foundation for privacy assessment is laid. More importantly, this solution innovatively considers the possibility that multiple field combinations may uniquely identify a specific player, and performs joint weight enhancement processing on the fields in these combinations, effectively avoiding the underestimation of privacy risks caused by "combined identification." When no such combination exists, the maximum single-field sensitivity weight is taken, ensuring comprehensive coverage of privacy risks. This mechanism makes the determination of privacy-sensitive attributes more accurate and comprehensive, and can more realistically reflect the degree to which player behavior data units can identify individuals. Therefore, the subsequent generation of desensitization depth parameters and processing priority labels will be more reasonable, effectively protecting user privacy while maximizing the preservation of data value and avoiding data utility loss due to excessive desensitization, thereby improving the robustness and effectiveness of the entire data analysis and management method.

[0073] In some embodiments described above in this application, the determination of privacy-sensitive attributes relies on a preset sensitivity mapping relationship. However, in practical applications, the privacy sensitivity of data fields is not static; it may dynamically adjust with the occurrence of data breaches, the evolution of attack techniques, and changes in users' awareness of privacy. If the sensitivity mapping relationship remains static, it may lead to insufficient privacy risk assessment of some fields, thereby increasing the risk of data breaches. Alternatively, it may result in excessive desensitization of other fields, causing unnecessary waste of computing resources and reduced data availability.

[0074] To address this, this application further proposes a method for constructing the preset sensitivity mapping relationship. For example, data field types are divided into multiple sensitivity levels based on their identifiability, including account identifier level, device identifier level, behavioral attribute level, and statistical count level. Sensitivity weights are assigned to each sensitivity level in descending order of identifiability, with values ​​ranging from 0 to 1. Furthermore, based on records of each field type being identified in historical data breaches, the sensitivity weights are adjusted online, increasing the sensitivity weights corresponding to the identified field types by a preset adjustment step size. The adjusted mapping relationship between each data field type and its corresponding sensitivity weight is then used as the preset sensitivity mapping relationship. In a specific example, the account identifier level is assigned a sensitivity weight of 0.95, the device identifier level 0.75, the behavioral attribute level 0.40, and the statistical count level 0.10. The preset correction step size is set to a fixed value between 0.05 and 0.15. When a certain field type is recorded as successfully identified in a historical data leakage event, its corresponding sensitivity weight is increased by the preset correction step size based on the original weight. The upper limit of the increased weight does not exceed 1.0.

[0075] Data fields are categorized into account identifier, device identifier, behavioral attribute, and statistical count levels based on their identifiability, aiming for refined management of different data field types. Account identifier fields, such as player IDs and registered phone numbers, directly and uniquely identify specific players, exhibiting the highest privacy sensitivity. Device identifier fields, such as device serial numbers and MAC addresses, uniquely identify the device used by a player, indirectly linking them to the player; their privacy sensitivity is second highest. Behavioral attribute fields, such as in-game operation records, spending records, and chat content, describe specific player behaviors or preferences. While difficult to directly identify a player on their own, they can be inferred from other information, exhibiting moderate privacy sensitivity. Statistical count fields, such as player online time and game activity statistics, are typically aggregated data, making it difficult to trace back to specific individuals; their privacy sensitivity is lowest. This tiered approach provides the foundation for subsequent sensitivity weighting.

[0076] Sensitivity levels are assigned decreasing sensitivity weights, ranging from 0 to 1, based on their identifiability from highest to lowest. This means that fields with higher identifiability have higher sensitivity weights, and vice versa. For example, account identifier fields might be assigned a weight close to 1, while statistical count fields might be assigned a weight close to 0. This weighting mechanism intuitively reflects the privacy priorities of different data fields, ensuring focused attention on high-risk data. The weight range of 0 to 1 facilitates subsequent normalization and standardized mathematical calculations.

[0077] Based on records where various field types were identified through reverse engineering in historical data breaches, the sensitivity weights are adjusted online. "Online adjustment" here means the system can dynamically adjust sensitivity weights according to actual privacy breaches or reverse engineering cases, rather than using a static configuration. "Records identified through reverse engineering" in historical data breaches refer to cases where, in actual attacks or analyses, seemingly insensitive fields were successfully used to identify specific players by combining them with other publicly available information or data. For example, a behavioral attribute field might be found to be correlated with publicly available social media information in a specific scenario, thereby identifying the player. This mechanism allows sensitivity assessment to adapt to the ever-changing privacy risk environment.

[0078] The sensitivity weight corresponding to the field type that has been reverse-identified is increased by a preset adjustment step size. When a field type is found to pose a risk of reverse-identification, its corresponding sensitivity weight will immediately increase by a preset fixed value. This adjustment aims to quickly improve the privacy protection level of the field, prompting the system to take stricter de-identification measures in subsequent processing, thereby effectively addressing newly emerging privacy risks.

[0079] The revised mapping relationship between each data field type and its corresponding sensitivity weight is used as the preset sensitivity mapping relationship. This means that after online correction, the original static mapping relationship is dynamically updated, forming a mapping table that can reflect the current privacy risk situation in real time. This updated mapping relationship will be applied to the subsequent determination of privacy-sensitive attributes to ensure that the system always performs data anonymization and management based on the latest risk assessment.

[0080] The above technical solution enables the construction of a dynamic and adaptive sensitivity mapping relationship. By classifying data field types into different sensitivity levels and assigning decreasing weights, a preliminary quantification of data privacy risks is achieved. More importantly, an online correction mechanism based on historical data breach events is introduced, allowing sensitivity weights to be dynamically adjusted according to actual reverse-engineering risk identification. When a field type is found to have a risk of being reverse-engineered, its sensitivity weight is promptly increased, ensuring that the field receives a higher level of privacy protection in subsequent processing. This effectively solves the problem that static sensitivity mapping relationships cannot adapt to dynamic privacy risk environments, avoids data breach risks caused by delayed risk assessment, and also avoids resource waste caused by excessive desensitization, improving the accuracy and efficiency of privacy protection and making the determination of privacy-sensitive attributes more precise and robust.

[0081] In some of the embodiments described above in this application, online semantic parsing of incoming player behavior data units is proposed to obtain content-sensitive and privacy-sensitive attributes. However, in subsequent processing, how to effectively transform these two different dimensions of attributes into specific and actionable guiding parameters to simultaneously consider data utilization value, privacy protection strength, and computational resource allocation efficiency is a technical problem that needs to be solved. Simply considering content-sensitive or privacy-sensitive attributes independently may lead to insufficient processing of high-value and high-privacy-risk data, or excessive investment of computational resources in low-value data, thereby affecting the accuracy, efficiency, and privacy compliance of the overall analysis.

[0082] To address this, this application further proposes a joint mapping between the content-sensitive attribute and the privacy-sensitive attribute to obtain the de-identification depth parameter and processing priority label. See [link to relevant documentation]. Figure 4 This includes the following steps.

[0083] 401. Perform a multiplicative fusion of the content sensitivity attribute and the privacy sensitivity attribute to obtain a joint index.

[0084] Multiplicative fusion is a method that multiplies two or more attribute values ​​to obtain a comprehensive index. Here, the content sensitivity attribute characterizes the expected impact of the player behavior data unit on the conclusions of the collaborative analysis task, while the privacy sensitivity attribute characterizes the degree to which the information it carries can be personally identifiable. By multiplicatively fusing these two attributes, a joint index can be obtained that simultaneously reflects the potential value and privacy risks of the data unit. For example, if both the content sensitivity and privacy sensitivity attributes are high, the joint index after multiplicative fusion will increase significantly, indicating that the data unit has both high analytical value and high privacy risks, requiring more refined processing. Conversely, if either attribute is low, the joint index will decrease accordingly, avoiding overestimation or underestimation of the comprehensive characteristics of the data unit by a single attribute. This fusion method ensures a more comprehensive and balanced evaluation of the data unit.

[0085] 402. Map the joint index to the preset cutoff layer depth range to obtain the desensitization depth parameter.

[0086] The lower limit of the cutoff layer depth range corresponds to the shallowest cutoff layer of the local feature extraction network, and the upper limit corresponds to the deepest cutoff layer. This joint index is a quantitative representation of the comprehensive value and risk of a data unit. To transform this abstract index into a concrete de-identification operation, this application maps it to a preset cutoff layer depth range. This range defines the possible locations of feature truncation in the local feature extraction network, from the shallowest cutoff layer (retaining more original information, low de-identification) to the deepest cutoff layer (extracting more abstract features, high de-identification). The mapping process can linearly or non-linearly transform the numerical range of the joint index to discrete or continuous values ​​of the cutoff layer depth parameter. For example, when the joint index is high, the mapped de-identification depth parameter will indicate a deeper cutoff layer, meaning that more feature extraction operations are performed on the local computing node, thereby generating a more abstract and more difficult-to-reverse-reference de-identified intermediate representation of the original data, thus enhancing privacy protection. Conversely, when the joint metric is low, it may be mapped to a shallower cutoff layer to reduce local computational overhead and preserve data utility as much as possible.

[0087] 403. Compare the content sensitivity attribute with the preset sensitivity threshold. When the content sensitivity attribute exceeds the preset sensitivity threshold, set the processing priority label to high priority level. When the content sensitivity attribute does not exceed the preset sensitivity threshold, set the processing priority label to low priority level.

[0088] To optimize the execution efficiency and resource allocation of collaborative analysis tasks, this application introduces a processing priority label. The label is primarily determined based on the content sensitivity attributes of player behavior data units. By comparing the content sensitivity attributes with a preset sensitivity threshold, the system can quickly determine the importance of the data unit to the current collaborative analysis task. If the content sensitivity attribute exceeds the threshold, it indicates that the data unit has a significant impact on the analysis conclusions and should be assigned a high priority level to ensure its priority scheduling in the computation queue, thereby accelerating the processing of key information. Conversely, if the content sensitivity attribute does not exceed the threshold, it is set to a low priority level. This priority allocation mechanism based on content sensitivity attributes allows computational resources to intelligently tilt towards data that contributes more to the analysis results, improving the overall system's response speed and task completion efficiency. The processing priority label is associated with corresponding computational resource weighting rules: high-priority de-identified intermediate representations are assigned a weight value W_high in the computation queue of the collaborative computing node, and low-priority de-identified intermediate representations are assigned a weight value W_low, where W_high > W_low > 0. Collaborative computing nodes are scheduled according to the weight values ​​corresponding to each de-identified intermediate representation using a weighted fair queuing method. The larger the weight value, the more computing time slices are allocated per unit of time. This computing resource weighting rule is part of the system policy parameter group and can be dynamically adjusted based on feedback signals.

[0089] The aforementioned technical solutions enable more refined management of player behavior data units, effectively balancing data utilization value, privacy protection, and computational resource efficiency. For example, by performing a multiplicative fusion of content-sensitive and privacy-sensitive attributes, a comprehensive joint indicator can be obtained. This indicator comprehensively reflects the potential value and privacy risks of data units, avoiding the one-sidedness of single-attribute evaluation. Based on the desensitization depth parameter obtained from this joint indicator mapping, the cutoff layer position of the local feature extraction network can be dynamically adjusted to ensure that high-value or high-privacy-risk data units receive deeper local desensitization processing, thereby maximizing the preservation of data utility for collaborative analysis while protecting user privacy. By comparing content-sensitive attributes with preset sensitivity thresholds and assigning high or low priority labels to data units, the system can intelligently allocate computational resources, prioritizing data with a greater impact on analysis conclusions, thus improving the response speed and overall efficiency of collaborative analysis tasks. This joint mapping mechanism allows data processing strategies to adaptively adjust according to the comprehensive characteristics of data units, thereby achieving better data management and resource utilization in complex game operation collaborative analysis scenarios.

[0090] In some of the embodiments described above in this application, a relatively simple mapping method may be used when jointly mapping content-sensitive attributes and privacy-sensitive attributes to obtain the de-identification depth parameter. However, such a simple mapping may not be able to adequately distinguish data units that simultaneously possess extremely high privacy risks and extremely high content value, resulting in insufficient or inaccurate de-identification depth for these key data units, thereby affecting data security and the effectiveness of collaborative analysis.

[0091] To address this, this application further proposes a specific method for mapping the joint indicator to a preset cutoff layer depth range to obtain the anonymization depth parameter. This includes: applying a nonlinear transformation to the joint indicator to obtain a transformed joint indicator; the nonlinear transformation applies a relatively larger amplification factor to joint indicators in the high-value range, allowing data units with overlapping high privacy risks and high content value to obtain a deeper cutoff layer position; linearly mapping the transformed joint indicator to the cutoff layer depth range to obtain the anonymization depth parameter; combining the adjustable parameter used in the nonlinear transformation with the upper and lower limits of the cutoff layer depth range to form an anonymization strategy parameter set; registering this anonymization strategy parameter set as an online parameter that can be updated by a feedback loop, which receives feedback signals from the game operation collaborative analysis conclusions.

[0092] For example, applying a nonlinear transformation to the joint indicator yields a transformed joint indicator. This nonlinear transformation applies a relatively larger amplification factor to the joint indicator in the high-value range, allowing data units with both high privacy risks and high content value to obtain a deeper cutoff layer. This nonlinear transformation aims to break the simple linear relationship between the joint indicator and the anonymization depth parameter, especially for player behavior data units with high joint indicator values. By applying a relatively larger amplification factor to the high-value range of the joint indicator, it ensures that data units with both high privacy risks and high content value are assigned a deeper cutoff layer. For example, nonlinear transformations can be implemented using exponential functions, polynomial functions, or piecewise functions. If an exponential function is used, the joint indicator can be used as the independent variable of the exponential function, and the amplification factor can be controlled by adjusting the base and coefficients, so that the larger the joint indicator, the faster its transformed value increases. If a piecewise function is used, after the joint indicator reaches a certain preset threshold, it can switch to a linear or nonlinear segment with a larger slope, thereby achieving additional amplification of the high-value range. This processing method can more accurately reflect the overall sensitivity of data units, avoiding insufficient or excessive desensitization caused by simple linear mapping.

[0093] After performing a nonlinear transformation on the joint index, the resulting transformed joint index may fall within an irregular numerical range. To convert it into a usable cutoff layer location in the local feature extraction network, the transformed joint index needs to be linearly mapped to the cutoff layer depth range to obtain the desensitization depth parameter. This linear mapping proportionally scales the current numerical range of the transformed joint index to the preset cutoff layer depth range. For example, if the range of the transformed joint index is [A, B], and the cutoff layer depth range is [min_depth, max_depth], the mapping can be achieved using the formula: desensitization depth parameter = min_depth + (transformed joint index - A)(max_depth - min_depth) / (BA). This ensures that the desensitization depth parameter always falls within the effective cutoff layer range supported by the local feature extraction network, thereby guiding subsequent hierarchical feature extraction operations.

[0094] To achieve dynamic adjustment of the anonymization strategy, the adjustable parameters used in the nonlinear transformation and the upper and lower limits of the cutoff layer depth range are combined into an anonymization strategy parameter set. This parameter set is registered as online parameters that can be updated by a feedback loop, which receives feedback signals from the game operation collaborative analysis conclusions. The anonymization strategy parameter set is a key configuration item for dynamically adjusting the anonymization behavior. It includes specific parameters of the nonlinear transformation (e.g., the base of the exponential function, the coefficients of the polynomial function, the inflection point and slope of the piecewise function, etc.) and specific boundary values ​​of the cutoff layer depth range. Combining these parameters into a set and registering them as parameters that can be updated online by the feedback loop means that the system can self-optimize and adjust based on actual operating results. The feedback loop evaluates the effectiveness of the current anonymization strategy by receiving feedback signals from the game operation collaborative analysis conclusions (e.g., the accuracy of the analysis results, privacy leakage risk assessment, model drift, etc.). When feedback signals indicate that the current strategy is insufficient, the feedback loop generates update instructions to adjust the adjustable parameters in the desensitization strategy parameter set and the upper and lower limits of the cutoff layer depth range. For example, if the feedback indicates that the privacy risk is too high, the feedback loop may adjust the parameters of the nonlinear transformation to make it amplify the high-value joint indicators more, thereby prompting a deeper cutoff.

[0095] By introducing nonlinear transformations through the aforementioned technical solutions, joint indicators can be distinguished more finely, especially data units that simultaneously possess high privacy risks and high content value, which can be amplified more significantly. This allows these key data units to obtain deeper cutoff layer locations, resulting in more thorough desensitization processing in the local feature extraction network, enhancing the strength and targeting of privacy protection. The adjustable parameters of the nonlinear transformation are combined with the upper and lower limits of the cutoff layer depth range to form a desensitization strategy parameter set, which can be updated online through a feedback loop, giving the desensitization strategy dynamic adaptability. The system can adjust the desensitization strategy in real time based on feedback signals from game operation collaborative analysis conclusions. For example, when an increase in potential privacy risks is detected, the amplification factor of the nonlinear transformation can be increased to promote deeper cutoff. When the analytical utility is impaired, appropriate adjustments can be made to balance the situation. This adaptive mechanism ensures that the desensitization strategy can be continuously optimized, maximizing the effectiveness of collaborative analysis while ensuring data privacy security, avoiding insufficient or excessive desensitization that may result from static strategies, and improving the robustness and efficiency of the entire data processing flow.

[0096] In some embodiments described above, this application proposes online semantic parsing of incoming player behavior data units to obtain content-sensitive attributes and privacy-sensitive attributes. A joint mapping is then performed on these content-sensitive and privacy-sensitive attributes to obtain a de-identification depth parameter and a processing priority label. Specifically, this joint mapping involves performing a multiplicative fusion of the content-sensitive and privacy-sensitive attributes to obtain a joint index, which is then mapped to a preset cutoff depth range to obtain the de-identification depth parameter. To enable data units with both high privacy risks and high content value to obtain deeper cutoff positions, a nonlinear transformation is applied to the joint index during the mapping process. The adjustable parameters used in this nonlinear transformation, along with the upper and lower limits of the cutoff depth range, are combined to form a de-identification strategy parameter set, which is registered as an online parameter that can be updated by the feedback loop. However, in practical applications, the risk level of specific data categories in the game operation collaborative analysis conclusions may change dynamically. If the adjustable parameters of the nonlinear transformation cannot respond to these changes in a timely and accurate manner, it may lead to a lag in the desensitization strategy, which may fail to effectively balance data utility and privacy protection. Insufficient desensitization may even increase the risk of privacy leakage, or excessive desensitization may reduce the accuracy of the analysis conclusions.

[0097] To address this, this application further proposes a method for online updating of the adjustable parameters of the nonlinear transformation. Specifically, the method includes: receiving a desensitization strategy update instruction from the feedback loop, triggered by a change in the risk level of a specific data category in the game operation collaborative analysis conclusion; increasing or decreasing the value of the adjustable parameter according to the direction of the risk level change to adjust the amplification factor of the nonlinear transformation on the joint indicators in the high-value range; applying the updated adjustable parameter to the nonlinear transformation of the joint indicators in the next cycle, causing a change in the cutoff layer position corresponding to the same joint indicator value in the next cycle. In one specific embodiment, the nonlinear transformation adopts an exponential function form f(x) = a × exp(b × x), where x is the joint indicator, and a and b are adjustable parameters. When the desensitization strategy update instruction indicates an increase in risk level, parameter b is increased by a preset step size delta_b (e.g., delta_b = 0.1), increasing the growth rate of the exp(b × x) term, thereby enhancing the amplification factor on the joint indicators in the high-value range. When the risk level decreases, parameter b is decreased by delta_b. The updated parameter b×=b+ / -delta_b is applied to the nonlinear transformation of the next cycle, causing the transformed value corresponding to the same joint index value to change, which in turn leads to a change in the location of the cutoff layer.

[0098] For example, upon receiving an anonymization strategy update instruction from this feedback loop, this step aims to ensure that the anonymization strategy responds to real-time feedback generated during system operation. This feedback loop can be an independent monitoring and decision-making module that continuously analyzes game operation collaborative analysis conclusions, such as by monitoring risk indicators (e.g., fraud rates, privacy breach reports, user complaints) for specific data categories (e.g., data involving user transactions, social interactions, or abnormal behavior). When these risk indicators change significantly, such as a data category's risk level changing from "medium" to "high," or from "low" to "medium," the feedback loop generates and sends an anonymization strategy update instruction. This instruction can include information such as the data category identifier triggering the update, the direction of the risk change (increase or decrease), and the magnitude of the change, and is passed to the parameter update module via mechanisms such as message queues, API interfaces, or shared memory.

[0099] Based on the direction of the risk level change, the value of the adjustable parameter is increased or decreased to adjust the amplification factor of the nonlinear transformation on the joint indicators in the high-value range. This step is crucial for achieving dynamic adaptive desensitization. The adjustable parameter is one or more coefficients in the nonlinear transformation function. For example, if the nonlinear transformation uses an exponential function form f(x) = aexp(bx) or a power function form f(x) = ax^b, then parameter a or b can be used as the adjustable parameter. When an instruction to increase the risk level is received, the system increases the value of the adjustable parameter, for example, by increasing the base of the exponential function or the exponent of the power function. This makes the slope of the nonlinear transformation curve steeper in the high-value range (i.e., higher joint indicator values, indicating high privacy risk and high content value), resulting in a greater amplification factor for the joint indicators. This means that even if the joint indicator value increases slightly, the desensitization depth parameter it maps will increase significantly, leading to a deeper cutoff layer and thus enhancing the desensitization strength. Conversely, when the risk level decreases, the system will reduce the value of the adjustable parameters, reduce the amplification factor of the nonlinear transformation, and appropriately weaken the desensitization intensity in order to retain the data utility as much as possible while ensuring privacy and security.

[0100] The updated adjustable parameters are applied to the nonlinear transformation of the joint index in the next cycle, causing the cutoff layer position corresponding to the same joint index value to change in the next cycle. This step ensures the smoothness and effectiveness of parameter updates. The updated adjustable parameters do not immediately affect the currently processed data unit, but are loaded and applied to the next cycle of system processing (e.g., the next data batch, the next time window, or the next task queue). This periodic update mechanism avoids system instability that may be caused by frequent parameter changes. By applying the updated parameters, even if subsequent incoming player behavior data units have the same joint index value as before, they will be mapped to different desensitization depth parameters due to the adjustment of the nonlinear transformation function, resulting in the selection of different cutoff layer positions in the local feature extraction network. This allows the system to dynamically adjust the desensitization strategy based on the latest risk assessment, achieving fine-grained management of data privacy protection and utility preservation.

[0101] Through the above technical solution, this application provides a mechanism for dynamically adjusting the desensitization strategy. When the game operation collaborative analysis conclusion indicates a change in the risk level of a specific data category, the system can promptly receive the desensitization strategy update instruction and adaptively adjust the adjustable parameters of the nonlinear transformation according to the direction of the risk change. This adjustment allows the amplification factor of the nonlinear transformation on the joint indicators in the high-value range to change dynamically. Thus, in the next cycle, even player behavior data units with the same content sensitivity and privacy sensitivity attributes can be mapped to a deeper or shallower cutoff layer position according to the latest risk situation. This effectively solves the problem that traditional desensitization strategies are difficult to adapt to dynamically changing risk environments, ensuring that privacy protection can be strengthened in a timely manner when the risk increases to avoid privacy leakage. When the risk decreases, the desensitization intensity can be appropriately relaxed to maximize the retention of data utility, thereby achieving a better effect in dynamically balancing privacy protection and data utilization, and improving the robustness and adaptability of the entire collaborative analysis system.

[0102] In some embodiments described above in this application, to balance the privacy protection and data utility of player behavior data units, it is necessary to perform hierarchical feature extraction and truncation in the local feature extraction network based on the de-identification depth parameter to obtain a de-identified intermediate representation. However, in the implementation process, how to accurately translate the dynamically changing de-identification depth parameter into the specific truncation behavior of the local feature extraction network, and ensure that only the necessary de-identified information is transmitted to the collaborative computing nodes while maintaining the effectiveness of collaborative analysis, is a key challenge in realizing this technical solution. Inappropriate selection of the truncation layer or inaccurate data transmission may lead to privacy leakage risks or loss of data utility.

[0103] To address this, this application proposes a more refined hierarchical feature extraction and truncation method. For example, based on the aforementioned anonymization depth parameter, the Kth layer in the local feature extraction network is determined as the truncation layer. This anonymization depth parameter is a dynamic value, reflecting the sensitivity and analytical value of the player behavior data unit. The local feature extraction network is typically a multi-layer deep learning model, with each layer extracting features at different levels of abstraction. Mapping the anonymization depth parameter to the Kth layer of the network as the truncation layer means that the system will choose at which layer of the network to stop local computation and transmit data based on the sensitivity of the data. A larger anonymization depth parameter indicates more sensitive data or a greater need for privacy protection; therefore, the deeper the truncation layer is in the local feature extraction network (i.e., the larger the K value), the more abstract the locally extracted features become, and the more difficult it is to infer the original data, thus providing stronger privacy protection. For example, a mapping function or lookup table can be predefined to map the range of the anonymization depth parameter to the layer index of the local feature extraction network, thereby achieving a dynamic mapping from the anonymization depth parameter to the truncation layer K.

[0104] The first K layers of the local feature extraction network are computed on a local compute node to obtain the intermediate activation features output by the Kth layer. The local compute node is where the data is generated or initially processed, such as on a player's device or in a game data analysis and management system for edge users. The first K layers of forward computation refer to the computation performed layer by layer in the local feature extraction network, starting from the input layer and proceeding up to the Kth layer. The intermediate activation features output by the Kth layer represent the representation of the original data after K layers of abstraction. This process ensures that the data has undergone a certain degree of desensitization before leaving the local compute node. The local feature extraction network can be a pre-trained neural network model with multiple layers, such as convolutional layers, pooling layers, or fully connected layers. Forward computation involves passing the input data layer by layer according to the network topology and computing the output of each layer. When the Kth layer is reached, its output tensor or vector is the intermediate activation feature.

[0105] The intermediate activation features output from the Kth layer are used as a de-identified intermediate representation, and only this de-identified intermediate representation is output to the collaborative computing node. The collaborative computing node is a central user game data analysis and management system or cloud platform responsible for aggregating and analyzing data from multiple local computing nodes. Outputting only the de-identified intermediate representation to the collaborative computing node emphasizes the principle of minimizing data transmission, i.e., transmitting only the necessary information after de-identification processing, rather than the original data or complete features. The collaborative computing node performs forward computation on the remaining layers after the truncated layer in the local feature extraction network based on this de-identified intermediate representation. This means that after completing the forward computation of the first K layers, the local computing node serializes the intermediate activation features and sends them to the collaborative computing node via the network. Upon receiving the data, the collaborative computing node deserializes it and uses it as input to the (K+1)th layer of the local feature extraction network, continuing forward computation until the next layer of the network.

[0106] Through the above technical solution, this application clarifies the specific application of the desensitization depth parameter in the local feature extraction network, namely, dynamically determining the truncation layer K based on the parameter value. This enables the local computing node to accurately perform the forward computation of the first K layers and transmit only the intermediate activation features, after preliminary abstraction and desensitization, as desensitized intermediate representations to the collaborative computing node. This hierarchical computation and selective transmission mechanism effectively avoids the risk of directly exposing the original sensitive data to the collaborative computing environment, improving the privacy protection level of player behavior data units. The collaborative computing node can continue to complete the feature extraction of the remaining layers based on the received desensitized intermediate representations, ensuring that the collaborative analysis task can obtain complete feature representations, thereby maintaining data utility and analysis accuracy while ensuring privacy. This distributed, on-demand desensitized feature extraction method optimizes data transmission efficiency, reduces the consumption of computing resources, and provides a secure and reliable data foundation for subsequent fusion analysis.

[0107] In some embodiments described above in this application, a technical solution is proposed to perform hierarchical feature extraction and truncation in a local feature extraction network based on the desensitization depth parameter to obtain a desensitized intermediate representation, and only output this desensitized intermediate representation to the collaborative computing node. However, in its implementation, for player behavior data units that simultaneously possess high content sensitivity and high privacy sensitivity attributes, even after hierarchical truncation, the obtained intermediate activation features may still contain a high degree of sensitive information, posing a certain risk of privacy leakage. Furthermore, the data volume may still be large, resulting in high transmission bandwidth consumption and computational resource consumption.

[0108] To address this, this application further proposes an optimization scheme. After obtaining the intermediate activation feature output from the Kth layer, it determines whether the low-rank decomposition triggering condition is met. The low-rank decomposition triggering condition is that the content-sensitive attribute exceeds a preset content threshold and the privacy-sensitive attribute exceeds a preset privacy threshold. When the low-rank decomposition triggering condition is met, low-rank decomposition is performed on the intermediate activation feature, splitting it into low-rank decomposition component groups. Each component in the low-rank decomposition component group is multiplicatively reconstructed to approximately restore the intermediate activation feature. These low-rank decomposition component groups replace the intermediate activation feature as the desensitized intermediate representation and are output to the collaborative computing node. When the low-rank decomposition triggering condition is not met, the low-rank decomposition is skipped, and the intermediate activation feature is directly used as the desensitized intermediate representation and output to the collaborative computing node.

[0109] For example, after completing the first K layers of forward computation of the local feature extraction network on the local computing node and obtaining the intermediate activation features output by the Kth layer, the system evaluates the sensitivity of the player behavior data unit corresponding to this intermediate activation feature. This evaluation process relies on two key indicators: content sensitivity and privacy sensitivity. The content sensitivity attribute characterizes the expected impact of the player behavior data unit on the conclusion of the current collaborative analysis task; the higher the value, the more important the data unit is to the analysis results. The privacy sensitivity attribute characterizes the degree to which the information carried by the player behavior data unit can be personally identified; the higher the value, the greater the privacy risk of the data unit. The system presets content thresholds and privacy thresholds. These thresholds are empirical values ​​configured based on actual application scenarios, privacy protection strategies, and data value requirements, used to define whether a data unit belongs to the category of "high content sensitivity" and "high privacy sensitivity." For example, it can be set that when the content sensitivity attribute value exceeds a certain percentile (e.g., 90%), it is considered high content sensitivity, and when the privacy sensitivity attribute value exceeds a certain percentile (e.g., 80%), it is considered high privacy sensitivity. The low-rank decomposition trigger condition is only considered met when both the content-sensitive attribute and the privacy-sensitive attribute exceed their respective preset thresholds.

[0110] When the low-rank decomposition trigger condition is met, the system will perform low-rank decomposition on the intermediate activation features output from the Kth layer. Low-rank decomposition is an effective data dimensionality reduction and privacy protection technique, which can be achieved using methods such as Singular Value Decomposition (SVD), Principal Component Analysis (PCA), or Non-negative Matrix Factorization (NMF). Taking SVD as an example, the intermediate activation feature matrix M can be decomposed into the product of three matrices U, S, and V. By retaining the largest r singular values ​​in the S matrix and their corresponding columns in the U and V matrices, the optimal rank r approximation of M can be obtained. These decomposed matrices or their key parts constitute the low-rank decomposition component groups. These component groups reduce the data dimensionality and data volume while retaining the main information of the original intermediate activation features, and due to their approximation, increase the difficulty of reverse engineering to restore the original sensitive information. The local computing node no longer directly transmits the original intermediate activation features, but instead sends these low-rank decomposition component groups as desensitized intermediate representations to the collaborative computing node. After receiving the low-rank decomposed component sets, the collaborative computing node can approximate the original intermediate activation features through multiplicative reconstruction (e.g., matrix multiplication) for use in subsequent forward computation of the remaining layers and the collaborative analysis model. In one specific implementation, assuming the intermediate activation feature matrix M has a dimension of d×n, where d is the feature dimension and n is the batch size, singular value decomposition (SVD) is performed on M: M = U × Sigma × V^T. The largest r singular values ​​in Sigma are retained, where r = min(d,n) × rho, and rho is the retention ratio, the value of which is determined by a joint index of content-sensitive and privacy-sensitive attributes. The higher the joint index, the smaller the value of rho (i.e., fewer singular values ​​are retained, and the decomposition is more thorough). The low-rank decomposed component sets consist of truncated U_r, Sigma_r, and V_r. The collaborative computing node approximates the intermediate activation features by calculating U_r × Sigma_r × V_r^T and then continues to perform subsequent layer computations.

[0111] Conversely, when the low-rank decomposition trigger condition is not met—that is, the content-sensitive or privacy-sensitive attributes of the player behavior data unit do not simultaneously reach the preset high sensitivity level—the system will skip the low-rank decomposition step. In this case, the intermediate activation features output from the Kth layer are directly used as the desensitized intermediate representation and output to the collaborative computing node. This avoids unnecessary computational overhead, because for data with low sensitivity, hierarchical feature extraction and truncation alone are sufficient to meet the requirements of privacy protection and data transmission efficiency.

[0112] The above technical solution intelligently identifies player behavior data units that possess both high content sensitivity and high privacy sensitivity. For this type of data, after conventional hierarchical feature extraction and truncation, low-rank decomposition is further performed, transforming the intermediate activation features into more compact and harder-to-reverse-reconstruct low-rank decomposed component groups. This enhances privacy protection capabilities, making it difficult for attackers to accurately reconstruct the original sensitive information even if they obtain the anonymized intermediate representation, thus effectively reducing the risk of privacy leakage. The data volume of the low-rank decomposed component groups is typically much smaller than the original intermediate activation features, effectively reducing the data transmission bandwidth consumption from local computing nodes to collaborative computing nodes and improving the overall communication efficiency of the system. This solution avoids performing low-rank decomposition on all data units, processing only the data that truly requires deep anonymization, thus achieving a good balance between privacy protection, data utility, and computational resource consumption. Although low-rank decomposition is performed, because it retains the main information of the intermediate activation features, collaborative computing nodes can still approximately reconstruct the original features based on these component groups for subsequent analysis, ensuring the accuracy of the game operation collaborative analysis conclusions.

[0113] To address this, this application further proposes a specific method for collaborative computing nodes to perform forward computation of the remaining layers after the truncated layer in the local feature extraction network based on the anonymized intermediate representation. In some embodiments of this application, online semantic parsing and joint mapping are performed on the incoming player behavior data units to obtain anonymization depth parameters and processing priority labels. Based on the anonymization depth parameters, hierarchical feature extraction is performed in the local feature extraction network and truncated to obtain an anonymized intermediate representation, which is then output only to the collaborative computing node. However, if these anonymized intermediate representations are not processed indiscriminately after being received by the collaborative computing node, it may lead to delays in the analysis of high-priority data units, affecting the real-time performance and effectiveness of the collaborative analysis task, especially under conditions of limited computing resources.

[0114] To address the aforementioned issues, this application proposes that after receiving the de-identified intermediate representation, the collaborative computing node determines the scheduling position of the de-identified intermediate representation in its computation queue based on the processing priority label. The higher-priority de-identified intermediate representation is prioritized over the lower-priority de-identified intermediate representation in the forward computation. The remaining layer forward computation is then performed on the collaborative computing node to obtain the complete feature representation. Finally, the collaborative analysis model performs cross-source feature aggregation and classification inference on this complete feature representation to obtain the cross-source collaborative inference result.

[0115] For example, a collaborative computing node is a centralized or distributed computing entity responsible for receiving anonymized intermediate representations from multiple local computing nodes and further performing feature extraction and collaborative analysis tasks. Upon receiving the anonymized intermediate representation, the node places it into its internal computing queue based on the processing priority tag transmitted with it. The processing priority tag is determined by comparing the content sensitivity attributes of the player behavior data unit with a preset sensitivity threshold, and is divided into high-priority and low-priority levels. The collaborative computing node uses a scheduling mechanism to ensure that high-priority anonymized intermediate representations enter the forward computing process before low-priority ones. This can be achieved through various strategies such as priority queues, preemptive scheduling, or time-slice allocation to ensure timely processing of critical data units. Data transmission between the local computing node and the collaborative computing node is achieved through the following mechanism: the local computing node encapsulates the anonymized intermediate representation, its corresponding processing priority tag, and cross-source association identifier into a structured data packet, and asynchronously sends it to the collaborative computing node via a secure network channel encrypted with TLS using gRPC or HTTP / 2 protocol. The collaborative computing nodes maintain a priority-aware receive buffer, sorting data packets according to their processing priority tags, with higher-priority packets entering the computing queue first. Within the computing queue, each collaborative computing node maintains a processing status record for each de-identified intermediate representation, including the source local computing node identifier, receive timestamp, processing priority tag, and current computing progress (number of remaining layers completed).

[0116] On the collaborative computing node, for the received anonymized intermediate representation, the remaining layers after the truncated layers in the local feature extraction network are further computed. According to the above implementation, the local computing node only completes the first K layers of computation in the local feature extraction network and uses the intermediate activation features output from the Kth layer as the anonymized intermediate representation. Therefore, the collaborative computing node needs to complete computation from the (K+1)th layer to the first layer to transform the anonymized intermediate representation into a complete feature representation containing richer and more abstract semantic information. This process typically involves subsequent layer operations in the neural network, such as convolution, pooling, and fully connected operations, aiming to extract more discriminative features from the intermediate representation for the analysis task.

[0117] After obtaining complete feature representations, the collaborative analysis model performs cross-source feature aggregation and classification inference on these representations. Cross-source feature aggregation refers to aligning and integrating complete feature representations from different players, time periods, or data sources based on their cross-source association identifiers (such as player identifiers, match identifiers, or time window identifiers). Aggregation methods can include feature concatenation, attention mechanism weighting, graph neural network aggregation, etc., aiming to fuse multi-source information to obtain a more comprehensive context. The collaborative analysis model then performs classification inference based on the aggregated cross-source fused features, for example, through deep learning classifiers, regression models, or clustering algorithms, to generate collaborative analysis conclusions on game operations, such as player churn probability, abnormal behavior markers, or payment tendency scores, at least one of these.

[0118] Through the aforementioned technical solution, collaborative computing nodes can intelligently schedule the received anonymized intermediate representations based on processing priority tags, ensuring that high-priority data units are processed first, thereby improving the real-time performance and response speed of collaborative analysis tasks. This priority scheduling mechanism effectively avoids analysis delays caused by resource contention, especially when processing massive amounts of player behavior data, ensuring timely insights into key events and high-value information. The remaining layer forward computation is completed on the collaborative computing nodes, generating complete feature representations. The collaborative analysis model then performs cross-source feature aggregation and classification inference, enabling the analysis process to fully utilize locally anonymized feature information and combine it with multi-source data for in-depth mining, generating accurate and comprehensive collaborative analysis conclusions on game operations. This not only optimizes the utilization efficiency of computing resources but also enhances the overall performance and decision support capabilities of the analysis system.

[0119] In some of the embodiments described above in this application, load-aware simplification of player behavior data units is proposed based on content-sensitive attributes and the real-time load status of local computing nodes to obtain retained data units. However, in practice, how to accurately and dynamically adjust the data simplification strategy according to the real-time resource status of local computing nodes, in order to effectively manage limited computing and communication resources while ensuring the quality of analysis tasks and avoiding resource overload or excessive data discarding, is a technical problem that needs to be solved.

[0120] To address this, this application further proposes a load-aware simplification method, specifically including: using the content-sensitive attribute as the utility score of the player behavior data unit. This content-sensitive attribute characterizes the expected impact of the player behavior data unit on the conclusions of the current collaborative analysis task. Directly quantifying it as a utility score provides a unified and intuitive measurement standard for subsequent data simplification decisions, clearly demonstrating the importance of the data unit. For example, the utility score can directly use the original calculation result of the content-sensitive attribute, or it can be normalized to adapt to a specific threshold comparison mechanism.

[0121] This application continuously monitors the real-time load status of the local computing node. This real-time load status is a key indicator reflecting the current resource availability of the local computing node, and may include at least one of the following: compute utilization, task queue depth, and uplink bandwidth reserve. For example, compute utilization can be obtained by periodically querying the system's CPU or GPU utilization. Task queue depth can be sensed in real time by monitoring the length of the local task scheduler or message queue. Uplink bandwidth reserve can be calculated using network interface traffic statistics and total bandwidth. Real-time monitoring of these indicators ensures that streamlined decision-making can respond promptly to the actual resource pressure on the system.

[0122] Based on this, this application dynamically adjusts the minimum utility threshold upwards or downwards according to the deviation of the real-time load status from the preset load reference value. This minimum utility threshold is a key parameter determining whether player behavior data units are retained. When the real-time load status indicates resource scarcity (e.g., excessively high computing power utilization, excessively deep task queues, or insufficient uplink communication bandwidth), the system will correspondingly increase the minimum utility threshold, meaning that only data units with higher utility scores will be retained. Conversely, when resources are abundant, the system will decrease the minimum utility threshold to retain more data. This dynamic adjustment mechanism is the core of load awareness, allowing the system to flexibly adjust its data retention strategy based on actual resource conditions.

[0123] This application downsamples or discards player behavior data units with utility scores below a minimum utility threshold, while retaining player behavior data units with utility scores at or above the minimum utility threshold. Data units with utility scores significantly below the dynamic threshold can be directly discarded to maximize resource efficiency. For data units with utility scores slightly below or close to the threshold, a downsampling strategy can be employed; for example, retaining one data unit every certain number of data units, or retaining only some key fields, thereby preserving data diversity as much as possible while minimizing resource consumption. In this way, the system can intelligently select data units that contribute significantly to the analysis task and have reasonable resource consumption for subsequent processing.

[0124] The aforementioned technical solution quantifies the content-sensitive attributes of player behavior data units into utility scores. Combined with precise monitoring of the real-time load status of local computing nodes, the minimum utility threshold can be dynamically adjusted. When local computing nodes face resource constraints, the system intelligently raises the minimum utility threshold, prioritizing downsampling or discarding data units with lower utility. This effectively alleviates system pressure, avoids resource overload, and ensures the smooth execution of critical analysis tasks. Conversely, when resources are abundant, the system lowers the minimum utility threshold, retaining as many player behavior data units as possible to maximize data utilization and the comprehensiveness of the analysis. This load-aware simplification mechanism allows the data processing process to flexibly adapt to constantly changing computing environments. It significantly optimizes the resource utilization efficiency of local computing nodes while ensuring the quality of game operation collaborative analysis conclusions, avoiding unnecessary resource waste or loss of critical data, thereby improving the robustness and efficiency of the entire data analysis and management system.

[0125] In some of the embodiments described above in this application, although it is proposed to adjust the minimum utility threshold based on the deviation between the real-time load status of the local computing node and the preset load reference value in order to achieve load-aware simplification, this adjustment method may not fully consider the actual pressure of task processing within the computing node and the strict latency requirements of collaborative analysis tasks. This may result in insufficient adaptability of the simplification strategy in scenarios with drastic load fluctuations or latency-sensitive tasks, which may lead to excessive discarding of important data or retention of too much redundant data, affecting the accuracy or real-time performance of collaborative analysis.

[0126] To address this, this application further proposes a method for adjusting the minimum utility threshold based on the deviation between the real-time load status and a preset load reference value. This method includes: determining the current task queue depth of the local computing node and a preset upper bound for latency assurance; determining a queue pressure factor based on the relative magnitude of the current task queue depth to the preset queue depth reference value; determining a latency urgency factor based on the urgency of the latency assurance upper bound relative to the preset latency reference value; and combining the queue pressure factor and the latency urgency factor into a load penalty term. A weighted summation is performed on the minimum utility threshold of the previous period and the load penalty term, and the weighted summation result is used as the minimum utility threshold for the current period. This minimum utility threshold for the current period is used for load perception simplification judgment of the next batch of player behavior data units.

[0127] For example, the current task queue depth of the local compute node and a preset latency guarantee upper bound are determined. The current task queue depth refers to the number of tasks currently waiting to be processed on the local compute node, a direct indicator of the node's real-time processing pressure. The local compute node can continuously monitor its internal task scheduler or the application programming interface (API) provided by the operating system to obtain the number of tasks currently waiting to be processed. The preset latency guarantee upper bound defines the tolerance of collaborative analysis tasks for data processing completion time, a key parameter for measuring task urgency. This latency guarantee upper bound can be preset based on the type, importance, or user configuration of the collaborative analysis task. For example, for anti-fraud analysis with high real-time requirements, the latency upper bound may be very short, while for offline trend analysis, it may be longer. These parameters can be loaded during task initialization or dynamically obtained through a configuration service.

[0128] The queue pressure factor is determined based on the relative magnitude of the current task queue depth compared to a preset queue depth reference value. The latency urgency factor is determined based on the urgency of the latency guarantee upper bound relative to a preset latency reference value. The queue pressure factor and latency urgency factor are then combined into a load penalty term. The queue pressure factor quantifies the degree of current task backlog and can be obtained by calculating the relative difference between the current task queue depth and the preset queue depth reference value. For example, it can be obtained by using (current task queue depth - preset queue depth reference value) / preset queue depth reference value or by using a non-linear function (such as the sigmoid function or exponential function) to map the relative magnitude to a specific range. This factor should increase significantly when the queue depth far exceeds the reference value. The latency urgency factor quantifies the time sensitivity of the task and can be obtained by calculating the relative difference between the preset latency reference value and the latency guarantee upper bound. For example, it can be obtained by using (preset latency reference value - latency guarantee upper bound) / preset latency reference value or by using a reverse mapping function. The smaller the latency guarantee upper bound (i.e., the more urgent the task), the larger this factor should be. The two factors are combined to form a load penalty term, which aims to comprehensively reflect the overall pressure faced by the local computing node and provide a more comprehensive basis for subsequent adjustments to the minimum utility threshold. The queue pressure factor and the latency urgency factor can be simply added together or weighted and summed. The weights can be adjusted according to actual business needs and experience to highlight the more critical factor between queue backlog and latency urgency.

[0129] The minimum utility threshold from the previous period is weighted and summed with a load penalty term. The weighted sum is then used as the minimum utility threshold for the current period, which is used for load-aware simplification decisions for the next batch of player behavior data units. This dynamic adjustment mechanism allows the minimum utility threshold to adaptively adjust based on real-time computational pressure and task latency requirements, thus more intelligently determining which data units should be retained, downsampled, or discarded. The minimum utility threshold can be updated iteratively or smoothly. For example, the current minimum utility threshold can be calculated as α (the minimum utility threshold from the previous period) + (1-α)(base threshold + load penalty term), where α is a smoothing factor (0 < α < 1), and the base threshold is a preset baseline value. Alternatively, more directly, the current minimum utility threshold = the minimum utility threshold from the previous period + a load penalty term adjustment coefficient. This update should be performed periodically, such as every second, every minute, or after each batch of data is processed. The updated threshold will be immediately applied to the simplification decisions for subsequent incoming player behavior data units.

[0130] By incorporating considerations of the current task queue depth and preset latency guarantee upper bound of local computing nodes, and transforming these into queue pressure factors and latency urgency factors, which are then superimposed to form a load penalty term, this application can more precisely and dynamically assess the actual load pressure of local computing nodes and the real-time requirements of tasks. Based on this, the load penalty term is weighted and superimposed with the minimum utility threshold of the previous period to generate the minimum utility threshold for the current period, enabling the data simplification strategy to adaptively respond to system load and latency requirements. When the system load is high or the task latency requirements are more urgent, the minimum utility threshold is increased, prompting the system to prioritize discarding data with lower utility, thereby effectively alleviating the processing pressure on local computing nodes, ensuring the timely completion of critical tasks, and avoiding the loss of important data due to blind simplification or system overload due to insufficient simplification. Conversely, when the system load is low, the minimum utility threshold is decreased, allowing more data to be retained, improving the accuracy and comprehensiveness of collaborative analysis. This mechanism ensures that the efficiency and effectiveness of data simplification achieve the best balance in resource-constrained and latency-sensitive scenarios.

[0131] In some embodiments described above, a load-aware simplification of player behavior data units is proposed based on content-sensitive attributes and the real-time load status of local computing nodes to obtain retained data units. Specifically, the current task queue depth of the local computing node and a preset upper bound for latency protection are determined and superimposed as a load penalty term to adjust the minimum utility threshold. However, in practical implementation, accurately quantifying queue pressure and latency urgency, and effectively integrating them to form a penalty term that accurately reflects the overall system load, thereby enabling dynamic and fine-grained adjustment of the minimum utility threshold, is a key challenge in ensuring the system can still operate efficiently and rationally select data units under different load conditions. A lack of a clear quantification mechanism may lead to inaccurate load assessment, thus affecting the effectiveness of the data simplification strategy.

[0132] In this regard, the present application further proposes the following specific implementations. The relative magnitude of the current task queue depth with respect to a preset queue depth reference value is taken as the queue pressure factor. For example, a local computing node continuously monitors its current task queue depth, and compares it with a preset queue depth reference value that represents an ideal or acceptable queue depth. The more the current task queue depth exceeds the preset queue depth reference value, the greater the immediate processing pressure faced by the local computing node, and the larger the value of the determined queue pressure factor at this time. This queue pressure factor intuitively quantifies the current task backlog degree of the system, and is an important indicator for measuring the immediate pressure of computing resources. For example, the queue pressure factor may be set to be proportional to the difference between the current task queue depth and the preset queue depth reference value, or mapped by means of a piecewise linear function, an exponential function or the like, so as to reflect different degrees of queue pressure more flexibly. In a specific implementation, the queue pressure factor P_queue is calculated by the following formula: P_queue=max(0,(Q_current-Q_ref) / Q_ref), where Q_current is the current task queue depth, and Q_ref is the preset queue depth reference value. The delay urgency factor P_delay is calculated by the following formula: P_delay=max(0,(T_ref-T_upper) / T_ref), where T_upper is the upper bound of delay guarantee, and T_ref is the preset delay reference value. The load penalty term P_load=alpha×P_queue+(1-alpha)×P_delay, where alpha is a weighting coefficient with a value range of [0,1] and a default value of 0.5. The minimum utility threshold theta_current of the current cycle=beta×theta_prev+(1-beta)×(theta_base+gamma×P_load), where theta_prev is the minimum utility threshold of the previous cycle, theta_base is the base utility threshold, beta is a smoothing factor (0<beta<1), and gamma is a penalty amplification coefficient.

[0133] The latency urgency factor is determined by showing a negative monotonically increasing correlation with the degree of looseness of the latency guarantee upper bound relative to the preset latency reference value. This means that the more stringent the latency requirements of the collaborative analysis task, i.e., the tighter the latency guarantee upper bound is relative to the preset latency reference value, the larger the determined latency urgency factor will be. For example, a function can be set such that the closer the latency guarantee upper bound is to or below the preset latency reference value, the larger the latency urgency factor is, and vice versa. This latency urgency factor reflects the task's time sensitivity requirements, ensuring that time-sensitive tasks are given priority in resource allocation and data simplification decisions. For example, a reciprocal relationship, a negative exponential relationship, or a piecewise function can be used to characterize this negative monotonically increasing correlation to ensure that the more urgent the latency requirements, the larger the factor value.

[0134] The queue pressure factor and latency urgency factor are weighted and summed, with the result used as a load penalty. By assigning different weights to the queue pressure factor and latency urgency factor, the system can flexibly adjust the emphasis on real-time processing pressure and time sensitivity according to the needs of actual application scenarios. For example, in scenarios with extremely high real-time requirements, the latency urgency factor can be given a higher weight. Conversely, in scenarios where computing resource bottlenecks are more prominent, the queue pressure factor can be given a higher weight. This load penalty comprehensively reflects the overall load situation faced by the local computing node; a higher value indicates a heavier system load.

[0135] The higher the value of this load penalty term, the higher the minimum utility threshold for the current period, and the more player behavior data units are downsampled or discarded. This means that when the system load is heavy, by increasing the minimum utility threshold, only player behavior data units with high content-sensitive attributes (i.e., utility scores) will be retained, while data units with low utility scores will be downsampled or discarded directly, thereby effectively reducing the system load.

[0136] Through the above technical solution, this application provides a more refined and dynamic load assessment mechanism. By clearly quantifying queue pressure factors and latency urgency factors, and weighting them to form a comprehensive load penalty term, the system can accurately capture the real-time processing pressure and time sensitivity requirements faced by local computing nodes. This precise load assessment makes the adjustment of the minimum utility threshold more reasonable and responsive. When the system load increases, the increase in the load penalty term can promptly raise the minimum utility threshold, prompting the system to prioritize retaining player behavior data units that contribute highly to the collaborative analysis conclusions and have high value, while downsampling or discarding data units with lower value, thereby effectively alleviating the pressure on computing resources and communication bandwidth and avoiding system overload. Conversely, when the system load is low, the minimum utility threshold will decrease accordingly, allowing more data to be retained to support more comprehensive and in-depth analysis. This mechanism not only optimizes the resource utilization efficiency of local computing nodes and ensures the processing priority of key data under resource constraints, but also effectively balances the relationship between data privacy protection and system performance while ensuring the quality of collaborative analysis conclusions, improving the robustness and adaptability of the entire game operation collaborative analysis method.

[0137] In some of the above implementations, the system performs load-aware simplification based on the content-sensitive attributes of player behavior data units and the real-time load status of local computing nodes to determine which data units to retain. However, when only a uniform minimum utility threshold is used for data simplification, it may not be possible to adequately distinguish the actual contribution of different data categories to the conclusions of game operation collaborative analysis. This could lead to some specific data categories that are crucial to the analysis being over-simplified, or data categories with low contribution to the analysis still occupying valuable computing resources, thereby affecting the accuracy and efficiency of collaborative analysis.

[0138] To address this, this application further proposes an optimization scheme: pre-setting a category bias for different data categories, which can be divided according to data source or event type. The minimum utility threshold is then superimposed with the corresponding category bias for each data category to obtain the effective utility threshold for that data category. Furthermore, the category bias is updated based on the game operation collaborative analysis conclusions. Specifically, the category bias for data categories whose contribution in the game operation collaborative analysis conclusions is lower than a pre-set standard is increased by a pre-set bias step size, making these data categories with lower contribution levels more easily downsampled or discarded in subsequent load-aware simplification.

[0139] For example, "data category" refers to the logical grouping of player behavior data units. This can be based on the data's source (e.g., client logs, user game data analysis and management system logs), event type (e.g., login events, purchase events, combat events, chat events), or player group characteristics (e.g., new player behavior, high-paying player behavior). "Category bias" is an adjustable numerical parameter associated with each data category, whose initial value can be set to zero or a preset small constant. This bias is used to adjust the utility threshold of a specific data category during load-aware slimming to reflect the specific importance or processing priority of that category of data in the overall analysis. For example, these category biases can be stored in the system's configuration database or a dynamically updated lookup table.

[0140] When calculating the effective utility threshold, the "minimum utility threshold" is a globally adjusted threshold dynamically based on the real-time load status of the local computing nodes, as described above. The "category bias" serves as a fine-tuning factor for this global minimum utility threshold. By superimposing the minimum utility threshold with the category bias of the corresponding data category, the system can generate a unique "effective utility threshold" for each data category. For example, if the category bias is positive, the effective utility threshold for that data category will be increased, requiring a higher utility score to retain the data in that category. If the category bias is negative, the effective utility threshold will be decreased, making it easier to retain the data in that category. This superposition mechanism allows the system to personalize the processing strategies for different data categories based on global load pressure.

[0141] To achieve adaptive optimization of the strategy, the system dynamically updates the "Category Bias" based on the "Game Operation Collaborative Analysis Conclusions." These conclusions are generated by fusing and analyzing anonymized intermediate representations with retained data units, such as player churn probability, abnormal behavior markers, or payment tendency scores. "Contribution" refers to the actual impact of a specific data category on the accuracy, insight, or model performance improvement of these analytical conclusions. The system continuously evaluates the "contribution" of different data categories to the analytical conclusions, using methods such as feature importance analysis, model sensitivity analysis, or A / B testing. When the contribution of a data category consistently falls below a "preset standard" (e.g., below a certain statistical significance level, information gain threshold, or minimum value standard set by the business), the system increases its corresponding "Category Bias" by a "preset bias step size." For example, if the contribution of chat log data to player churn prediction is consistently below the preset standard, its category bias will be increased. This upward adjustment directly raises the "effective utility threshold" for that data category in subsequent load-aware simplification, making it easier for the system to classify that category of data as low-value and downsample or discard it when facing computational resource constraints. This dynamic update mechanism forms a feedback loop, enabling the data simplification strategy to adaptively adjust based on actual analysis results.

[0142] The above technical solution overcomes the limitations of a single global minimum utility threshold when processing heterogeneous data categories. By introducing and dynamically adjusting category bias, the system can intelligently adjust its data simplification strategy based on the actual contribution of different data categories to the collaborative analysis conclusions of game operations. For example, for data categories that contribute less to the analysis conclusions, their effective utility threshold is dynamically increased, making them easier to downsample or discard when resources are limited, thereby effectively reducing unnecessary data processing burdens and optimizing the resource utilization efficiency of local computing nodes. For data categories that may have low utility scores in certain scenarios but are potentially important to the overall analysis, their category bias can be adjusted to ensure they are properly preserved, avoiding the accidental loss of valuable information. This adaptive, category-aware load simplification mechanism allows the collaborative analysis conclusions of game operations to focus more on high-value information, improving the accuracy and reliability of the analysis, and ensuring that the system can continuously produce high-quality operational insights under limited computing resources.

[0143] While some of the above implementations propose fusing and analyzing the cross-source association identifiers carried by the de-identified intermediate representations and retained data units to generate collaborative analysis conclusions for game operations, in practical applications, game data often comes from diverse sources, has different formats, and may have complex relationships between different data sources. How to efficiently and accurately integrate these heterogeneous data and extract specific analytical conclusions that are instructive for game operations is a key challenge in achieving collaborative analysis.

[0144] To address this, this application further proposes a method for fusing and analyzing the cross-source association identifiers carried by the anonymized intermediate representations and retained data units. Specifically, this method includes: aligning the anonymized intermediate representations from different data sources according to the cross-source association identifiers, which include player identifiers, match identifiers, or time window identifiers. Then, performing splicing and attention-weighted aggregation on the aligned multi-source anonymized intermediate representations in a shared feature space to obtain cross-source fusion features. A collaborative analysis model then performs classification inference based on these cross-source fusion features to obtain at least one analytical conclusion. This at least one analytical conclusion is used as the collaborative analysis conclusion for game operation, and the analytical conclusion includes at least one of player churn probability, abnormal behavior markers, or payment tendency scores.

[0145] For example, anonymized intermediate representations from different data sources are aligned using cross-source association identifiers, such as player identifiers, match identifiers, or time window identifiers. This step aims to address the primary challenge in multi-source heterogeneous data fusion—data alignment. In game operation scenarios, player behavior data may originate from multiple data sources, including client logs, user game data analysis and management system logs, payment systems, and social platforms. Events recorded by these data sources may involve the same player, the same match, or the same time period. For meaningful collaborative analysis, these disparate data sets must be linked. Cross-source association identifiers, such as player identifiers (e.g., user ID, OpenID), match identifiers (e.g., room ID, game ID), and time window identifiers (e.g., timestamps of events, session IDs), are common association keys. The alignment process can be achieved through database JOIN operations, window aggregation in streaming data processing, or entity links based on graph databases. For instance, the system can maintain a global player ID mapping table to unify the unique identifiers of players from different data sources. When receiving anonymized intermediate representations from different data sources, the system extracts the player identifier, match identifier, or time window identifier carried within them. Then, in a unified data storage or processing layer, it logically associates data fragments belonging to the same entity or event based on these identifiers. This ensures that subsequent analysis can be conducted within a complete context.

[0146] Building upon this, the aligned multi-source desensitized intermediate representations are concatenated and attention-weighted aggregated in a shared feature space to obtain cross-source fusion features. The collaborative analysis model then performs classification inference based on these cross-source fusion features to arrive at at least one analytical conclusion. This step is crucial for achieving deep fusion and advanced analysis. After data alignment, desensitized intermediate representations from different data sources may have different dimensions and semantics. To enable the collaborative analysis model to process them uniformly, they need to be mapped to a shared feature space and effectively aggregated. The shared feature space can be transformed into a unified vector representation using pre-trained embedding models (such as Word2Vec and Transformer encoders). Concatenation and attention-weighted aggregation are key to feature fusion. Concatenation is the most direct method, directly connecting the feature vectors of multiple aligned desensitized intermediate representations to form a longer feature vector. Considering that features from different data sources or at different time points may contribute differently to the analytical conclusions, an attention mechanism can be introduced for weighted aggregation. For example, self-attention or cross-attention mechanisms can be used to dynamically assign weights to each data source or feature dimension based on the importance of the current analytical task. Features with higher weights will have a greater impact on the fused features. Once formed, the fused features can be input into various machine learning or deep learning models for classification inference. For example, these could be multilayer perceptrons (MLPs), recurrent neural networks (RNNs), convolutional neural networks (CNNs), or more complex graph neural networks (GNNs), depending on the nature of the analysis task. The model will learn the complex relationships between these fused features and the predefined analysis objective, and output classification results or predicted values.

[0147] At least one analytical conclusion is used as the collaborative analysis conclusion for game operations. This conclusion includes at least one of the following: player churn probability, abnormal behavior markers, or payment tendency score. This step clarifies the output of the collaborative analysis and its application value. Through the aforementioned deep integration and model reasoning, the system can generate conclusions that have direct guiding significance for game operations. For example, player churn probability refers to the model's ability to predict the likelihood of players churning in the future, which helps the operations team identify high-risk players in advance and take retention measures. Abnormal behavior markers refer to the model's ability to identify activities that significantly deviate from normal player behavior patterns, such as cheating, malicious ranking manipulation, and account theft, which is crucial for maintaining game fairness and security. Payment tendency score refers to the model's ability to assess the likelihood and potential amount of money a player is willing to spend, which helps the operations team conduct precise marketing and personalized recommendations. These analytical conclusions are usually output in the form of probability values, binary labels, or continuous scores, and can be directly used to trigger operational strategies, such as sending retention emails, banning accounts, or pushing exclusive gift packs.

[0148] By precisely aligning anonymized intermediate representations from different data sources according to cross-source association identifiers such as player identifiers, match identifiers, or time window identifiers, this application effectively solves the challenge of integrating heterogeneous game data, ensuring data consistency in semantic and temporal dimensions. Based on this, performing splicing and attention-weighted aggregation in a shared feature space fully leverages the complementarity of multi-source information and dynamically captures the key contributions of different data sources to the analysis task through an attention mechanism, thereby generating more representative and discriminative cross-source fusion features. The collaborative analysis model performs classification inference based on these high-quality fusion features, accurately outputting specific and actionable analytical conclusions such as player churn probability, abnormal behavior markers, or payment tendency scores. This enables game operators to efficiently extract insights directly guiding business decisions from massive amounts of data, improving the refinement and intelligence of game operations and effectively solving the challenge of extracting specific and guiding analytical conclusions from complex multi-source data.

[0149] In some embodiments described above, game operation collaborative analysis conclusions can be generated by fusing and analyzing the desensitized intermediate representation with the cross-source association identifier carried by the retained data units. However, in actual operation, due to changes in player behavior patterns, game environment, or data distribution, the collaborative analysis model may experience performance degradation or reduced prediction accuracy, i.e., model drift. If this model drift is not detected and corrected in time, it will reduce the reliability of subsequent data processing and analysis results, affecting the effectiveness of game operation decisions.

[0150] To address this, this application further proposes a method that includes monitoring the output distribution changes of the aforementioned collaborative analysis model within a preset time window to obtain the model drift amount. This model drift amount is then compared with a preset drift threshold. When the model drift amount exceeds the preset drift threshold, the grading rules for content-sensitive attributes in the online semantic parsing are updated. Based on the direction of the model drift, the weight of the fusion dimension corresponding to that drift direction is increased, while the weights of other fusion dimensions are decreased, causing the content-sensitive attribute values ​​of the same player behavior data units to change in the next period.

[0151] For example, to monitor changes in the output distribution of the aforementioned collaborative analysis model within a preset time window and obtain the model drift, various statistical or machine learning methods can be employed. For instance, the system can periodically (e.g., hourly, daily, or weekly) collect statistical characteristics of the analytical conclusions (e.g., player churn probability, abnormal behavior markers, or payment propensity scores) output by the collaborative analysis model within the preset time window, such as mean, variance, quantiles, and entropy. These statistical characteristics are then compared with historical baseline data or the statistical characteristics of the previous time window to quantify the degree of change in the output distribution. Commonly used drift detection metrics include Kullback-Leibler divergence (KL divergence), Jensen-Shannon divergence (JS divergence), Population Stability Index (PSI), or Characteristic Stability Index (CSI). These metrics provide a quantifiable measure of model drift, reflecting the difference between the current model output distribution and the reference distribution. In one specific implementation, the model drift is calculated using the Population Stability Index (PSI): PSI = sum[(P_actual_i - P_expected_i) × ln(P_actual_i / P_expected_i)], where P_expected_i is the sample proportion of the i-th bin in the historical baseline data, and P_actual_i is the sample proportion of the i-th bin within the current preset time window. Typically, the output value range is divided into 10 equal bins. When the PSI value exceeds 0.25, it is considered a significant model drift, triggering a grading rule update.

[0152] After obtaining the model drift amount, the system compares it with a preset drift threshold. This preset drift threshold is a critical value used to determine whether the model drift has reached a point where intervention is necessary. This threshold can be set based on historical data, business requirements, tolerance for model performance metrics (such as accuracy, recall, F1 score, etc.), or expert experience. For example, when the calculated KL divergence exceeds a certain empirical value, or when the model's key performance indicators on the real-time validation set decrease by more than a preset percentage, it can be determined that the model drift amount has exceeded the preset drift threshold.

[0153] When the model drift exceeds a preset drift threshold, the system triggers an update to the grading rules for content-sensitive attributes in the online semantic parsing. Content-sensitive attributes are typically obtained by weighted fusion of multiple dimensions, including event rarity, behavioral impact, real-time requirement components, and pattern completion contribution. To adjust based on the direction of model drift, the system needs to determine the specific direction of the drift. This can be achieved by analyzing performance changes across different data subsets, prediction bias for specific output categories, or identifying key input feature dimensions causing the drift through feature importance analysis. For example, if the model's accuracy in predicting churn for a specific player group (such as new or highly active players) decreases, or its recognition rate for a specific type of abnormal behavior decreases, the drift direction can be determined to be related to these specific data dimensions or behavioral patterns. Once the drift direction is determined, the system will correspondingly increase the weight of the fusion dimension corresponding to that drift direction and decrease the weight of the remaining fusion dimensions. For example, if drift analysis indicates that the model is not sensitive enough to "event rarity," leading to the underestimation of rare but important events, the system will increase the weight of event rarity in the weighted fusion of content-sensitive attributes. Conversely, if a dimension (such as "pattern completion contribution") contributes less to model drift, or its importance decreases under the new data distribution, its weight can be reduced. This dynamic adjustment aims to make the calculation of content-sensitive attributes more reflective of the current data environment and the needs of collaborative analysis tasks. In this way, even for the same player behavior data unit, the calculation results of its content-sensitive attributes will change in the next cycle due to the change in the weight of the fusion dimension, thus affecting the determination of subsequent desensitization depth parameters and processing priority labels.

[0154] By introducing a monitoring mechanism for changes in the output distribution of the collaborative analysis model, this application can promptly detect model drift. When the model drift exceeds a preset threshold, the system can dynamically adjust the grading rules of content-sensitive attributes in online semantic parsing based on the specific direction of the drift, particularly by adaptively increasing or decreasing the weights of each fusion dimension constituting the content-sensitive attributes. This dynamic adjustment ensures that the calculation of content-sensitive attributes can continuously and accurately reflect the true value of player behavior data units and their impact on collaborative analysis tasks, thereby avoiding inaccurate analysis conclusions due to model drift. Furthermore, this makes the subsequent determination of desensitization depth parameters and processing priority labels more precise, ensuring a balance between data privacy protection and analytical utility, optimizing the allocation of computing resources, and improving the robustness and accuracy of the entire game operation collaborative analysis method.

[0155] In some embodiments described above, online semantic parsing, joint mapping, hierarchical feature extraction, and load-aware simplification are performed on incoming player behavior data units, followed by fusion analysis to generate game operation collaborative analysis conclusions. However, in actual game operation environments, player behavior patterns, data sensitivity, and the focus of collaborative analysis tasks may dynamically change over time. If the grading rules used to determine content sensitivity attributes, the mapping relationship of desensitization depth parameters, and the criteria for determining processing priority labels remain static, the data processing strategy may fail to adapt to new business needs or risk situations, thereby affecting the accuracy of collaborative analysis conclusions, the effectiveness of data privacy protection, and the optimal utilization of computing resources.

[0156] To address this, this application proposes an adaptive optimization mechanism, which further includes: extracting feedback signals from the game operation collaborative analysis conclusions, wherein the feedback signals include at least one of the following: changes in the risk level of a specific data category, statistical contribution of a specific data category to the game operation collaborative analysis conclusions, or prediction bias of the collaborative analysis model. Based on the feedback signals, the grading rules used to determine content-sensitive attributes in the online semantic parsing are updated, increasing the baseline of content-sensitive attributes corresponding to data categories with increased risk levels by a preset step size, and decreasing the baseline of content-sensitive attributes corresponding to data categories with contribution levels below a preset contribution threshold by the preset step size. Based on the updated grading rules, the mapping relationship of the desensitization depth parameter and the basis for determining the processing priority label are adjusted, so that the same player behavior data units in the next cycle obtain different cutoff layer positions and different computational resource contention weights.

[0157] For example, extracting feedback signals from the conclusions of collaborative game operation analysis is a key input for the system to self-adjust and optimize. These feedback signals aim to objectively reflect the effectiveness and potential problems of the current data processing strategy in practical applications. For instance, changes in the risk level of a specific data category can be determined by continuously monitoring the frequency or severity of that data category being marked as anomalies, fraud, or high-risk events in the collaborative analysis conclusions. If a certain type of transaction data is frequently identified as anomaly, its risk level can be considered to have increased. The contribution of a specific data category to the conclusions of collaborative game operation analysis can be quantified by calculating the marginal contribution or information gain of different data categories to the analysis conclusions (such as player churn prediction, payment tendency rating, etc.). For example, interpretable artificial intelligence methods can be used to evaluate the impact of each data feature on the model's prediction results. The prediction bias of the collaborative analysis model can be measured by comparing the differences between the model's prediction results and the actual observation results. For example, for player churn prediction, the degree of match between the model's predicted churn players and the actual churn players can be compared. These feedback signals can be periodically obtained from the collaborative analysis platform, aggregated, and quantified to provide a basis for subsequent strategy adjustments.

[0158] Updating the grading rules used to determine content-sensitive attributes in online semantic parsing based on this feedback signal is the core mechanism for achieving adaptive adjustment of the system. The grading rules for content-sensitive attributes determine the importance of data units in subsequent desensitization and prioritization. When calculating content-sensitive attributes, there is usually a base value or weight used to measure the inherent importance of a specific data category; this can be considered the baseline for content-sensitive attributes. When the feedback signal indicates that the risk level of a specific data category is increasing, it means that this type of data has a greater potential impact or risk on the analysis conclusions, thus requiring an increase in its content-sensitive attribute to achieve deeper desensitization or higher processing priority in subsequent processing. This can be achieved by increasing its weight in the content-sensitive attribute calculation or directly raising its baseline value, for example, by increasing the preset step size. Conversely, when the feedback signal indicates that the contribution of a specific data category is consistently below a preset contribution threshold, it means that this type of data has lower value for the current collaborative analysis task, and its content-sensitive attribute can be appropriately reduced, thus prioritizing more valuable data when resources are limited. This can be achieved by decreasing its weight in the content-sensitive attribute calculation or lowering its baseline value, for example, by lowering the preset step size. These adjustments can be linear or non-linear, and the specific step size and adjustment function can be optimized based on experience or through reinforcement learning.

[0159] Adjusting the mapping relationship of the anonymization depth parameters and the criteria for determining processing priority labels, based on the updated grading rules, is a key step in transforming the dynamic adjustment of content-sensitive attributes into specific data processing strategies. The mapping relationship of the anonymization depth parameters, such as mapping the joint indicator to the cutoff layer depth range, as well as the adjustable parameters in the nonlinear transformation and the upper and lower limits of the cutoff layer depth range, can all be dynamically adjusted. When the grading rules for content-sensitive attributes are updated, the new content-sensitive attribute values ​​will affect the joint indicator, and thus the anonymization depth parameters. The parameters of the mapping function itself can also be adjusted. For example, if the baseline of the content-sensitive attribute for a data category with increased risk is raised, it may be necessary to adjust the amplification factor of the nonlinear transformation so that even a slight increase in the content-sensitive attribute can lead to a significant increase in the anonymization depth parameters, thereby obtaining a deeper cutoff layer position. The criteria for determining processing priority labels, such as comparing the content-sensitive attribute with a preset sensitivity threshold, can also be adjusted. When the grading rules for content-sensitive attributes are updated, the content-sensitive attribute values ​​of the same player behavior data unit may change, potentially crossing the preset sensitivity threshold, causing its priority label to change from low priority to high priority, and vice versa. The preset sensitivity threshold itself can also be used as an adjustable parameter, dynamically adjusted based on feedback signals. These adjustments ensure the flexibility and adaptability of the data processing strategy, enabling the system to self-optimize based on actual operating conditions. This allows the same player behavior data unit in the next cycle to obtain a cutoff layer position and computational resource contention weight that matches its current value and risk.

[0160] The aforementioned technical solution enables the extraction of multi-dimensional feedback signals from game operation collaborative analysis conclusions. These signals objectively reflect the effectiveness and potential problems of the current data processing strategy. Based on these signals, the system can dynamically update the grading rules used in online semantic parsing to determine content sensitivity attributes. For example, it can raise the baseline of content sensitivity attributes for data categories with increased risk levels, while lowering the baseline for data categories with lower contribution. This adaptive rule adjustment mechanism allows the system to adjust the value assessment and risk consideration of data units in real time according to the dynamic changes in the game operation environment. Furthermore, based on the updated grading rules, the system can dynamically adjust the mapping relationship of desensitization depth parameters and the basis for determining processing priority labels, ensuring that the same player behavior data unit obtains a cutoff layer position and computational resource contention weight that matches its current value and risk in different periods. This effectively solves the problem that static rules cannot adapt to dynamic environments, improves the accuracy of data privacy protection, the accuracy of collaborative analysis conclusions, and the efficiency of computational resource utilization, and achieves continuous optimization and adaptation of data processing strategies.

[0161] In some of the embodiments described above in this application, feedback signals are extracted from the conclusions of collaborative analysis of game operations. These feedback signals are then used to update the grading rules for determining content sensitivity attributes in online semantic parsing, as well as to adjust the mapping relationship of desensitization depth parameters and the basis for determining processing priority tags. However, in actual implementation, how to specifically and effectively identify and extract instructive feedback signals, such as changes in risk level or data contribution statistics, from complex collaborative analysis conclusions to ensure the accuracy and effectiveness of subsequent parameter adjustments is a technical detail that needs clarification.

[0162] To address this, this application further proposes a specific method for extracting feedback signals from the conclusions of game operation collaborative analysis. This method includes: extracting data category identifiers marked as high-risk or abnormal from the game operation collaborative analysis conclusions, and using these data category identifiers as the carrier of changes in risk levels; statistically analyzing the marginal contribution of various data types to the game operation collaborative analysis conclusions within a preset evaluation period, and using these marginal contribution values ​​as the carrier of contribution statistics; and combining the data category identifiers with the marginal contribution values ​​to form the feedback signal, which is used to drive the update of the grading rules and the adjustment of the desensitization depth parameter mapping relationship.

[0163] For example, a "game operation collaborative analysis conclusion" is a comprehensive judgment derived from in-depth analysis of player behavior data. It takes various forms, including player churn probability, abnormal behavior markers, and payment tendency scores. To extract feedback on changes in risk levels, this application proposes identifying and extracting "data category identifiers" marked as high-risk or abnormal from these conclusions. These "data category identifiers" can be unique identifiers for specific event types that trigger high-risk or abnormal conclusions (e.g., large-scale recharge followed by immediate withdrawal, login from a different location, use of cheats, etc.), player groups (e.g., newly registered users, high-value users, users with churn tendency), or data sources (e.g., data from specific channels, third-party plugin data). Through a preset risk assessment model or rules, when the collaborative analysis conclusion indicates that a certain data category has reached or exceeded a preset risk threshold, the identifier of that data category is extracted as a concrete carrier of risk level changes, thus concretizing abstract risk changes into actionable identifier information.

[0164] To quantify the importance of different data categories to the analytical conclusions, this application further proposes to statistically measure the "marginal contribution" of each data category to the "game operation collaborative analysis conclusions" within a "preset evaluation period." This "preset evaluation period" can be a fixed time window, such as daily, weekly, or after each batch of analysis tasks is completed. The "marginal contribution" refers to the incremental impact or importance of a specific data category on the collaborative analysis conclusions. For example, it can be calculated using model interpretability techniques (such as Shapley values, LIME, etc.) or feature importance assessment methods. For instance, it can assess the degree of change in the collaborative analysis model's predicted results (such as player churn probability) when data from a certain data category is included or excluded, or the average weight of features in that data category in the model's decision path. Statistically aggregating these quantitative results yields the "marginal contribution" of that data category within a specific period, which, as a carrier of "contribution statistics," intuitively reflects the value of that data category to the overall analysis task.

[0165] To establish a complete and operational feedback mechanism, this application effectively combines the extracted "data category identifiers" and the statistically obtained "marginal contribution" to form a structured "feedback signal." This "feedback signal" can be a data structure containing data category, risk status (e.g., "high risk," "abnormal"), and contribution value. For example, a feedback signal might represent "Data category: remote login, risk status: high risk, marginal contribution: 0.2." This structured "feedback signal" is used to drive the dynamic adjustment of internal system parameters. Specifically, it can be used for "updating hierarchical rules," i.e., adjusting the weight or threshold of content-sensitive attributes determined in online semantic parsing. It is also used for "adjusting the mapping relationship of desensitization depth parameters," i.e., changing the functional relationship between content-sensitive attributes and privacy-sensitive attributes jointly mapped to desensitization depth parameters. This combined feedback signal ensures that the system can make refined and adaptive strategy adjustments based on specific data risks and data value, thereby optimizing the data processing flow.

[0166] Through the aforementioned technical solution, this application clarifies a specific method for extracting feedback signals from complex game operation collaborative analysis conclusions, resolving the technical problem of unclear feedback signal extraction. By using high-risk or abnormal "data category identifiers" as the carrier of risk level changes, and using the "marginal contribution" of various data types as the carrier of contribution statistics, the system can obtain specific, quantifiable, and instructive feedback information. This structured feedback signal makes subsequent adjustments to the content sensitivity attribute grading rules and the mapping relationship of desensitization depth parameters more accurate and effective. For example, when a data category is identified as high-risk, the system can specifically raise its content sensitivity attribute baseline, thereby prompting it to achieve a deeper desensitization depth in subsequent processing and enhancing privacy protection. For data categories with low marginal contribution, the system can appropriately lower its content sensitivity attribute baseline, optimize resource allocation, and avoid unnecessary computational overhead. This dynamic, data-driven feedback mechanism improves the system's adaptability and overall performance in balancing data privacy, data utility, and computational resources.

[0167] In some embodiments described above in this application, while it is proposed to update the grading rules used to determine content-sensitive attributes in online semantic parsing based on feedback signals, such as by adjusting the baseline of content-sensitive attributes corresponding to specific data categories upwards or downwards, this simple baseline adjustment may not fully reflect the profound impact of different feedback signals on the calculation of content-sensitive attributes. Especially when content-sensitive attributes are obtained through weighted fusion of multiple dimensions, directly adjusting the baseline may lack sufficient flexibility and granularity, making it difficult to effectively guide subsequent desensitization depth and resource allocation strategies, thereby affecting the accuracy and efficiency of collaborative analysis.

[0168] To address this, this application further proposes increasing the fusion weight of the content sensitivity attribute baseline corresponding to a specific data category in the weighted fusion when the feedback signal indicates an increased risk level for that specific data category. Conversely, decreasing the fusion weight of the content sensitivity attribute baseline corresponding to that specific data category in the weighted fusion when the feedback signal indicates that the marginal contribution of a specific data category remains below a preset contribution threshold. The updated fusion weight is then applied to the content sensitivity attribute calculation in the next cycle, enabling data categories with increased risk levels to obtain higher content sensitivity attribute values ​​in the next cycle, thereby achieving a deeper cutoff layer position.

[0169] For example, when feedback signals indicate an increased risk level for a specific data category, it means that information from the game operation collaborative analysis suggests that certain types of data (e.g., data involving sensitive user actions, high-value transactions, or abnormal behavior) pose a higher risk in terms of privacy exposure, security risks, or potential misleading of analysis results. To address this risk, this application enhances the influence of this specific data category on the content sensitivity attribute value by increasing its fusion weight in the weighted fusion calculation of content sensitivity attributes. Content sensitivity attributes are the result of a weighted fusion of multiple components (such as event rarity, behavioral impact, real-time requirement component, and pattern completion contribution). Increasing the fusion weight of a data category means that that data category is given greater importance when calculating content sensitivity attributes.

[0170] Correspondingly, when the feedback signal indicates that the marginal contribution of a specific data category is consistently below a preset contribution threshold, this suggests that certain data categories provide little incremental value or information for generating collaborative analysis conclusions on game operations, or that their positive impact on the analysis results is insignificant. To optimize resource utilization and avoid unnecessary processing of low-value data, this application reduces the fusion weight of that specific data category in the weighted fusion calculation of content-sensitive attributes. Reducing the fusion weight decreases the impact of that data category on the value of content-sensitive attributes, resulting in a relatively lower value for those attributes.

[0171] The aforementioned adjustments to the fusion weights do not take effect immediately; instead, the updated fusion weights are applied to the content-sensitive attribute calculation in the next cycle. This periodic update mechanism ensures the stability and controllability of the system strategy. In this way, data categories with higher risk levels will receive higher content-sensitive attribute values ​​in the next cycle. According to the aforementioned joint mapping mechanism, higher content-sensitive attribute values ​​will lead to deeper desensitization depth parameters, thereby enabling the local feature extraction network to perform deeper truncation when processing such data. For example, if the local feature extraction network is a multi-layer neural network, deeper truncation layers mean that more forward computation layers are performed on the local computing nodes, resulting in a more abstract and harder-to-reverse-reconstruct desensitized intermediate representation, effectively enhancing the privacy protection of high-risk data.

[0172] By dynamically adjusting the fusion weights in the weighted fusion of content-sensitive attributes, this application can respond more precisely to feedback signals from game operation collaborative analysis conclusions. When the risk level of a specific data category increases, its weight in the calculation of content-sensitive attributes is increased, resulting in a higher content-sensitive attribute value for that data category. This leads to a deeper desensitization depth parameter obtained through the joint mapping mechanism, enabling deeper truncation in the local feature extraction network and effectively enhancing the privacy protection of high-risk data. Conversely, when the marginal contribution of a specific data category remains below a preset threshold, its fusion weight is reduced, lowering its content-sensitive attribute value. This may result in a shallower truncation layer or easier downsampling or discarding during load-aware simplification, thereby optimizing the allocation of computing resources and avoiding over-processing of low-value data. This feedback mechanism based on fusion weight adjustment makes the calculation of content-sensitive attributes more flexible and adaptive, more accurately balancing privacy protection, data value, and computational efficiency, and improving the intelligence and robustness of the entire user game data analysis and management method.

[0173] In some of the embodiments described above in this application, although the grading rules for determining content-sensitive attributes in online semantic parsing are updated based on feedback signals to adapt to changes in the conclusions of game operation collaborative analysis, in practical applications, there are multiple interrelated strategy parameters within the system, such as the mapping relationship of the desensitization depth parameter, the utility threshold for load-aware simplification, and the computational resource weighting rules for processing priority tags. If these parameters are adjusted independently, it may lead to a lack of synergy between strategies, making it impossible to optimize the overall system performance, and may even cause new performance bottlenecks or resource waste.

[0174] To address this, this application further proposes a more comprehensive strategy management mechanism, which includes: combining the hierarchical rules, the mapping relationship of the anonymization depth parameter, the utility threshold used in the load-aware simplification process to determine whether to retain or discard the player behavior data unit, and the computational resource weight rule of the processing priority label into a system strategy parameter group. The feedback signal is mapped to a strategy update increment, and this update increment is applied to the system strategy parameter group to obtain an updated system strategy parameter group. The updated system strategy parameter group replaces the previous cycle's system strategy parameter group and is applied to the player behavior data unit processing in the next cycle, ensuring that the hierarchical rules, the mapping relationship of the anonymization depth parameter, the utility threshold, and the computational resource weight rule work synergistically in the next cycle.

[0175] For example, the grading rules refer to the rules used to determine content-sensitive attributes. For instance, in the above implementation, the baseline of the content-sensitive attributes or the weights in the weighted fusion are adjusted based on changes in risk level or contribution statistics. These rules can be judgment logic based on preset thresholds, a set of coefficients for weighted fusion, or a more complex parameterized model. The mapping relationship of the desensitization depth parameter refers to the function or lookup table that maps the joint indicator (obtained by multiplicatively fusing content-sensitive attributes and privacy-sensitive attributes) to a preset cutoff layer depth range. This mapping relationship can use linear functions, nonlinear transformations (such as power functions, exponential functions, etc.), and its internal adjustable parameters (such as the amplification factor of the nonlinear transformation, the upper and lower limits of the cutoff layer depth range) jointly define the mapping relationship. The utility threshold used in the load-aware simplification process to determine whether a player behavior data unit should be retained or discarded refers to the critical value used to determine whether a player behavior data unit should be retained or discarded during load-aware simplification, such as the minimum utility threshold and the category bias for different data categories. These thresholds directly affect the strength of data simplification. The computational resource weighting rules for priority labels refer to the rules used to differentiate the computational resource allocation weights or scheduling order that de-identified intermediate representations of different priority levels should receive when scheduling them in collaborative computing nodes. For example, higher-priority de-identified intermediate representations may be assigned higher scheduling weights, thus gaining priority access to computational resources. Integrating all the aforementioned interrelated policy parameters into a system policy parameter group, forming a unified policy management unit, helps to achieve coordinated control and optimization of the entire system's behavior.

[0176] The feedback signal, as described in the above implementation, may include changes in the risk level of a specific data category, statistical analysis of the contribution of a specific data category to the collaborative analysis conclusions of game operations, or prediction biases in the collaborative analysis model. Mapping this feedback signal to a strategy update increment means that there exists a mechanism or algorithm capable of transforming this high-level feedback information into specific, operable parameter adjustment instructions, i.e., a strategy update increment. For example, when the feedback signal indicates an increased privacy risk for a certain data category, the strategy update increment may indicate the need to increase the baseline of the content sensitivity attribute for that data category and simultaneously adjust the mapping relationship of the anonymization depth parameters to correspond to a deeper cutoff layer. Applying this strategy update increment to the system strategy parameter set means applying the calculated increment to the corresponding parameters in the system strategy parameter set, thereby changing the current values ​​of these parameters. For example, the updated system strategy parameter set may include adjusted grading rule weights, new anonymization depth mapping function parameters, and revised utility thresholds.

[0177] By replacing the previous system strategy parameter set with the updated set and applying it to the player behavior data unit processing in the next cycle, this application ensures the dynamic adaptability of the system strategy. This holistic strategy management and update mechanism allows the mapping relationship between hierarchical rules, desensitization depth parameters, utility thresholds, and computational resource weighting rules to take effect collaboratively in the next cycle. This means that when game operation collaborative analysis conclusions indicate that the system needs adjustment, such as facing higher privacy risks or stricter real-time requirements, the system strategy parameter set can respond in a coordinated manner as a whole. This allows the determination of content-sensitive attributes, the depth of data desensitization, the intensity of data simplification, and the allocation of computational resources to work together to adapt to the new operating environment and business needs. In this way, this application overcomes the strategy conflicts and local optima problems that may result from independent adjustment of single parameters, ensuring that the system can dynamically balance data privacy protection, data analysis value, and computational resource consumption when processing player behavior data units, thereby achieving global optimization and continuous adaptability of system performance.

[0178] In some embodiments of this application, the system policy parameter group (including hierarchical rules, mapping relationships of desensitization depth parameters, utility thresholds used in load-aware simplification to determine whether to retain or discard player behavior data units, and computational resource weight rules for processing priority labels) is updated by mapping feedback signals to policy update increments. However, in practical applications, simply mapping feedback signals to policy update increments may not fully balance multiple objectives such as system performance, resource consumption, and privacy protection. This could lead to policy updates failing to achieve optimal results, or even negatively impacting other key objectives while optimizing one, such as excessively consuming computational resources to improve analysis accuracy or increasing privacy exposure risks.

[0179] To address this, this application further proposes a method for mapping the feedback signal into a policy update increment. This method includes: constructing a comprehensive evaluation index comprising a latency cost component, a communication bandwidth cost component, and a privacy exposure risk cost component. The latency cost component is determined based on the degree to which the end-to-end latency of the key collaborative analysis task exceeds a preset latency upper limit. The communication bandwidth cost component is determined based on the uplink data transmission volume. The privacy exposure risk cost component is determined based on the reversibility of the anonymized intermediate representation. Under the condition that the latency cost component does not exceed a preset latency constraint, the policy parameter update amount that optimizes the comprehensive evaluation index is determined. This policy parameter update amount is used as the policy update increment, and it simultaneously affects the mapping relationship between the hierarchical rule, the anonymization depth parameter, and the utility threshold.

[0180] For example, constructing comprehensive evaluation metrics aims to quantify the system's overall performance in terms of performance, resources, and privacy. The latency cost component measures the extent to which the system's end-to-end latency exceeds a preset latency limit when processing critical collaborative analysis tasks. For instance, a penalty function can be set so that when the actual end-to-end latency exceeds the preset limit, the value of this component increases sharply with the degree of exceedance, reflecting the severity of the violation of real-time requirements. The communication bandwidth cost component assesses the network resource consumption of data transmission. It can be determined based on the amount of data transmitted uplink from the local computing node (e.g., the total size of the anonymized intermediate representation and retained data units), or compared to a preset bandwidth capacity. When the transmission volume approaches or exceeds the capacity, the value of this component increases accordingly. The privacy exposure risk cost component assesses the likelihood and extent to which the anonymized intermediate representation can be reverse engineered to recover the original sensitive information. This can be quantified by simulating the ability of an attacker to infer identifiable fields in the original player behavior data units from the anonymized intermediate representation using reverse engineering techniques. For example, the recovery rate of identifiable fields can be calculated; the higher the recovery rate, the larger the privacy exposure risk cost component.

[0181] Based on this, and under the condition that the delay cost component does not exceed a preset delay constraint, the policy parameter update amount that optimizes the comprehensive evaluation index is determined. This typically involves a multi-objective optimization problem, aiming to find a set of policy parameter update amounts that minimize the comprehensive evaluation index. For example, optimization algorithms such as gradient descent, genetic algorithms, or reinforcement learning can be used to search for the optimal solution. The delay cost component not exceeding the preset delay constraint can be treated as a hard constraint, handled by the Lagrange multiplier method, penalty function method, or by excluding solutions that do not meet the constraint in the search space. The output of the optimization algorithm is the specific adjustment amount for the mapping relationship between the hierarchical rule, the desensitization depth parameter, and the utility threshold.

[0182] The update amount of this strategy parameter is used as the strategy update increment, and it is applied simultaneously to the hierarchical rule, the mapping relationship of the desensitization depth parameter, and the utility threshold. This means that after receiving this set of updates, the system will synchronously apply them to the corresponding strategy parameters, ensuring that the updates of the strategy parameters are coordinated and holistic, rather than isolated. For example, if the optimization result indicates that a certain weight of the hierarchical rule needs to be increased, a certain parameter of the desensitization depth parameter mapping relationship needs to be adjusted, and the utility threshold also needs to be increased, the system will execute these three adjustments simultaneously to achieve overall optimization of the system behavior.

[0183] The above technical solutions effectively balance the potential conflicts between data processing latency, communication bandwidth consumption, and user privacy protection, avoiding the negative impacts of single-objective optimization. The introduction of comprehensive evaluation indicators allows the system to fully weigh performance, resources, and privacy during policy adjustments, leading to more informed decisions. The setting of latency cost components and latency constraints ensures that the real-time requirements of key collaborative analysis tasks are met, avoiding task delays caused by policy adjustments. Consideration of communication bandwidth cost components enables the system to more intelligently control data transmission volume and optimize network resource utilization. The introduction of privacy exposure risk cost components allows the system to proactively assess and control the risk of privacy leakage during policy adjustments, providing stronger privacy protection. The policy parameter update simultaneously affects the mapping relationship between hierarchical rules, desensitization depth parameters, and utility thresholds, achieving collaborative optimization of policies across system modules, making the entire data analysis and management process more efficient and robust. Combined with feedback signals, the system can adaptively adjust policies based on actual operating conditions and changes in the external environment, continuously optimizing system performance.

[0184] In some embodiments described above, this application proposes using a comprehensive evaluation index that includes latency cost, communication bandwidth cost, and privacy exposure risk cost to drive the update of system policy parameters. However, in its implementation, accurately and dynamically quantifying the privacy exposure risk brought about by the anonymized intermediate representation to ensure that the adjustment of policy parameters can effectively balance data utility and privacy protection is a specific technical problem that needs to be solved. If the privacy risk assessment is inaccurate, it may lead to over-anonymization that sacrifices data value, or insufficient anonymization that increases the risk of privacy leakage.

[0185] To address this, this application further proposes determining the privacy exposure risk cost component as follows: A pre-defined reverse reconstruction attack simulation is performed on the de-identified intermediate representation to obtain reconstructed data units. The recognizable field restoration rate between the reconstructed data units and the original player behavior data units is calculated. This recognizable field restoration rate is used as the privacy exposure risk cost component; a higher recognizable field restoration rate results in a larger privacy exposure risk cost component, and a greater adjustment range for the lower limit of the corresponding cutoff layer depth value range in the policy parameter update.

[0186] For example, to accurately assess the privacy exposure risk of the anonymized intermediate representation, this application performs a pre-defined reverse reconstruction attack simulation on the anonymized intermediate representation to obtain reconstructed data units. This step aims to objectively assess the privacy security of the anonymized intermediate representation by simulating potential privacy attacks. For example, various machine learning models or statistical methods can be used, such as training an inverse decoder (e.g., an autoencoder or a generator in a generative adversarial network), which takes the anonymized intermediate representation as input and attempts to reconstruct the original player behavior data units. The pre-defined reverse reconstruction attack simulation can include a series of known attack models and strategies for specific data types or feature representations, which can be selected and configured based on historical attack events, domain knowledge, or privacy protection standards. By performing such simulations, the system can obtain a reconstructed data unit that approximates the original data structure, providing a basis for subsequent privacy risk quantification. In one specific embodiment, the inverse reconstruction attack simulation uses a pre-trained inverse decoder network based on an autoencoder architecture, where the encoder is the first K layers of a local feature extraction network, and the decoder consists of several deconvolutional layers and upsampling layers. The reconstruct rate of identifiable fields is R_reconstruct = N_matched / N_total, where N_matched is the number of identifiable fields in the reconstructed data unit that are completely identical to the corresponding field values ​​in the original data unit or whose edit distance is less than the preset tolerance, and N_total is the total number of identifiable fields in the original data unit. When R_reconstruct exceeds the preset security threshold (e.g., 0.3), the system determines that the current desensitization depth is insufficient and increases the lower limit of the cutoff layer depth value range in the policy parameter update by increasing the privacy exposure risk cost component.

[0187] Based on this, this application calculates the identifiable field restoration rate between the reconstructed data unit and the original player behavior data unit. This step is used to quantify the success rate of reverse reconstruction attacks, i.e., the ability to recover original sensitive information from the anonymized intermediate representation. Specifically, the system identifies fields with unique identifiers or high sensitivity in the original player behavior data unit (e.g., player ID, device ID, precise timestamp, specific behavior sequence, etc.) and compares them with the corresponding fields in the reconstructed data unit. The identifiable field restoration rate can be defined as the proportion of successfully reconstructed identifiable fields that highly match the original data to the total number of identifiable fields in the original data, or it can be measured by calculating similarity metrics between these fields (e.g., edit distance, Jaccard similarity coefficient, etc.). This restoration rate directly reflects the risk of the anonymized intermediate representation being reverse-engineered and identified.

[0188] This application uses the identifiable field restoration rate as a privacy exposure risk cost component. A higher identifiable field restoration rate results in a larger privacy exposure risk cost component, and a greater adjustment range for the lower limit of the corresponding cutoff layer depth value range in the policy parameter update. This step directly integrates the quantified privacy leakage risk into the system policy optimization feedback loop. By using the identifiable field restoration rate as a privacy exposure risk cost component, the system can obtain a dynamic risk metric based on actual attack effects. When the restoration rate is high, it indicates insufficient privacy protection in the de-identified intermediate representation, and the corresponding privacy exposure risk cost component will increase accordingly. This increased cost component will further affect the policy parameter update, particularly prompting the system to adjust the lower limit of the cutoff layer depth value range to a greater extent. This means that, to reduce privacy risks, the system will tend to choose deeper cutoff layer positions, thereby performing more feature extraction and de-identification processing on local computing nodes, reducing the amount of original information transmitted to collaborative computing nodes, and further enhancing privacy protection.

[0189] Through the above technical solution, this application provides a specific and quantifiable method to determine the privacy exposure risk cost component. By performing a preset reverse reconstruction attack simulation on the anonymized intermediate representation and calculating the identifiable field restoration rate between the reconstructed data unit and the original player behavior data unit, the system can obtain a dynamic and accurate privacy risk metric based on the actual attack effect. This accurate risk quantification avoids evaluation biases that may arise from relying on static or heuristic rules, thus enabling the system to more accurately assess the effectiveness of the current anonymization strategy. When the identifiable field restoration rate is high, the system can promptly detect the increase in privacy risk and accordingly increase the privacy exposure risk cost component, thereby driving a greater adjustment to the lower limit of the cutoff layer depth range in the strategy parameter update. This prompts the system to select a deeper cutoff layer in subsequent processing, enhancing local anonymization and effectively reducing the privacy leakage risk of data transmitted to collaborative computing nodes. Therefore, this application can achieve a more refined and adaptive balance between privacy protection and data utility, ensuring maximum protection of user data privacy while meeting the needs of collaborative analysis tasks.

[0190] A large-scale multiplayer online game system deploys multiple local computing nodes (game user data analysis and management systems) globally to continuously collect player behavior data. A central collaborative computing node is responsible for real-time collaborative analysis of this data to generate game operation collaborative analysis conclusions such as player churn prediction, abnormal behavior markers, and payment tendency scores. When processing this type of data, related technologies typically configure privacy anonymization, load simplification, and task scheduling independently. This leads to critical analysis tasks failing to guarantee timely processing under sudden loads due to low-value data consuming computing and communication resources. Simultaneously, anonymizing highly privacy-sensitive data and low privacy-sensitive data at the same granularity results in insufficient protection for the former or wasted communication bandwidth for the latter.

[0191] The following example will provide a more detailed explanation of the above technical solution: When player A performs the action of "purchasing rare items" on a local computing node, the data unit of this action is captured by the system.

[0192] The system performs online semantic analysis on the incoming player behavior data unit to obtain content-sensitive attributes and privacy-sensitive attributes.

[0193] Regarding the data unit on the behavior of "purchasing rare items": Determining Content Sensitive Attributes: The system determines the event rarity of a behavioral data unit based on its frequency of occurrence among historical players in the same group. For example, the event "purchasing rare items" occurs less frequently among players in the same group, thus its event rarity is high. The system inputs this behavioral data unit into a distilled simplified version of the main model corresponding to the current collaborative analysis task (e.g., player churn prediction) to perform forward inference, estimating the behavioral impact of the data unit on the churn prediction conclusion, thus obtaining the behavioral impact score. For example, purchasing rare items may indicate that a player is highly invested in the game and has a low probability of churn in the short term, thus the behavioral impact score is high. Based on the upper bound of the latency of the current churn prediction task (e.g., requiring prediction to be completed within 500 milliseconds), a real-time requirement component is determined. The shorter the upper bound of the latency, the larger the real-time requirement component. The system also determines whether the behavioral data unit completes a specific detected behavioral pattern (e.g., a key node of a "high-paying player" pattern). If the purchase behavior completes a high-value pattern, a high pattern completion contribution score is obtained. The system performs multiplicative modulation on the impact of the behavior based on the event's rarity, and then weights and fuses the modulated impact, the real-time requirement component, and the contribution to pattern completion to obtain the content sensitivity attribute of the "purchasing rare items" behavior data unit. It is assumed that this behavior is assigned a high content sensitivity attribute value due to its rarity, high impact on churn prediction, and completion of a high-value pattern.

[0194] Determining Privacy-Sensitive Attributes: The system decomposes the behavioral data unit into fields, obtaining a set of field types, such as "Player Account Identifier," "Device Identifier," "Purchase Amount," and "Item Name." Based on a preset sensitivity mapping relationship (e.g., account identifier has the highest sensitivity weight, followed by device identifier, and behavioral attribute has the lowest), the system determines the sensitivity weight corresponding to each field type. Since the combination of "Player Account Identifier" and "Device Identifier" can uniquely identify a specific player, the system performs a joint enhancement process on the sensitivity weights corresponding to these two field types, and uses the maximum value of the enhanced sensitivity weights as the privacy-sensitive attribute of the behavioral data unit. For example, if the behavioral data unit contains both a player account identifier and a device identifier, it is assigned a higher privacy-sensitive attribute value.

[0195] The system performs a joint mapping between the content-sensitive attributes and the privacy-sensitive attributes to obtain the de-identification depth parameter and the processing priority label.

[0196] The system performs a multiplicative fusion of high content sensitivity attributes and high privacy sensitivity attributes to obtain a higher joint index. This joint index is mapped to a preset range of truncation layer depth values ​​to obtain the desensitization depth parameter. For example, because the joint index of this behavioral data unit is high, the system assigns it a deeper truncation layer position. This means that more forward computation layers need to be executed in the local feature extraction network to extract more abstract and harder-to-reverse-engineer features.

[0197] The system compares the content sensitivity attribute with a preset sensitivity threshold. Since the content sensitivity attribute exceeds the threshold, the behavioral data unit is set as a high-priority processing priority label.

[0198] Unlike related technologies that use a uniform desensitization granularity for all data, this solution dynamically adjusts the desensitization depth based on the actual content value and privacy risks of the data. This avoids the waste of resources caused by excessive desensitization of low-privacy data and ensures stronger protection for high-privacy data.

[0199] Based on the desensitization depth parameter, the local computing node performs hierarchical feature extraction and truncation in the local feature extraction network to obtain the desensitized intermediate representation.

[0200] Based on the determined depth parameters for desensitization, the local computing node selects the Kth layer in the local feature extraction network as the cutoff layer. The local computing node performs forward computation on the first K layers of this network to obtain the intermediate activation features output by the Kth layer. These intermediate activation features serve as the intermediate desensitization representation and are only sent to the collaborative computing nodes.

[0201] If the intermediate activation feature simultaneously meets the low-rank decomposition triggering condition that the content-sensitive attribute exceeds the preset content threshold and the privacy-sensitive attribute exceeds the preset privacy threshold, the local computing node will also perform low-rank decomposition on the intermediate activation feature, splitting it into low-rank decomposition component groups, and using these component groups to replace the original intermediate activation feature as the desensitized intermediate representation output, further enhancing privacy protection.

[0202] Based on the content sensitivity attributes and the real-time load status of the local computing node, the system performs load-aware simplification on the player behavior data unit to obtain the retained data unit.

[0203] The content-sensitive attributes of the "purchasing rare items" behavior data unit are used as its utility score. The local compute node continuously monitors its real-time load status, such as compute utilization, task queue depth, and uplink communication bandwidth capacity. If the current task queue depth of the local compute node significantly exceeds the preset queue depth reference value, it indicates a high load. Based on this deviation, the system adjusts the minimum utility threshold upwards.

[0204] Because the utility score (content-sensitive attribute) of the "purchasing rare items" behavior data unit is high and not lower than the adjusted minimum utility threshold, this data unit is reserved and will not be downsampled or discarded.

[0205] Unlike related technologies that adjust the sampling rate based solely on node resource occupancy, this solution uses the expected contribution of data to the analysis task (content-sensitive attributes) as the key basis for simplification decisions. This ensures that high-value data is prioritized for retention when the load is high, avoiding the problem that critical analysis tasks cannot guarantee processing timeliness due to low-value data occupying resources.

[0206] The collaborative computing nodes perform fusion analysis on the desensitized intermediate representation and the cross-source association identifier carried by the retained data units to generate collaborative analysis conclusions on game operation.

[0207] The collaborative computing node receives the anonymized intermediate representation from the local computing node. Based on its high-priority processing priority label, this anonymized intermediate representation is preferentially entered into the collaborative computing node's computing queue for forward computation. The collaborative computing node then performs forward computation on the remaining layers after the truncated layer in its local feature extraction network based on this anonymized intermediate representation to obtain the complete feature representation.

[0208] Collaborative computing nodes align anonymized intermediate representations from different data sources (including the anonymized intermediate representation of the "purchasing rare items" behavior) according to cross-source association identifiers such as player identifier, match identifier, or time window identifier. The aligned multi-source anonymized intermediate representations are then spliced ​​and attention-weighted aggregated in a shared feature space to obtain cross-source fusion features. The collaborative analysis model performs classification inference based on these cross-source fusion features to generate collaborative analysis conclusions for game operations; for example, marking player A's churn probability as "low" and their payment tendency score as "high".

[0209] Unlike related technologies that rely on preset static priorities, this solution dynamically adjusts the allocation of computing resources based on data semantics, ensuring that high-value and time-sensitive data can be processed first, thereby improving the response speed and accuracy of analysis tasks.

[0210] The system also extracts feedback signals from the game operation collaborative analysis conclusions. For example, if it finds that the contribution of behaviors like "purchasing rare items" to churn prediction is consistently lower than expected, the system will use this as a feedback signal. Based on this feedback signal, the system will update the classification rules used in online semantic parsing to determine content-sensitive attributes online. For example, it will reduce the fusion weight of the baseline content-sensitive attribute corresponding to that data category in the weighted fusion. This will result in the same "purchasing rare items" behavior data unit receiving a lower content-sensitive attribute value in the next cycle, which may lead to a shallower cutoff layer position or make it easier to downsample or discard during load simplification. This closed-loop feedback mechanism enables the entire system to adaptively optimize resource allocation and privacy protection strategies, further improving efficiency and accuracy.

[0211] As demonstrated by the examples above, this method achieves differentiated privacy protection, intelligent load management, and dynamic task scheduling for player behavior data through the synergistic effects of online semantic parsing, joint mapping, hierarchical feature extraction, and load-aware simplification. This solves the multi-dimensional resource trade-off mismatch caused by the independent configuration of privacy anonymization, load simplification, and task scheduling in related technologies, improving the efficiency, accuracy, and privacy security of cross-source game data collaborative analysis.

[0212] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0213] Figure 5This is a schematic diagram of the structure of a user game data analysis and management system 500 provided in an embodiment of this application. The user game data analysis and management system 500 can vary significantly due to different configurations or performance. It may include one or more processors (Central Processing Units, CPUs) 501 and one or more memories 502. The one or more memories 502 store at least one computer program, which is loaded and executed by the one or more processors 501 to implement the methods provided in the above-described method embodiments. Of course, the user game data analysis and management system 500 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The user game data analysis and management system 500 may also include other components for implementing device functions, which will not be elaborated here.

[0214] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for analyzing and managing user game data, characterized in that, The method includes: Online semantic analysis is performed on the incoming player behavior data units to obtain content-sensitive attributes and privacy-sensitive attributes. The content-sensitive attributes characterize the expected impact of the player behavior data unit on the conclusion of the current collaborative analysis task, and the privacy-sensitive attributes characterize the degree to which the information carried by the player behavior data unit can identify an individual. The content-sensitive attribute and the privacy-sensitive attribute are jointly mapped to obtain the desensitization depth parameter and the processing priority label. The desensitization depth parameter determines the truncation layer position of the local feature extraction network, and the processing priority label distinguishes the computational resource contention weight. Based on the desensitization depth parameter, hierarchical feature extraction and truncation are performed in the local feature extraction network to obtain the desensitized intermediate representation. Based on the content sensitivity attribute and the real-time load status of the local computing node, load-aware simplification is performed on the player behavior data unit to obtain the retained data unit. The desensitized intermediate representation and the cross-source association identifier carried by the retained data unit are fused and analyzed to generate game operation collaborative analysis conclusions.

2. The method according to claim 1, characterized in that, The online semantic analysis of the incoming player behavior data units yields content-sensitive attributes, including: Based on the frequency of occurrence of the player behavior data unit in the historical distribution of the same group, the event rarity is determined, and the event rarity reflects the scarcity of the event type corresponding to the player behavior data unit; Based on the distilled simplified version of the master model corresponding to the current collaborative analysis task, forward inference is performed on the player behavior data unit to estimate the degree of behavioral influence of the player behavior data unit on the conclusion of the current collaborative analysis task, and the behavioral influence degree is obtained. The real-time requirement component is determined based on the upper limit of the latency of the current collaborative analysis task. The shorter the upper limit of latency, the larger the real-time requirement component. Determine the pattern completion degree of the player behavior data unit in the detected behavior patterns to obtain the pattern completion contribution. The event rarity is used to perform multiplicative modulation on the behavior impact, and the modulated behavior impact, the real-time requirement component, and the mode completion contribution are weighted and fused to obtain the content sensitivity attribute.

3. The method according to claim 2, characterized in that, The determination of event rarity based on the frequency of occurrence of the player behavior data unit in the historical distribution of the same group, and the determination of real-time requirement components based on the upper bound of the latency of the current collaborative analysis task, include: The total number of times the event type corresponding to the player behavior data unit appears in the same group of players within a preset historical time window is counted. The total number of active events of the same group of players within the preset historical time window is determined. The relative size of the total number of occurrences to the total number of active events is used as the occurrence frequency. The event rarity is determined in a way that is negatively monotonically correlated with the occurrence frequency. Based on the type identifier of the current collaborative analysis task, a preset latency upper bound mapping table is queried to determine the latency upper bound of the current collaborative analysis task. The basic real-time metric is determined in a way that is negatively monotonically correlated with the latency upper bound. The basic real-time metric is normalized to a preset numerical range to obtain the real-time requirement component. The event rarity and the real-time requirement component are used as input values ​​for the corresponding dimensions in the weighted fusion of the content-sensitive attributes, wherein the event rarity is also used as a modulation factor for the multiplicative modulation of the behavior influence.

4. The method according to claim 1, characterized in that, The privacy-sensitive attributes are determined in the following ways: The player behavior data unit is decomposed into fields to obtain a set of field types that make up the player behavior data unit; Based on a preset sensitivity mapping relationship, determine the sensitivity weight corresponding to each field type in the field type set; When there is a combination of multiple field types in the set of field types that can uniquely identify a specific player, a joint boosting process is performed on the sensitivity weights corresponding to each field type in the combination, and the maximum value of the boosted sensitivity weights is taken as the privacy sensitive attribute. When the combination does not exist in the set of field types, the maximum value of the sensitivity weight corresponding to each field type in the set of field types is taken as the privacy sensitive attribute.

5. The method according to claim 1, characterized in that, The joint mapping of the content-sensitive attributes and the privacy-sensitive attributes to obtain the de-identification depth parameter and processing priority label includes: Perform a multiplicative fusion of the content-sensitive attribute and the privacy-sensitive attribute to obtain a joint index; The joint index is mapped to a preset cutoff layer depth range to obtain the desensitization depth parameter. The lower limit of the cutoff layer depth range corresponds to the shallowest cutoff layer of the local feature extraction network, and the upper limit of the cutoff layer depth range corresponds to the deepest cutoff layer of the local feature extraction network. The content sensitivity attribute is compared with a preset sensitivity threshold. When the content sensitivity attribute exceeds the preset sensitivity threshold, the processing priority label is set to a high priority level. When the content sensitivity attribute does not exceed the preset sensitivity threshold, the processing priority label is set to a low priority level.

6. The method according to claim 1, characterized in that, The step of performing hierarchical feature extraction and truncation in the local feature extraction network based on the desensitization depth parameter to obtain the desensitized intermediate representation includes: Based on the desensitization depth parameter, the Kth layer in the local feature extraction network is determined as the cutoff layer. The larger the value of the desensitization depth parameter, the deeper the cutoff layer is in the local feature extraction network. Perform the first K layers of forward computation of the local feature extraction network on the local computing node to obtain the intermediate activation features output by the Kth layer; The intermediate activation features output from the Kth layer are used as the desensitized intermediate representation. Only the desensitized intermediate representation is output to the collaborative computing node. The collaborative computing node performs forward computation of the remaining layers after the truncated layer in the local feature extraction network based on the desensitized intermediate representation.

7. The method according to claim 1, characterized in that, The process of performing load-aware simplification on the player behavior data unit based on the content-sensitive attributes and the real-time load status of the local computing node to obtain the retained data unit includes: The content sensitivity attribute is used as the utility score of the player behavior data unit; Monitor the real-time load status of the local computing node, the real-time load status including at least one of computing power utilization, task queue depth and uplink communication bandwidth margin; The minimum utility threshold is adjusted up or down based on the degree of deviation between the real-time load status and the preset load reference value. Player behavior data units with utility scores lower than the minimum utility threshold are downsampled or discarded, while player behavior data units with utility scores not lower than the minimum utility threshold are retained as the data units.

8. The method according to claim 1, characterized in that, The process of fusing and analyzing the desensitized intermediate representation with the cross-source association identifier carried by the retained data unit to generate game operation collaborative analysis conclusions includes: The de-identified intermediate representations from different data sources are aligned according to the cross-source association identifier, which includes player identifier, game identifier, or time window identifier. The aligned multi-source desensitized intermediate representations are spliced ​​and attention-weighted aggregated in a shared feature space to obtain cross-source fusion features. The collaborative analysis model performs classification reasoning based on the cross-source fusion features to obtain at least one analytical conclusion. The at least one analytical conclusion shall be used as the game operation collaborative analysis conclusion, and the analytical conclusion shall include at least one of player churn probability, abnormal behavior markers, or payment tendency scores.

9. The method according to claim 1, characterized in that, The method further includes: Feedback signals are extracted from the game operation collaborative analysis conclusions. The feedback signals include at least one of the following: risk level changes of a specific data category, contribution statistics of a specific data category to the game operation collaborative analysis conclusions, or prediction bias of the collaborative analysis model. Based on the feedback signal, the classification rules used to determine content sensitive attributes in the online semantic parsing are updated, the baseline of the content sensitive attributes corresponding to the data categories with increased risk levels is increased by a preset step size, and the baseline of the content sensitive attributes corresponding to the data categories with contribution levels below a preset contribution threshold is decreased by the preset step size. Based on the updated grading rules, the mapping relationship of the desensitization depth parameters and the determination basis of the processing priority labels are adjusted so that the same player behavior data units in the next cycle obtain different cut-off layer positions and different computational resource contention weights.

10. A user game data analysis and management system, characterized in that, The system includes a processor and a memory, the processor being configured to execute the user game data analysis and management method according to any one of claims 1-9.

Images