Audio product function optimization method based on user behavior burying and related equipment

Abstract

The present application relates to the technical field of electric digital data processing, and discloses an audio product function optimization method based on user behavior burying and related equipment, which is used for improving the efficiency of audio product function optimization. The method comprises the following steps: obtaining user behavior burying data, device state data and abnormal event data to construct a standardized time sequence data set, and establishing a user task state machine model and a burying event association graph based on the standardized time sequence data set. The actual behavior sequence of the user is compared with the normal path and characteristics in the state machine model to identify abnormal sticking points, and then candidate factors are extracted from the event graph to perform causal analysis to obtain causes. An optimization scheme is generated based on the causes, and the optimization scheme is verified. The verified scheme is put online and recorded in a knowledge base to form a closed-loop iterative optimization process.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a method and related equipment for optimizing audio product functions based on user behavior tracking. Background Technology

[0002] Currently, while smart audio products have widely integrated user behavior tracking, log reporting, and basic data dashboards at the data collection level, the data application of most products remains at a rudimentary stage of "data collection" and "post-event problem review," failing to form a complete closed loop from tracking analysis to feature optimization, effect verification, and continuous iteration. When optimizing complex audio functions, product teams still heavily rely on human experience to judge whether a function is difficult to use, whether user configuration has failed, or whether users have churned. This makes it difficult to effectively transform the complex usage paths of multi-stage, multi-branch, and multi-parameter functions in audio products into quantifiable and executable optimization tasks. This reliance on manual optimization leads to difficulties in verifying the effects of feature redesigns through objective data, long version iteration cycles, and the recurrence of similar experience problems. Summary of the Invention

[0003] This invention provides a method and related equipment for optimizing audio product functions based on user behavior tracking, in order to solve the problem of low efficiency in audio product function optimization caused by the high dependence of audio product function optimization on human experience and the lack of a closed-loop iterative mechanism driven by objective data in the prior art.

[0004] The first aspect of this invention provides a method for optimizing audio product functions based on user behavior tracking, comprising: acquiring user behavior tracking data, device status data, and abnormal event data during the use of the audio product to obtain a standardized time-series dataset; constructing a user task state machine model and a tracking event correlation graph based on the standardized time-series dataset; inputting the user's actual behavior sequence into the user task state machine model, comparing the user's actual behavior sequence with the normal behavior path and normal behavior characteristics represented by the user task state machine model, and identifying abnormal bottlenecks in the user's actual behavior sequence based on the comparison results; in response to the identification of the abnormal bottlenecks, extracting candidate influencing factors related to the abnormal bottlenecks from the tracking event correlation graph, performing causal analysis on the candidate influencing factors to obtain the causal causes of the abnormal bottlenecks; generating a product optimization plan based on the causal causes, and verifying the effectiveness of the product optimization plan; launching the verified product optimization plan online, recording optimization information, and forming a product optimization knowledge base.

[0005] In one feasible implementation, before constructing the user task state machine model and the event association map based on the standardized time-series dataset, a training sample pre-screening step is included: Based on the system resource occupancy time-series sequence and network transmission delay time-series sequence in the device status data, the peak resource occupancy and average network delay of each successful session record during the execution of the target task are extracted; the peak and average values ​​are weighted and summed to calculate the external environmental interference coefficient of the successful session record; based on the crash event identifier, reconnection event identifier, and decoding failure event identifier in the abnormal event data, it is detected whether each successful session record in the standardized time-series dataset contains unexpected interruption events during the execution of the target task; if so, the successful session record is marked as an abnormal interruption sample and removed from the training sample set; the successful session records remaining after the dual screening of the device status data and the abnormal event data, which do not contain environmental interference and do not contain unexpected interruptions, are determined as pure training samples for statistical analysis of legal transition paths between state nodes.

[0006] In one feasible implementation, the step of constructing a user task state machine model and a tracking event association graph based on the standardized time-series dataset includes: defining multiple consecutive task stages according to the functional modules of the audio product, with each task stage corresponding to a state node in the state machine model, and assigning a corresponding node identifier to each state node; using the clean training samples as a statistical benchmark, statistically analyzing the transition directions and number of transitions between each state node, determining the legal transition paths between each state node based on the statistical results, and compiling them into a set of legal transition paths; wherein, the clean training samples are successful session records that have completed the target task and have been filtered and removed from environmental interference samples and abnormal interruption samples by the device status data and the abnormal event data; Based on the distribution of behavioral feature parameters under each state node in the pure training samples, the mean and standard deviation of the behavioral feature parameters of each state node are calculated. The mean ± N times the standard deviation is taken as the expected behavioral feature interval of the state node, where N is a preset tolerance coefficient. Using various types of embedded events as event graph nodes, the temporal following relationship between events as directed edges, and the time interval between adjacent events as edge attributes, the embedded event association graph is constructed. Each state node, the set of legal transition paths, and the expected behavioral feature interval corresponding to each state node are associated and stored in the state machine model to complete the construction of the user task state machine model, and a mapping relationship is established between each state node in the state machine model and each embedded event in the embedded event association graph.

[0007] In one feasible implementation, the user's actual behavior sequence is compared with the normal behavior path and normal behavior features represented by the user task state machine model, and abnormal bottlenecks in the user's actual behavior sequence are identified based on the comparison results. This includes: dividing the user's actual behavior sequence into multiple actual behavior sub-sequences according to the user session boundary; mapping each event in each actual behavior sub-sequence to the corresponding state node in the state machine model in sequence to generate a user actual path sequence and actual behavior features at each state node; comparing the user's actual path sequence with the normal behavior path represented by the state machine model to obtain a path comparison result; comparing the actual behavior features at each state node with the corresponding expected behavior features in the state machine model to obtain a feature comparison result; and determining whether there are abnormal state nodes in the user's actual behavior sequence that deviate from the normal behavior path or the normal behavior features based on the path comparison result and the feature comparison result, and marking the abnormal state nodes that are determined to deviate as abnormal bottlenecks.

[0008] In one feasible implementation, the step of dividing the user's actual behavior sequence into multiple actual behavior sub-sequences according to the user session boundary, and mapping each event in each actual behavior sub-sequence to the corresponding state node in the state machine model to generate the user's actual path sequence and actual behavior features at each state node includes: detecting session boundary signals in the user's actual behavior sequence; dividing the user's actual behavior sequence into multiple actual behavior sub-sequences using the session boundary signals as dividing points; mapping each event in each actual behavior sub-sequence to the corresponding state node in the state machine model according to a preset mapping rule between the embedded events and state nodes to generate the user's actual path sequence; recording the event timestamp information at each state node, and generating the actual dwell time at each state node based on the event timestamp information as the actual behavior feature.

[0009] In one feasible implementation, comparing the user's actual path sequence with the normal behavior path represented by the state machine model to obtain a path comparison result includes: obtaining a set of legal transition paths in the state machine model, the set of legal transition paths being used to represent the normal behavior path; extracting adjacent state node pairs from the user's actual path sequence, and sequentially determining whether each adjacent state node pair exists in the set of legal transition paths; counting the number and distribution of adjacent state node pairs that do not exist in the set of legal transition paths, and generating a path comparison result based on the statistical results.

[0010] In one feasible implementation, the step of comparing the actual behavioral features at each state node with the corresponding expected behavioral features in the state machine model to obtain feature comparison results includes: obtaining the expected behavioral feature intervals corresponding to each state node in the state machine model, wherein the expected behavioral feature intervals are used to characterize the normal behavioral features; extracting the actual behavioral feature values ​​at each state node from the user's actual path sequence, comparing the actual behavioral feature values ​​at each state node with the expected behavioral feature intervals corresponding to that state node one by one; determining whether the actual behavioral feature values ​​at each state node fall within the corresponding expected behavioral feature intervals, and generating feature comparison results based on the determination results of each state node.

[0011] In one feasible implementation, determining whether there are abnormal state nodes in the user's actual behavior sequence that deviate from the normal behavior path or the normal behavior characteristics based on the path comparison result and the feature comparison result, and marking the identified abnormal state nodes as abnormal checkpoints, includes: based on the path comparison result, detecting whether there are adjacent state node pairs in the user's actual path sequence that deviate from the set of legal transfer paths; if so, determining the latter state node in the adjacent state node pair as a first type of abnormal state node; based on the feature comparison result, detecting whether there are state nodes in each state node whose actual behavior feature value does not fall into the corresponding expected behavior feature range; if so, determining the state node as a second type of abnormal state node; merging the first type of abnormal state nodes and the second type of abnormal state nodes into an abnormal state node set, and marking each state node in the abnormal state node set as an abnormal checkpoint.

[0012] In one feasible implementation, the step of responding to the identification of the abnormal checkpoint by extracting candidate influencing factors related to the abnormal checkpoint from the event association graph and performing causal analysis on the candidate influencing factors to obtain the causal cause of the abnormal checkpoint includes: taking the mapping event corresponding to the state node of the abnormal checkpoint in the event association graph as the starting point, tracing back the associated events at a preset level in chronological order, extracting the device status parameters and user operation parameters corresponding to each associated event as candidate factors, and compiling them into a set of candidate influencing factors; sequentially taking out each candidate factor from the set of candidate influencing factors, dividing the user group into a first group and a second group according to the value of the candidate factor, and obtaining the behavioral performance differences between the first group and the second group at the abnormal checkpoint; judging whether each candidate factor has a statistically significant impact on the occurrence of the abnormal checkpoint based on the behavioral performance differences, and determining the candidate factors with significant impact as the causal cause of the abnormal checkpoint.

[0013] In one feasible implementation, the step of sequentially extracting candidate factors from the candidate influencing factor set, dividing the user group into a first group and a second group according to the value of the candidate factor, and obtaining the behavioral performance difference between the first group and the second group at the abnormal checkpoint includes: extracting a candidate factor from the candidate influencing factor set, obtaining the value distribution of the candidate factor in historical data, and determining the target value range of the candidate factor based on the value distribution; dividing users in the user group whose value of the candidate factor falls within the target value range into a first group, and dividing users whose value of the candidate factor falls outside the target value range into a second group; respectively counting the total number of users passing the abnormal checkpoint and the number of users passing the abnormal checkpoint in the first group and the second group, calculating the checkpoint pass rate of the first group and the second group, and taking the difference between the checkpoint pass rate of the first group and the checkpoint pass rate of the second group as the behavioral performance difference.

[0014] In one feasible implementation, the step of determining whether each candidate factor has a statistically significant impact on the occurrence of the abnormal checkpoint based on the behavioral performance differences, and determining the candidate factors with significant impact as the causal causes of the abnormal checkpoint, includes: obtaining the difference in checkpoint pass rates between the first group and the second group at the abnormal checkpoint; calculating the significance level corresponding to the difference based on the sample size of the first group and the second group; comparing the significance level with a preset significance threshold; when the significance level is less than or equal to the significance threshold, determining that the candidate factor has a statistically significant impact on the occurrence of the abnormal checkpoint; outputting the candidate factors determined to have a significant impact as the causal causes of the abnormal checkpoint; traversing all candidate factors in the candidate influencing factor set; and outputting the causal cause determination result corresponding to each candidate factor.

[0015] In one feasible implementation, the step of generating a product optimization plan based on the causal causes and verifying the effectiveness of the product optimization plan includes: determining the optimization direction based on the factor type to which the causal causes belong; determining the optimization position based on the task stage where the abnormal bottleneck is located; generating a product optimization plan based on the optimization direction and the optimization position; and binding a verification indicator for measuring the optimization effect to the product optimization plan; identifying the product optimization plan as an optimization plan to be verified; dividing users into an experimental group and a control group according to a preset traffic ratio; applying the optimization plan to be verified to the experimental group and not applying the optimization plan to the control group; collecting the indicator values ​​of the experimental group and the control group on the verification indicator respectively; calculating the indicator difference of the experimental group relative to the control group; and determining that the optimization plan to be verified has passed verification when the indicator difference meets a preset benefit condition and the performance of the experimental group on the preset negative monitoring indicator is not significantly worse than that of the control group; wherein, the preset negative monitoring indicator is used to measure the unexpected negative impact of the product optimization plan on the overall stability of the audio product or the user experience.

[0016] A second aspect of this invention provides an audio product function optimization device based on user behavior tracking, comprising: an acquisition module for acquiring user behavior tracking data, device status data, and abnormal event data during the use of the audio product to obtain a standardized time-series dataset; a construction module for constructing a user task state machine model and a tracking event correlation graph based on the standardized time-series dataset; an identification module for inputting the user's actual behavior sequence into the user task state machine model, comparing the user's actual behavior sequence with the normal behavior path and normal behavior characteristics represented by the user task state machine model, and identifying abnormal bottlenecks in the user's actual behavior sequence based on the comparison results; an analysis module for extracting candidate influencing factors related to the abnormal bottleneck from the tracking event correlation graph in response to the identification of the abnormal bottleneck, performing causal analysis on the candidate influencing factors, and obtaining the causal causes of the abnormal bottleneck; a verification module for generating a product optimization scheme based on the causal causes, and verifying the effectiveness of the product optimization scheme; and a processing module for launching the verified product optimization scheme online, recording optimization information, and forming a product optimization knowledge base.

[0017] A third aspect of the present invention provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to execute the above-described audio product function optimization method based on user behavior tracking.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described audio product function optimization method based on user behavior tracking.

[0019] In this embodiment of the invention, by constructing a user task state machine model and a tracking event correlation graph, the complex multi-stage and multi-branch functional usage paths in audio products are quantified into calculable state nodes and transition paths, achieving automated identification of abnormal bottlenecks in user behavior sequences. Based on this, by tracing candidate influencing factors backward from the tracking event correlation graph and performing causal analysis, the causal relationship between bottleneck phenomena and root causes is elevated from subjective experience-based inference to objective quantitative judgment based on data statistics, improving the accuracy and interpretability of abnormal root cause localization. Furthermore, by automatically generating optimization solutions based on causal causes and conducting A / B testing verification, functional optimization is transformed from "post-event manual repair" to a closed-loop iterative process of "data-driven - solution generation - effect verification - knowledge accumulation," effectively shortening the cycle of problem localization and solution verification, reducing the probability of similar problems recurring, and improving the efficiency and effectiveness of audio product functional optimization. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an embodiment of the audio product function optimization method based on user behavior tracking in this invention. Figure 2 This is a schematic diagram of another embodiment of the audio product function optimization method based on user behavior tracking in this invention. Figure 3 This is a schematic diagram of an embodiment of the audio product function optimization device based on user behavior tracking points in this invention. Figure 4 This is a schematic diagram of another embodiment of the audio product function optimization device based on user behavior tracking points in this invention. Figure 5 This is a schematic diagram of one embodiment of the electronic device in this invention. Detailed Implementation

[0021] This invention provides a method and related equipment for optimizing audio product functions based on user behavior tracking. By constructing a user task state machine model, the method performs dual-dimensional anomaly detection on user behavior sequences based on path and feature, and generates an effective optimization scheme based on causal analysis of anomaly points, thereby improving the efficiency of audio product function optimization.

[0022] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] It is understood that the executing entity of this invention can be an audio product function optimization device based on user behavior tracking, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0024] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the audio product function optimization method based on user behavior tracking in this invention includes: 101. Obtain user behavior tracking data, device status data, and abnormal event data during the use of audio products to obtain a standardized time-series dataset; The server continuously receives user interaction event logs reported by the audio product front-end through the data collection interface, pulls runtime status snapshots from the device management service, and obtains crash and lag reports from the anomaly monitoring system. The server performs time window alignment and field normalization on the multi-source heterogeneous data, filters out invalid records with collection anomalies or missing fields, and sorts and aggregates the remaining valid data according to user and time dimensions to generate a standardized time-series dataset.

[0025] 102. Based on standardized time-series datasets, construct a user task state machine model and a correlation graph of embedded events; Based on a standardized time-series dataset, the server constructs a user task state machine model using directed graph modeling. It defines a set of state nodes for each typical user task and uses historical normal samples to statistically analyze the distribution of transition frequencies and dwell times between states, fitting the probability distribution of normal behavior paths and their feature tolerance intervals. Simultaneously, the server employs a frequent pattern mining algorithm to analyze the sequential dependencies and concurrent associations between tracked events, constructing a tracked event association graph and generating embedding vectors for each node and edge.

[0026] 103. Input the user's actual behavior sequence into the user task state machine model, compare the user's actual behavior sequence with the normal behavior path and normal behavior characteristics represented by the user task state machine model, and identify abnormal bottlenecks in the user's actual behavior sequence based on the comparison results. After receiving real-time reported user behavior sequences, the server inputs them into the loaded task state machine model, performs state alignment operations, matches the state node to which the current behavior belongs one by one, and calculates the sequence edit distance and state dwell deviation between the actual transition path and the model's standard path. The server compares the deviation value with the preset normal feature thresholds in the model item by item. When the combined deviation of multiple consecutive state nodes exceeds the tolerance limit, the server determines that the user has encountered an abnormal bottleneck in that task phase, immediately marks the timestamp, state location, and associated context parameters of the bottleneck, and triggers the subsequent cause analysis process without waiting for user feedback.

[0027] 104. In response to the identification of abnormal checkpoints, candidate influencing factors related to the abnormal checkpoints are extracted from the event correlation graph, and causal analysis is performed on the candidate influencing factors to obtain the causal causes of the abnormal checkpoints. After identifying an abnormal bottleneck, the server performs a neighborhood search in the event association graph based on the corresponding event identifier. This extracts candidate event nodes and environmental attribute nodes that are temporally adjacent to or frequently co-occur with the identified bottleneck, forming a set of candidate influencing factors. The server then executes a time-series causal inference algorithm on this candidate set. By controlling for confounding variables and comparing the probability of bottleneck occurrence with and without the factor, the server identifies core factors with significant causal effects. Finally, the server outputs a causal report, clearly pointing to specific root causes such as server-side interface response delays, data parsing anomalies, or client-side resource loading bottlenecks.

[0028] 105. Generate product optimization plans based on causal relationships and verify the effectiveness of the product optimization plans; Based on the type of causal cause, the server matches corresponding adjustment actions from its built-in optimization strategy knowledge base. These actions might include modifying the interface timeout retry mechanism, adjusting the data prefetch window, optimizing recommendation ranking weights, or dynamically downgrading non-core functions. The server first replays historical anomaly sequences in an offline simulation environment to simulate changes in behavior paths after the strategy takes effect, evaluating the effectiveness of mitigation measures. The strategy validated through simulation is then used by the server to generate a canary deployment configuration, which is distributed to experimental group users through the configuration center. The server continuously collects and compares the occurrence rate of bottlenecks and the task completion rate between the experimental and control groups. The server automatically compares the differences in experimental metrics; if the expected improvement threshold is reached, a full rollout is performed; otherwise, the strategy is rolled back and iterated.

[0029] 106. Implementation of verified product optimization solutions, recording optimization information, and forming a product optimization knowledge base.

[0030] In this embodiment of the invention, by constructing a user task state machine model and a tracking event correlation graph, the complex multi-stage and multi-branch functional usage paths in audio products are quantified into calculable state nodes and transition paths, achieving automated identification of abnormal bottlenecks in user behavior sequences. Based on this, by tracing candidate influencing factors backward from the tracking event correlation graph and performing causal analysis, the causal relationship between bottleneck phenomena and root causes is elevated from subjective experience-based inference to objective quantitative judgment based on data statistics, improving the accuracy and interpretability of abnormal root cause localization. Furthermore, by automatically generating optimization solutions based on causal causes and conducting A / B testing verification, functional optimization is transformed from "post-event manual repair" to a closed-loop iterative process of "data-driven - solution generation - effect verification - knowledge accumulation," effectively shortening the cycle of problem localization and solution verification, reducing the probability of similar problems recurring, and improving the efficiency and effectiveness of audio product functional optimization.

[0031] Please see Figure 2 Another embodiment of the audio product function optimization method based on user behavior tracking points in this invention includes: 201. Obtain user behavior tracking data, device status data, and abnormal event data during the use of audio products to obtain a standardized time-series dataset; After acquiring the standardized time-series dataset, the server first groups the data according to session identifier and task identifier, and extracts the data window between the start time stamp and end time stamp of the target task. For the system resource utilization time-series sequence within this window, the server uses a sliding window extreme value detection algorithm, sets the window step size to a fixed sampling period, compares and records the maximum value of resource utilization as the peak value of resource utilization for this session. For the network transmission delay time-series sequence, after removing outliers caused by first packet delay and retransmission, the server uses an arithmetic mean to calculate the mean network delay for the remaining delay values. The server configures preset weighting coefficients for peak resource utilization and average network latency. The weighting coefficient for peak resource utilization is determined based on the audio product's sensitivity to CPU-intensive tasks, while the weighting coefficient for average network latency is determined based on the audio product's dependence on streaming transmission. The external environmental interference coefficient for the session is obtained by multiplying the peak value by its corresponding weighting coefficient and the average value by its corresponding weighting coefficient, and then summing the results. Both the weighting coefficients for peak resource utilization and average network latency are non-negative real numbers, and their values ​​are determined based on the audio product's sensitivity to local computing resource consumption and its dependence on real-time network transmission, respectively. The higher the sensitivity or dependence, the larger the corresponding weighting coefficient value, and the two are configured independently. The server compares this coefficient with a preset environmental interference threshold. If the threshold is exceeded, the session is marked as a high-environment-interference sample and removed from the training sample set. Meanwhile, the server searches the standardized time-series dataset for each successful session record to determine if it exhibits at least one of the following markers during the execution of the target task: a crash event marker, a reconnection event marker, or a decoding failure event marker. If any of these markers are present, the session is deemed to contain an unexpected interruption event, and the server marks the record as an abnormal interruption sample and removes it from the training sample set. If the same session simultaneously meets both the environmental interference removal condition and the interruption event removal condition, the server prioritizes performing the removal operation once and records the double-hit marker in the log. After performing the above dual screening on all successful session records, the server identifies the remaining successful session records that are neither marked as high environmental interference samples nor as abnormal interruption samples as pure training samples (i.e., completed task sessions that do not contain environmental interference or unexpected interruptions). The server also stores the removed samples separately in an abnormal sample database and records the removal reason label.

[0032] 202. Based on standardized time-series datasets, construct a user task state machine model and a correlation graph of embedded events; Based on the audio product's functional menu hierarchy and business operation process, the server divides the product's functional modules into an audio search module, a content browsing module, a playback control module, a playlist management module, and an account interaction module. Following the order of the functional modules experienced by the user in the target task, the search input stage, result display stage, playback loading stage, playback stage, and post-playback operation stage are defined as consecutive task stages. Each task stage corresponds to a state node in the state machine model. The server assigns a globally unique integer node identifier to each state node and adds a stage name label and module affixing label to each node. The server reads each session record from the clean training samples, extracts the state node sequence corresponding to each event and the transition pairs between adjacent nodes, and constructs a two-dimensional counting matrix using the adjacency matrix method. The row index of the matrix represents the starting node of the transition and the column index represents the ending node of the transition. The server traverses all adjacent node pairs in the clean training samples. Each time a transition from node i to node j occurs, the count value in the i-th row and j-th column of the matrix is ​​incremented by 1. After the traversal is completed, the server checks each cell in the matrix whose count value is greater than the preset minimum support threshold, determines the corresponding node pair as a valid transition path, and gathers all valid node pairs into a set of valid transition paths. The minimum support threshold is dynamically configured according to the total number of clean training samples and the functional complexity of the audio product to balance path coverage and noise filtering effect. The server further groups the clean training samples by state node, extracting behavioral feature parameter values ​​for all samples under each state node. These behavioral feature parameters include the actual dwell time at that node and the number of operations performed by preceding events. The server calculates the arithmetic mean and standard deviation of each parameter value, taking the mean ± N times the standard deviation as the expected behavioral feature interval for that state node, where N is a preset tolerance coefficient. The value range is configured according to the sensitivity of the audio product to operation latency at different task stages, with the tolerance coefficient being greater in the search stage than in the playback stage, to adapt to the fluctuation characteristics of user behavior in different scenarios. Simultaneously, the server uses various embedded events in the standardized time-series dataset as event graph nodes, traversing each session record, sorting by timestamp, and detecting the temporal following relationship between adjacent events. If event B occurs within a preset time window after event A, a directed edge from A to B is added to the graph, with the statistical distribution parameters of the time interval between adjacent events used as the attributes of this directed edge. The edge attributes include the mean, standard deviation, and maximum and minimum intervals.The server serializes and stores each state node, the set of legal transition paths and their transition count weights, and the expected behavior feature range of each state node in the state machine model configuration file. It also establishes a bidirectional mapping table between state node identifiers and event names, which is stored in a hash structure. This allows for quick lookup of associated event sets by state node identifiers and reverse lookup of state nodes by event names. After the model is built, the server loads the state machine model into the memory cache.

[0033] 203. Divide the user's actual behavior sequence into multiple actual behavior subsequences according to the user session boundary, and map each event in each actual behavior subsequence to the corresponding state node in the state machine model in sequence to generate the user's actual path sequence and the actual behavior features at each state node. When the server receives the user's actual behavior sequence, it simultaneously listens for session lifecycle signals reported by the application layer. These signals include user login success events, application suspension events when switching from foreground to background, application resumption events when switching from background to foreground, and user logout events. The server uses the login success and resumption events in the session boundary signals as session start markers, and the suspension and logout events in the session boundary signals as session end markers. Whenever the server detects a session start marker, it starts caching subsequent behavior events from that position until it detects the next session end marker. The cached event sequence is then divided into an actual behavior subsequence, thereby dividing the continuously reported event stream into multiple independent subsequences according to the user's actual operation session. After sorting the events in each actual behavior subsequence in ascending order by timestamp, the server reads the event name of each event sequentially. By querying the pre-established bidirectional hash mapping table between status nodes and event names, each event is mapped to its corresponding status node identifier. If an event does not find a corresponding node in the mapping table, the server assigns it to the status node of the nearest mapped event in time based on the time context of the event's occurrence and generates a mapping log for subsequent maintenance personnel to review. The server arranges the mapped status node identifiers in the original event order to generate the user's actual path sequence corresponding to the subsequence. The path sequence is stored in the form of an ordered array, with array elements recording each status node identifier and its entry timestamp in sequence. When recording event timestamps at each state node, the server, for each state node in the path sequence, obtains the timestamp of the first event under that node as the entry time and the timestamp of the last event under that node as the exit time. The difference between the exit time and the entry time is taken as the actual dwell time of that state node. If a state node contains only one event in the path sequence, the server subtracts the current event timestamp from the duration field of that event or the timestamp of the first event of the next state node to obtain the dwell time. The server uses the entry timestamp, exit timestamp, and actual dwell time of each state node as the actual behavioral characteristics of that node, and stores them in the user session feature table indexed by the state node identifier for subsequent feature comparison and analysis. The server also temporarily stores incomplete subsequences that fail to match the start and end identifiers after segmentation in a buffer queue and sets a waiting timeout. If the corresponding boundary signal is not received after the timeout, the incomplete sequence is discarded and a boundary loss log is recorded to avoid introducing noise data from incomplete sessions into the comparison process.

[0034] 204. Compare the user's actual path sequence with the normal behavior path represented by the state machine model to obtain the path comparison results; The server reads a pre-built set of valid transition paths from the state machine model in the memory cache. This set is stored in a hash set structure, where each valid transition pair exists as a string key in the form of "start node identifier_end node identifier". The value corresponding to each key records the transition frequency and probability in the clean training samples. The server iterates through the ordered array of the user's actual path sequence, starting from the second element. Each time, it takes the current element and its preceding element to form an adjacent state node pair, and formats this pair with the same string key. The server searches for this key in the valid transition path hash set. If a match is found, the transition is marked as valid; otherwise, it is marked as invalid, and the invalid transition pair is stored in the abnormal transition list. The server also records the position index and timestamp of the transition pair in the actual path sequence. After traversing all adjacent node pairs, the server counts the number of elements in the abnormal transfer list as the total number of illegal transfers. The server further performs distribution statistics on abnormal transfers according to task stages. Specifically, it groups and counts illegal transfers based on the task stage label of the starting node in each illegal transfer pair, generating a frequency distribution vector for each stage. Simultaneously, it calculates the average deviation of illegal transfers based on the jump distance between the starting and ending node identifiers (i.e., the absolute value of the difference between node identifier numbers). The server packages the total number of illegal transfers, the distribution vectors for each stage, and the average deviation into a path comparison result data structure. This structure also includes a Boolean anomaly flag; the flag is true when the total number of illegal transfers is greater than zero, and false otherwise. The server stores this path comparison result in the comparison record table of the current session and passes it to subsequent anomaly checkpoint judgment processes. The server also performs cluster analysis on consecutive illegal transfer pairs during path comparison. If consecutive illegal transfer pairs form an abnormal sub-path, the entire sub-path is marked as a path deviation segment, and the starting and ending nodes and duration of this segment are recorded.

[0035] 205. Compare the actual behavioral features at each state node with the corresponding expected behavioral features in the state machine model to obtain the feature comparison results; The server reads the expected behavioral feature intervals corresponding to each state node from the state machine model configuration file. These intervals are stored in a structured configuration table in the form of "mean ± N standard deviations". Each state node is associated with a lower limit and an upper limit of the feature interval, which are statistically derived from the distribution of behavioral feature parameters of that state node in the clean training samples. Simultaneously, the server reads the configured tolerance coefficient N for each state node. This tolerance coefficient is configured independently across different state nodes to accommodate the differentiated tolerance requirements for user behavior fluctuations at different task stages. The server extracts the actual dwell time values ​​corresponding to each state node in the current user's actual path sequence from the user session feature table, and builds an index based on the state node identifier to construct a list of actual feature values. The server sequentially retrieves each state node and compares its actual dwell time with the lower and upper limits of the expected behavioral feature interval. If the actual value is greater than or equal to the lower limit and less than or equal to the upper limit, the server determines that the node's feature is normal and marks it as a feature-passed node. If the actual value is less than the lower limit, the server determines it as an abnormally fast node, records its deviation direction as negative deviation, and calculates the difference between the lower limit and the actual value divided by the interval width to obtain the relative deviation degree. If the actual value is greater than the upper limit, the server determines it as an abnormally slow node, records its deviation direction as positive deviation, and calculates the difference between the actual value and the upper limit divided by the interval width to obtain the relative deviation degree. After completing the comparison of all status nodes, the server counts the number of feature-passed nodes, the number of abnormally fast nodes, and the number of abnormally slow nodes, and compiles the node identifier, deviation direction, relative deviation degree, and actual feature value of each abnormal node into a feature comparison result data structure in node index order. This data structure also includes an abnormal flag bit, which is set to true when the number of abnormally fast nodes or abnormally slow nodes is greater than zero. The server associates the feature comparison results with the path comparison results and stores them in the same session comparison record. It also marks overlapping nodes between the two types of anomalies, i.e., nodes that simultaneously exhibit path deviation and feature deviation. These nodes are prioritized for output as high-confidence anomaly nodes in subsequent anomaly checkpoint determination. The server also marks nodes whose actual values ​​are close to the interval boundary (within a preset percentage range near the boundary) as critical nodes during the feature comparison process. These are recorded separately as potential anomaly warning signals in the extended field of the comparison results.

[0036] 206. Based on the path comparison results and feature comparison results, determine whether there are any abnormal state nodes in the user's actual behavior sequence that deviate from the normal behavior path or normal behavior characteristics, and mark the abnormal state nodes that are determined to have deviations as abnormal checkpoints. The server reads the path comparison result data structure from the session comparison record and extracts the illegal transition list. Each item in this list records the starting node identifier, ending node identifier, and position index of an illegal adjacent state node pair in the actual path sequence. The server traverses this illegal transition list. For each illegal node pair in the list, it takes its ending node (i.e., the next state node) as the first type of abnormal state node and records the node identifier, task stage label, starting node information of the illegal transition, and position index of the node in the first type of abnormal node list. If the same state node is pointed to by multiple illegal transition pairs or appears repeatedly as the ending node, the server only retains the first appearance record and accumulates the number of times the node is hit, as an additional record of the path deviation severity index of the node. The server also reads the feature comparison result data structure from the session comparison record and extracts the abnormally fast node list and the abnormally slow node list. After merging and deduplicating these two lists, it obtains all state nodes whose actual behavioral feature values ​​do not fall within the expected range. The server records the node identifier, deviation direction (positive or negative deviation), relative deviation degree, and actual feature value of each node in the second type of abnormal node list. The server merges the first and second types of abnormal node lists using node identifiers as keys. First, it takes the union of the two lists to form an abnormal state node set. Nodes appearing only in the first type list are marked as path deviation type; nodes appearing only in the second type list are marked as feature deviation type; nodes appearing in both lists are marked as composite deviation type and assigned a composite anomaly priority label, prioritizing these high-confidence abnormal nodes in subsequent causal analysis. The server generates an abnormal checkpoint record for each state node in the abnormal state node set. Each record includes a node identifier, task stage label, abnormal type label, position index in the actual path sequence, deviation direction, and deviation degree value. All abnormal checkpoint records are sorted by position index and stored in an abnormal checkpoint list. The server passes the abnormal checkpoint list as output to the subsequent causal analysis process, simultaneously calculating the total number of abnormal checkpoints and the distribution ratio of abnormal types. If the abnormal checkpoint list is empty, subsequent steps are skipped and a normal session identifier is recorded; otherwise, the causal cause analysis process is triggered. The server also merges adjacent nodes in the abnormal checkpoint list that appear consecutively in the location index. If multiple consecutive abnormal checkpoints are detected to be adjacent in the path sequence and have the same abnormality type, they are aggregated into an abnormal checkpoint segment. The start and end nodes of this segment are used as the whole checkpoint block for subsequent processing to reduce the computational redundancy caused by repeatedly processing adjacent abnormal nodes in causal analysis.

[0037] The server reads the number of times the node has been hit by an illegal transfer. (That is, the number of illegal adjacent state node pairs of this node as the terminating node, used to characterize the severity of path deviation) and the degree of relative characteristic deviation of this node. (That is, the distance by which the actual dwell time exceeds the boundary of the expected behavioral feature interval divided by the interval width, used to characterize the severity of feature deviation); the server deviates from the preset path sensitivity coefficient. Sensitivity coefficient of deviation from features Calculate the path deviation sub-scores respectively. Deviation sub-scores from features Both are exponentially saturated and their values ​​fall within the range [0,1). The server then employs a noisy-OR fusion structure to combine the path deviation sub-score and the feature deviation sub-score into a comprehensive confidence score. The score also falls within the [0,1) range, with a higher confidence level indicating that the node is a genuine anomaly. The server then calculates the overall confidence score for each node. Compared with the preset confidence threshold (Values ​​between 0 and 1) are compared, and... Greater than or equal to The status node was determined to be an abnormal bottleneck; simultaneously, the server... Nodes identified as abnormal checkpoints are sorted from highest to lowest priority to form a priority list of abnormal checkpoints. This list serves as input for subsequent causal analysis. Nodes with higher priority are prioritized for candidate influencing factor extraction and causal analysis in step 207. Furthermore, the server can assign a comprehensive confidence score based on the criticality of the task stage to which the node belongs. Multiply by the preset criticality weight of the task stage in which the node is located. (Values ​​range from 0 to 1), resulting in a weighted overall confidence score. This approach prioritizes the identification of anomaly points during critical task phases such as search loading and playback startup, enabling the scoring model to flexibly adapt to the varying anomaly sensitivity requirements of different business scenarios in audio products. Through this comprehensive confidence scoring mechanism, the server integrates anomaly signals from both path deviation and feature deviation dimensions into a single, continuous, and sortable quantitative indicator, replacing the original hard-coded priority markers. This makes the criteria for identifying anomaly points more objective and interpretable. Simultaneously, by employing exponential saturation, it suppresses the excessive dominance of extreme outliers in a single dimension on the scoring, improving comparability and robustness across nodes and sessions.

[0038] For example, three candidate abnormal nodes are identified, and a preset path deviation sensitivity coefficient is used. Characteristic deviation sensitivity coefficient Confidence threshold Node i is located during the playback loading phase and is the termination node of both illegal transfer paths. The actual dwell time exceeded the upper limit of the expected characteristic interval by 1.5 times the interval width. This belongs to both path and feature deviation, and its comprehensive confidence score is... Node j only exhibits feature deviation ( , ), its rating Node k deviates from the path only once. , ), its rating Therefore, it can be seen that node i, which deviates in both the path and feature dimensions, receives a significantly higher confidence score than nodes j and k, which deviate in only one dimension, at the threshold. In this case, only node i is identified as a high-confidence abnormal checkpoint and is prioritized for causal analysis, while nodes j and k are downgraded as low-confidence nodes.

[0039] 207. In response to the identification of abnormal checkpoints, candidate influencing factors related to the abnormal checkpoints are extracted from the event correlation graph, and causal analysis is performed on the candidate influencing factors to obtain the causal causes of the abnormal checkpoints. For each abnormal checkpoint, the server traces its preceding event node backwards in the association graph through the mapping relationship between status nodes and event tracking. The specific process is as follows: After obtaining the list of abnormal checkpoints, the server sequentially extracts the status node identifier corresponding to each abnormal checkpoint. Through the bidirectional hash mapping table between status nodes and event tracking, it queries the corresponding mapping event node in the event association graph. The server uses this mapping event node as the starting node and performs hierarchical backtracking along the reverse direction of the directed edges in the graph. The fixed number of backtracking levels is set to 3. During each backtracking, only adjacent preceding event nodes with an edge weight greater than the global average edge weight are selected into the candidate set. If no node in a certain level satisfies the weight requirement, the server will proceed with the backtracking. For nodes with high thresholds, the backtracking of that branch is terminated early to avoid introducing weak correlation noise. After deduplicating all traced preceding event nodes, the server extracts the device resource utilization, network transmission latency, and available device memory at the time of occurrence from the standardized time-series dataset for each event node as device status parameters. At the same time, it extracts the user click type, swipe direction, and operation frequency before the event as user operation parameters. The server sorts all extracted parameters in reverse order of event occurrence time and constructs a candidate factor dictionary with parameter name as key and parameter value as value. All dictionaries are then combined into a set of candidate influencing factors, and the event node level and original timestamp from which each candidate factor originates are recorded.

[0040] The server sequentially selects a candidate factor from the set of candidate influencing factors. First, it identifies the data type of the candidate factor. If it is a continuous numerical factor (such as the average network transmission delay, peak device resource utilization, etc.), the server obtains the value distribution of the factor in all historical clean training samples, calculates its arithmetic mean and standard deviation, and uses the mean plus or minus one standard deviation as the target value range of the factor. If it is a discrete categorical factor (such as the type of pre-operation, network connection method, etc.), the server counts the frequency of each value of the factor in historical samples, takes the value with the highest frequency as the target value, and classifies samples with the same value as the target range and samples with different values ​​as outside the target range. Based on historical session data of all currently online users, the server categorizes users whose values ​​fall within the target range for the candidate factor into a first group and those whose values ​​fall outside the target range into a second group. The server then counts the total number of users who passed the abnormal checkpoint state node in both the first and second groups, as well as the number of users who successfully passed the abnormal checkpoint (i.e., subsequent state nodes are valid and the characteristics are normal). The server calculates the checkpoint pass rate for each group by dividing the number of successful users by the total number of users who passed the checkpoint. The server subtracts the pass rate of the second group from the pass rate of the first group to obtain the behavioral performance difference value. A positive difference value indicates that users within the target value range perform better at the checkpoint, while a negative difference value indicates that users outside the target value range perform better. The larger the absolute value of the difference, the stronger the distinguishing ability of the candidate factor for checkpoint pass rate.

[0041] After calculating the pass rates for the first and second groups, the server obtains the difference between the two pass rates as the raw effect size. The server reads the user sample size n1 of the first group and n2 of the second group, as well as the number of users x1 and x2 who passed the abnormal checkpoints in each group, and constructs a 2×2 contingency table. The four cells represent the number of users who passed in the first group (x1), the number who failed in the first group (n1-x1), the number who passed in the second group (x2), and the number who failed in the second group (n2-x2). The server uses a chi-square test to calculate the chi-square statistic of this contingency table, and obtains the corresponding p-value as the significance level based on the chi-square value and degrees of freedom (1). If the sample size of any group is less than 30, the server uses Fisher's exact test, directly calculating the one-sided cumulative probability of the four-fold table under a fixed marginal distribution based on the hypergeometric distribution, multiplied by 2. The p-values ​​are two-sided. The server compares the calculated p-values ​​with a preset significance threshold (fixed at 0.05). If the p-value is less than or equal to 0.05, the candidate factor is determined to have a statistically significant impact on the occurrence of abnormal checkpoints. After traversing all candidate factors, the server uses the Bonferroni method to perform multiple comparison correction on all p-values. Each p-value is multiplied by the total number of candidate factors to obtain the corrected p-value. The corrected p-value is then compared with 0.05, and only factors that are still less than or equal to 0.05 after correction are retained as causal causes. The server sorts the finally determined causal causes in descending order of absolute effect size and outputs the factor name, group pass rate difference, original p-value, corrected p-value, and direction of significant influence (positively promoting pass or negatively hindering pass) for each causal cause, forming a structured list of causal cause determination results.

[0042] 208. Generate product optimization solutions based on causal relationships and verify the effectiveness of the product optimization solutions; The server extracts the factor name and its type for each cause from the causal analysis result list. Factor types include four categories: device resource, network transmission, user operation, and content data. Based on the factor type, the server matches the corresponding optimization direction template: device resource type matches resource release and frequency reduction optimization; network transmission type matches connection strategy and timeout retry optimization; user operation type matches interface guidance and interactive feedback optimization; and content data type matches preloading and caching strategy optimization. Simultaneously, the server reads the task stage tag of the abnormal checkpoint from the record, including five categories: search input stage, result display stage, playback loading stage, playback stage, and post-playback operation stage. Based on the task stage, the server determines the optimization location, i.e., which functional module and business process link to optimize. The server combines the optimization direction and location, retrieves matching optimization actions from the preset strategy knowledge base, generates a product optimization plan containing specific operation parameters, and binds three verification indicators to the plan: the target indicator is the increase in the pass rate of the checkpoint; the auxiliary indicator is the change in user task completion rate; and the negative monitoring indicators include application crash rate and average response latency increment. The server uses a preset traffic ratio (e.g., 10% for the experimental group, 10% for the control group, and the remaining 80% not participating in the experiment) to divide daily active users into experimental, control, and observation groups using a user identifier hash modulo method. For experimental group users, the optimized strategy configuration is issued and activated at the corresponding task stage, while for control group users, the original product logic remains unchanged. During the experiment period (e.g., 7 consecutive calendar days), the server collects daily values ​​for each verification indicator for each user in both the experimental and control groups. For target and auxiliary indicators, the arithmetic mean of the daily values ​​for each user during the experiment period is taken as the indicator value for that user; for negative monitoring indicators, the maximum daily value during the experiment period is taken as the indicator value for that user. The server calculates the mean difference between the experimental group and the control group on the target indicator. If the difference is greater than the preset minimum benefit threshold (e.g., fixed at 3 percentage points), the benefit condition is considered met. The server also uses an independent samples t-test on the negative monitoring indicators to calculate the p-values ​​between the experimental group and the control group on the crash rate and response latency increment. If the p-value is greater than 0.05, the experimental group is considered not significantly worse than the control group. When both the benefit condition and the negative monitoring condition are met, the server determines that the optimization scheme to be verified has passed the verification; otherwise, it determines that it has failed and triggers the scheme rollback, while recording the reason log for the failure.

[0043] 209. Implementation of verified product optimization solutions, recording optimization information, and forming a product optimization knowledge base.

[0044] In this embodiment of the invention, the actual behavior sequence is segmented into sub-sequences according to the user session boundary and mapped to state nodes to generate the user's actual path sequence and the actual behavior features of each state node, thus realizing a structured representation of the user's complete task flow. Based on this, a dual comparison is performed from both the path and feature dimensions. The actual path sequence is compared with the set of legal transfer paths to detect deviations at the path level, and the actual behavior features of each state node are compared with the expected behavior feature range to detect deviations at the feature level. Anomalies are identified based on a comprehensive judgment of the two comparison results. Through this dual comparison mechanism, the identification of anomalies can not only capture path anomalies where the state node transfer order does not conform to the standard process, but also feature anomalies where the state node transfer order conforms to the standard but the node dwell time exceeds the normal fluctuation range. Compared with single-dimensional anomaly detection methods, this scheme significantly improves the recall and precision of anomaly identification through dual-dimensional cross-validation of path and features, effectively avoiding the problem of missed or false alarms caused by unreasonable single-dimensional threshold settings, and providing a more complete and accurate input of anomaly nodes for subsequent causal analysis.

[0045] The above describes the audio product function optimization method based on user behavior tracking in the embodiments of the present invention. The following describes the audio product function optimization device based on user behavior tracking in the embodiments of the present invention. Please refer to [link / reference]. Figure 3 One embodiment of the audio product function optimization device based on user behavior tracking points in this invention includes: The acquisition module 301 is used to acquire user behavior tracking data, device status data and abnormal event data during the use of audio products, and obtain a standardized time series dataset. Module 302 is used to build a user task state machine model and an event association graph based on a standardized time-series dataset. The identification module 303 is used to input the user's actual behavior sequence into the user task state machine model, compare the user's actual behavior sequence with the normal behavior path and normal behavior features represented by the user task state machine model, and identify abnormal bottlenecks in the user's actual behavior sequence based on the comparison results. Analysis module 304 is used to extract candidate influencing factors related to the abnormal checkpoint from the event association graph of the embedded point in response to the identification of the abnormal checkpoint, perform causal analysis on the candidate influencing factors, and obtain the causal causes of the abnormal checkpoint. The verification module 305 is used to generate product optimization solutions based on causal causes and to verify the effectiveness of the product optimization solutions. Processing module 306 is used to deploy the verified product optimization solutions online and record optimization information to form a product optimization knowledge base.

[0046] Please see Figure 4 Another embodiment of the audio product function optimization device based on user behavior tracking points in this invention includes, in addition to the aforementioned... Figure 3 In addition to the modules in the document, it also includes: Optionally, the filtering module 307 can be specifically used for: Based on the system resource utilization time series and network transmission delay time series in the device status data, the peak resource utilization and average network delay of each successful session record during the execution of the target task are extracted. The peak and average values ​​are weighted and summed to calculate the external environmental interference coefficient of the successful session record. Based on the crash event identifier, reconnection event identifier, and decoding failure event identifier in the abnormal event data, it is detected whether each successful session record in the standardized time series dataset contains unexpected interruption events during the execution of the target task. If it does, the successful session record is marked as an abnormal interruption sample and removed from the training sample set. The successful session records that remain after the dual screening of device status data and abnormal event data, which do not contain environmental interference and do not contain unexpected interruptions, are determined as pure training samples for statistical analysis of legal transition paths between state nodes.

[0047] Optionally, building module 302 can be specifically used for: Based on the functional modules of the audio product, multiple consecutive task stages are defined, each task stage corresponds to a state node in the state machine model, and a corresponding node identifier is assigned to each state node. Using clean training samples as a statistical benchmark, the transition directions and number of transitions between each state node are statistically analyzed. Based on the statistical results, the legal transition paths between each state node are determined and compiled into a set of legal transition paths. The clean training samples are successful session records that have completed the target task and have been filtered and removed from environmental interference samples and abnormal interruption samples by filtering out equipment status data and abnormal event data. Based on the distribution of behavioral feature parameters under each state node in the clean training samples, the mean and standard deviation of the behavioral feature parameters of each state node are calculated. The mean ± N times the standard deviation is used as the expected behavioral feature interval of the state node, where N is a preset tolerance coefficient. Using various types of embedded events as nodes in the event graph, the temporal following relationship between events is used as directed edges, and the time interval between adjacent events is used as edge attributes to construct an embedded event association graph. Each state node, the set of legal transition paths, and the expected behavioral feature interval corresponding to each state node are associated and stored in the state machine model to complete the construction of the user task state machine model and establish the mapping relationship between each state node in the state machine model and each embedded event in the embedded event association graph.

[0048] Optionally, the recognition module 303 includes: The generation unit 3031 is used to divide the user's actual behavior sequence into multiple actual behavior sub-sequences according to the user session boundary, and to map each event in each actual behavior sub-sequence to the corresponding state node in the state machine model in sequence, thereby generating the user's actual path sequence and the actual behavior features at each state node. The first comparison unit 3032 is used to compare the user's actual path sequence with the normal behavior path represented by the state machine model to obtain the path comparison result. The second comparison unit 3033 is used to compare the actual behavioral features at each state node with the corresponding expected behavioral features in the state machine model to obtain the feature comparison results. The processing unit 3034 is used to determine whether there are abnormal state nodes in the user's actual behavior sequence that deviate from the normal behavior path or normal behavior characteristics based on the path comparison results and feature comparison results, and to mark the abnormal state nodes that are determined to have deviations as abnormal checkpoints.

[0049] Optionally, the generation unit 3031 includes: The system detects session boundary signals in the user's actual behavior sequence and divides the user's actual behavior sequence into multiple actual behavior sub-sequences using the session boundary signals as the dividing points. According to the preset mapping rules between the event tracking points and state nodes, each event in each actual behavior sub-sequence is mapped to the corresponding state node in the state machine model to generate the user's actual path sequence. The system records the event timestamp information at each state node and generates the actual dwell time at each state node based on the event timestamp information, which serves as the actual behavior feature.

[0050] Optionally, the first comparison unit 3032 can be specifically used for: Obtain the set of legal transition paths in the state machine model, which is used to represent normal behavior paths; extract adjacent state node pairs from the user's actual path sequence, and determine whether each adjacent state node pair exists in the set of legal transition paths; count the number and distribution of adjacent state node pairs that do not exist in the set of legal transition paths, and generate path comparison results based on the statistical results.

[0051] Optionally, the second comparison unit 3033 can be specifically used for: Obtain the expected behavior feature intervals corresponding to each state node in the state machine model. The expected behavior feature intervals are used to represent normal behavior features. Extract the actual behavior feature values ​​at each state node from the user's actual path sequence. Compare the actual behavior feature values ​​at each state node with the expected behavior feature intervals corresponding to that state node one by one. Determine whether the actual behavior feature values ​​at each state node fall within the corresponding expected behavior feature intervals. Generate feature comparison results based on the judgment results of each state node.

[0052] Optionally, the processing unit 3034 may be specifically used for: Based on the path comparison results, detect whether there are adjacent state node pairs in the user's actual path sequence that deviate from the set of legal transfer paths. If so, the latter state node in the adjacent state node pair is identified as a first-type abnormal state node. Based on the feature comparison results, detect whether there are state nodes in each state node whose actual behavior feature value does not fall into the corresponding expected behavior feature range. If so, the state node is identified as a second-type abnormal state node. The first-type and second-type abnormal state nodes are merged into an abnormal state node set, and each state node in the abnormal state node set is marked as an abnormal checkpoint.

[0053] Optionally, analysis module 304 includes: Extraction unit 3041 is used to take the mapping event corresponding to the status node of the abnormal checkpoint in the embedded event association map as the starting point, trace the associated events of the preset level in reverse according to the time sequence, extract the device status parameters and user operation parameters corresponding to each associated event as candidate factors, and gather them into a set of candidate influencing factors. The acquisition unit 3042 is used to sequentially extract each candidate factor from the candidate influencing factor set, divide the user group into a first group and a second group according to the value of the candidate factor, and obtain the behavioral differences between the first group and the second group at the abnormal checkpoints. The judgment unit 3043 is used to determine whether each candidate factor has a statistically significant impact on the occurrence of abnormal checkpoints based on the differences in behavioral performance, and to determine the candidate factors with significant impact as the causal causes of abnormal checkpoints.

[0054] Optionally, the acquisition unit 3042 can be specifically used for: Select a candidate factor from the set of candidate influencing factors, obtain the value distribution of the candidate factor in historical data, and determine the target value range of the candidate factor based on the value distribution; divide users whose values ​​of the candidate factor fall within the target value range into the first group, and divide users whose values ​​of the candidate factor fall outside the target value range into the second group; count the total number of users who passed the abnormal checkpoints in the first group and the second group respectively, calculate the checkpoint pass rate of the first group and the second group respectively, and take the difference between the checkpoint pass rate of the first group and the checkpoint pass rate of the second group as the behavioral performance difference.

[0055] Optionally, the determination unit 3043 can be specifically used for: Obtain the difference in pass rate at abnormal checkpoints between the first group and the second group, and calculate the significance level corresponding to the difference based on the sample size of the first group and the second group; compare the significance level with a preset significance threshold, and determine that the candidate factor has a statistically significant impact on the occurrence of abnormal checkpoints when the significance level is less than or equal to the significance threshold; output the candidate factors determined to have a significant impact as the causal causes of abnormal checkpoints, traverse all candidate factors in the candidate influencing factor set, and output the causal cause determination result corresponding to each candidate factor.

[0056] Optionally, the verification module 305 can be specifically used for: The optimization direction is determined based on the type of factors to which the causal causes belong, and the optimization position is determined based on the task stage where the abnormal bottleneck is located. A product optimization plan is generated based on the optimization direction and position, and a verification indicator is bound to the product optimization plan to measure the optimization effect. The product optimization plan is designated as the optimization plan to be verified. Users are divided into an experimental group and a control group according to a preset traffic ratio. The optimization plan to be verified is applied to the experimental group, while it is not applied to the control group. The indicator values ​​of the experimental group and the control group on the verification indicators are collected separately, and the indicator difference between the experimental group and the control group is calculated. When the indicator difference meets the preset benefit conditions, and the performance of the experimental group on the preset negative monitoring indicators is not significantly worse than that of the control group, the optimization plan to be verified is deemed to have passed verification. The preset negative monitoring indicators are used to measure the unexpected negative impact of the product optimization plan on the overall stability of the audio product or the user experience.

[0057] above Figure 3 and Figure 4 The audio product function optimization device based on user behavior tracking points in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The electronic device in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0058] See Figure 5 As shown, the electronic device includes a processor 500 and a memory 501. The memory 501 stores machine-executable instructions that can be executed by the processor 500. The processor 500 executes the machine-executable instructions to implement the above-mentioned audio product function optimization method based on user behavior tracking.

[0059] Furthermore, Figure 5 The electronic device shown also includes a bus 502 and a communication interface 503. The processor 500, the communication interface 503 and the memory 501 are connected via the bus 502.

[0060] The memory 501 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 503 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 502 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0061] The processor 500 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 500 or by instructions in software form. The processor 500 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 501. The processor 500 reads the information in memory 501 and, in conjunction with its hardware, completes the method steps of the aforementioned embodiment.

[0062] The present invention also provides an electronic device, the computer device including a memory and a processor, the memory storing computer-readable instructions, which, when executed by the processor, cause the processor to perform the steps of the audio product function optimization method based on user behavior tracking points described in the above embodiments.

[0063] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the audio product function optimization method based on user behavior tracking.

[0064] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0065] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing audio product functions based on user behavior tracking, characterized in that, The method includes: Acquire user behavior tracking data, device status data, and abnormal event data during the use of audio products to obtain a standardized time-series dataset; Based on the standardized time-series dataset, a user task state machine model and a tracking event association graph are constructed. The user's actual behavior sequence is input into the user task state machine model. The user's actual behavior sequence is compared with the normal behavior path and normal behavior features represented by the user task state machine model. Based on the comparison results, abnormal bottlenecks in the user's actual behavior sequence are identified. In response to the identification of the abnormal checkpoint, candidate influencing factors related to the abnormal checkpoint are extracted from the event association map, and causal analysis is performed on the candidate influencing factors to obtain the causal causes of the abnormal checkpoint. Based on the aforementioned causal causes, a product optimization plan is generated, and the effectiveness of the product optimization plan is verified. Once the product optimization solutions that have been verified are implemented, the optimization information will be recorded to form a product optimization knowledge base.

2. The method for optimizing audio product functions based on user behavior tracking as described in claim 1, characterized in that, Before constructing the user task state machine model and the event association graph based on the standardized time-series dataset, a training sample pre-screening step is also included: Based on the system resource utilization time series and network transmission delay time series in the device status data, the peak resource utilization and average network delay of each successful session record during the execution of the target task are extracted. The peak value and average value are weighted and summed to calculate the external environment interference coefficient of the successful session record. Based on the crash event identifier, reconnection event identifier, and decoding failure event identifier in the abnormal event data, detect whether each successful session record in the standardized time series dataset contains an unexpected interruption event during the execution of the target task; if it does, mark the successful session record as an abnormal interruption sample and remove it from the training sample set. Successful session records that remain after being filtered by both the device status data and the abnormal event data, and that do not contain environmental interference or unexpected interruptions, are determined as pure training samples for statistical analysis of legitimate transition paths between state nodes.

3. The method for optimizing audio product functions based on user behavior tracking as described in claim 2, characterized in that, The construction of a user task state machine model and a tracking event association graph based on the standardized time-series dataset includes: Based on the functional modules of the audio product, multiple consecutive task stages are defined, each task stage corresponds to a state node in the state machine model, and a corresponding node identifier is assigned to each state node. Using the clean training samples as a statistical benchmark, the transition direction and number of transitions between each state node are statistically analyzed. Based on the statistical results, the legal transition paths between each state node are determined and aggregated into a set of legal transition paths. The clean training samples are successful session records that have completed the target task and have been filtered and removed from environmental interference samples and abnormal interruption samples by the device status data and the abnormal event data. Based on the distribution of behavioral feature parameters of each state node in the pure training samples, the mean and standard deviation of the behavioral feature parameters of each state node are calculated, and the mean ± N times the standard deviation is taken as the expected behavioral feature interval of the state node, where N is a preset tolerance coefficient. Using various types of event tracking points as nodes in the event graph, the temporal sequence relationship between events as directed edges, and the time interval between adjacent events as edge attributes, the event tracking point association graph is constructed. Each state node, the set of legal transition paths, and the expected behavior feature interval corresponding to each state node are associated and stored in the state machine model to complete the construction of the user task state machine model, and a mapping relationship is established between each state node in the state machine model and each event in the event association graph.

4. The method for optimizing audio product functions based on user behavior tracking as described in claim 1, characterized in that, The actual user behavior sequence is compared with the normal behavior path and normal behavior features represented by the user task state machine model. Based on the comparison results, abnormal bottlenecks in the actual user behavior sequence are identified, including: The user's actual behavior sequence is divided into multiple actual behavior sub-sequences according to the user session boundary. Each event in each actual behavior sub-sequence is sequentially mapped to the corresponding state node in the state machine model to generate the user's actual path sequence and the actual behavior features at each state node. The actual user path sequence is compared with the normal behavior path represented by the state machine model to obtain the path comparison result; The actual behavioral features at each state node are compared with the corresponding expected behavioral features in the state machine model to obtain the feature comparison results. Based on the path comparison results and the feature comparison results, it is determined whether there are any abnormal state nodes in the user's actual behavior sequence that deviate from the normal behavior path or the normal behavior characteristics, and the abnormal state nodes that are determined to deviate are marked as abnormal checkpoints.

5. The method for optimizing audio product functions based on user behavior tracking as described in claim 4, characterized in that, The step of dividing the user's actual behavior sequence into multiple actual behavior sub-sequences according to the user session boundary, and mapping each event in each actual behavior sub-sequence to the corresponding state node in the state machine model in sequence, generates the user's actual path sequence and the actual behavior features at each state node, including: Detect the session boundary signal in the user's actual behavior sequence, and use the session boundary signal as the dividing point to divide the user's actual behavior sequence into multiple actual behavior subsequences; According to the preset mapping rules between the event tracking points and the state nodes, each event in each actual behavior subsequence is mapped to the corresponding state node in the state machine model to generate the user's actual path sequence. Record the event timestamp information at each state node, and generate the actual dwell time at each state node based on the event timestamp information, as an actual behavior feature.

6. The method for optimizing audio product functions based on user behavior tracking as described in claim 4, characterized in that, The step of comparing the user's actual path sequence with the normal behavior path represented by the state machine model to obtain the path comparison result includes: Obtain the set of legal transition paths in the state machine model, the set of legal transition paths being used to characterize the normal behavior path; Extract adjacent state node pairs from the user's actual path sequence, and sequentially determine whether each adjacent state node pair exists in the set of legal transfer paths; The number and distribution of adjacent state node pairs that do not exist in the set of legal transition paths are counted, and path comparison results are generated based on the statistical results.

7. The method for optimizing audio product functions based on user behavior tracking as described in claim 4, characterized in that, The step of comparing the actual behavioral features at each state node with the corresponding expected behavioral features in the state machine model to obtain the feature comparison results includes: Obtain the expected behavior feature interval corresponding to each state node in the state machine model, and the expected behavior feature interval is used to characterize the normal behavior feature; Extract the actual behavior feature values ​​at each state node from the user's actual path sequence, and compare the actual behavior feature values ​​at each state node with the expected behavior feature interval corresponding to that state node one by one; Determine whether the actual behavior feature value at each state node falls within the corresponding expected behavior feature range, and generate feature comparison results based on the determination results of each state node.

8. The method for optimizing audio product functions based on user behavior tracking as described in claim 4, characterized in that, The step of determining whether there are any abnormal state nodes in the user's actual behavior sequence that deviate from the normal behavior path or the normal behavior characteristics based on the path comparison result and the feature comparison result, and marking the determined abnormal state nodes as abnormal checkpoints, includes: Based on the path comparison results, detect whether there are adjacent state node pairs in the user's actual path sequence that deviate from the set of legal transfer paths. If so, determine the latter state node in the adjacent state node pair as a first type of abnormal state node. Based on the feature comparison results, detect whether there are any state nodes in each state node whose actual behavior feature value does not fall into the corresponding expected behavior feature range. If so, the state node is identified as a second type of abnormal state node. The first type of abnormal state nodes and the second type of abnormal state nodes are merged into an abnormal state node set, and each state node in the abnormal state node set is marked as an abnormal checkpoint.

9. The method for optimizing audio product functions based on user behavior tracking as described in claim 1, characterized in that, In response to identifying the abnormal checkpoint, the process involves extracting candidate influencing factors related to the abnormal checkpoint from the event correlation graph, performing causal analysis on the candidate influencing factors to obtain the causal causes of the abnormal checkpoint, including: Starting from the mapping event corresponding to the status node of the abnormal checkpoint in the event association map, the associated events of the preset level are traced back in reverse order according to the time sequence, and the device status parameters and user operation parameters corresponding to each associated event are extracted as candidate factors and aggregated into a set of candidate influencing factors. Each candidate factor is sequentially extracted from the set of candidate influencing factors, and the user group is divided into a first group and a second group according to the value of the candidate factor. The behavioral differences between the first group and the second group at the abnormal checkpoint are obtained respectively. Based on the differences in behavioral performance, it is determined whether each candidate factor has a statistically significant impact on the occurrence of the abnormal checkpoint, and the candidate factors with significant impact are identified as the causal causes of the abnormal checkpoint.

10. The method for optimizing audio product functions based on user behavior tracking points according to claim 9, characterized in that, The step of sequentially extracting candidate factors from the set of candidate influencing factors, dividing the user group into a first group and a second group according to the value of the candidate factor, and obtaining the behavioral differences between the first group and the second group at the abnormal checkpoint includes: Take a candidate factor from the set of candidate influencing factors, obtain the value distribution of the candidate factor in historical data, and determine the target value range of the candidate factor based on the value distribution; Users whose values ​​for the candidate factor fall within the target value range are divided into a first group, and users whose values ​​for the candidate factor fall outside the target value range are divided into a second group. The total number of users who passed the abnormal checkpoints in the first group and the second group are counted respectively. The checkpoint pass rate of the first group and the second group is calculated. The difference between the checkpoint pass rate of the first group and the checkpoint pass rate of the second group is taken as the behavioral performance difference.

11. The method for optimizing audio product functions based on user behavior tracking points according to claim 9, characterized in that, The step of determining whether each candidate factor has a statistically significant impact on the occurrence of the abnormal checkpoint based on the differences in behavioral performance, and identifying the candidate factors with significant impact as the causal causes of the abnormal checkpoint, includes: Obtain the difference in checkpoint pass rate between the first group and the second group at the abnormal checkpoint, and calculate the significance level corresponding to the difference based on the sample size of the first group and the second group; The significance level is compared with a preset significance threshold. When the significance level is less than or equal to the significance threshold, the candidate factor is determined to have a statistically significant impact on the occurrence of the abnormal checkpoint. The candidate factors determined to have a significant impact are output as the causal causes of the abnormal checkpoints. All candidate factors in the set of candidate influencing factors are traversed, and the causal cause determination results corresponding to each candidate factor are output.

12. The method for optimizing audio product functions based on user behavior tracking as described in claim 1, characterized in that, The step of generating a product optimization plan based on the causal causes and verifying the effectiveness of the product optimization plan includes: The optimization direction is determined based on the factor type to which the cause and effect belong, the optimization position is determined based on the task stage where the abnormal bottleneck is located, a product optimization plan is generated based on the optimization direction and the optimization position, and a verification indicator for measuring the optimization effect is bound to the product optimization plan. The product optimization scheme is identified as the optimization scheme to be verified. Users are divided into an experimental group and a control group according to a preset traffic ratio. The optimization scheme to be verified is applied to the experimental group, while the optimization scheme to be verified is not applied to the control group. The indicator values ​​of the experimental group and the control group on the verification indicator are collected respectively, and the indicator difference of the experimental group relative to the control group is calculated. When the indicator difference meets the preset benefit condition, and the performance of the experimental group on the preset negative monitoring indicator is not significantly worse than that of the control group, the optimization scheme to be verified is determined to be verified. The preset negative monitoring indicator is used to measure the unexpected negative impact of the product optimization scheme on the overall stability of the audio product or the user experience.

13. An audio product function optimization device based on user behavior tracking, characterized in that, The audio product function optimization device based on user behavior tracking includes: The acquisition module is used to acquire user behavior tracking data, device status data, and abnormal event data during the use of audio products, and obtain a standardized time-series dataset. The construction module is used to build a user task state machine model and an event association graph based on the standardized time-series dataset. The identification module is used to input the user's actual behavior sequence into the user task state machine model, compare the user's actual behavior sequence with the normal behavior path and normal behavior features represented by the user task state machine model, and identify abnormal bottlenecks in the user's actual behavior sequence based on the comparison results. The analysis module is used to, in response to the identification of the abnormal checkpoint, extract candidate influencing factors related to the abnormal checkpoint from the event association map, perform causal analysis on the candidate influencing factors, and obtain the causal causes of the abnormal checkpoint. The verification module is used to generate a product optimization plan based on the causal causes and to verify the effectiveness of the product optimization plan. The processing module is used to deploy verified product optimization solutions online and record optimization information to form a product optimization knowledge base.

14. An electronic device, characterized in that, The electronic device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the electronic device to execute the audio product function optimization method based on user behavior tracking as described in any one of claims 1-12.

15. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the audio product function optimization method based on user behavior tracking as described in any one of claims 1-12.

Images