A service recommendation method and device, a storage medium and an electronic device

Abstract

The embodiment of the specification discloses a service recommendation method and device, a storage medium and an electronic equipment, wherein the method comprises the following steps: converting a user original behavior log into a semantic structure unit containing service item attributes and action features by using a preset interaction data collection burying point identification system; constructing an interleaved behavior sequence representing a user intention evolution track by alternately coding the extracted attributes and action features according to time; injecting candidate service item features into the sequence to form a target perception transduction sequence; reasoning a target perception intention feature; and finally, integrating the feature as an enhanced input into a preset recommendation model to output a prediction score. The embodiment of the specification effectively improves the intention capturing ability and prediction accuracy of the recommendation system by interleaved sequence modeling and target perception reasoning.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a service recommendation method, apparatus, storage medium, and electronic device. Background Technology

[0002] With the booming development of mobile internet service platforms, recommending content of interest to users from a vast array of services has become crucial for improving user experience and platform distribution efficiency. Recommendation systems in related technologies typically rely on deep learning models (such as models based on collaborative filtering or recurrent neural networks) to process users' historical interaction logs. By extracting features from users' past clicks, exposures, and other behavioral data, they predict users' future behavioral preferences and thus recommend services.

[0003] However, this method of service recommendation, which relies on deep learning models, still has limitations when dealing with complex interactive data. In practical applications, user behavior is often regarded as a simple ID sequence or a static feature set, making it difficult to accurately capture the dynamic evolution of user intent over time. As a result, the recommendation results often fail to accurately match the user's immediate needs in a specific context. Summary of the Invention

[0004] This specification provides a service recommendation method, apparatus, storage medium, and electronic device, the technical solutions of which are as follows:

[0005] Firstly, embodiments of this specification provide a service recommendation method, the method comprising:

[0006] Obtain the original interaction behavior logs of the target user for the target platform service, and use a preset interaction data collection point identification system to collect semantic structure units including service item attribute features and interaction action features from the original behavior logs;

[0007] Extract service item attribute features and interaction action features from each semantic structural unit, and encode the service item attribute features and interaction action features alternately over time to construct an interleaved behavior sequence that represents the user's intention evolution trajectory;

[0008] Injecting candidate service item features to be recommended into the interleaved behavior sequence yields a target perception transduction sequence, and performing behavioral intent reasoning on the target perception transduction sequence yields target perception intent features;

[0009] When the preset service recommendation model is invoked to recommend candidate services to the target user, the target perceived intent features are inserted into the feature input encoding layer of the preset service recommendation model, and the preset service recommendation model is used to evaluate the candidate service items to obtain a predicted service recommendation score.

[0010] Secondly, embodiments of this specification provide a service recommendation device, the device comprising:

[0011] The log processing module is used to obtain the original interaction behavior logs of the target user for the target platform service, and to collect semantic structure units including service item attribute features and interaction action features from the original behavior logs using a preset interaction data collection point identification system.

[0012] The sequence construction module is used to extract service item attribute features and interaction action features from each of the semantic structural units, and to alternately encode the service item attribute features and interaction action features over time to construct an interleaved behavior sequence that represents the user's intention evolution trajectory.

[0013] The intent reasoning module is used to inject candidate service item features to be recommended into the interleaved behavior sequence to obtain a target perception transduction sequence, and to perform behavioral intent reasoning on the target perception transduction sequence to obtain target perception intent features;

[0014] The service recommendation module is used to insert the target perceived intent features into the feature input encoding layer of the preset service recommendation model when the preset service recommendation model is invoked to recommend candidate services to the target user, and to obtain a predicted service recommendation score by recommending and evaluating the candidate service items through the preset service recommendation model.

[0015] Thirdly, embodiments of this specification provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the above-described method steps.

[0016] Fourthly, this specification provides a computer program product storing at least one instruction adapted to be loaded by a processor and to execute the method steps of one or more embodiments of this specification.

[0017] Fifthly, this specification provides a computer program product storing at least one instruction adapted to be loaded by a processor and to execute the method steps of one or more embodiments of this specification.

[0018] Fifthly, embodiments of this specification provide an electronic device that may include: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the above-described method steps.

[0019] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following:

[0020] In one or more embodiments of this specification, the original unstructured logs are parsed into semantic structural units containing both service item content and interactive actions by employing a preset interactive data collection and tracking system. A behavioral sequence with alternating content and actions is constructed along the time dimension. Furthermore, the features of candidate service items to be recommended are injected into the sequence to trigger target perception reasoning in a hierarchical sequence transduction model. Finally, the generated dynamic intent features are integrated into a preset recommendation model for evaluation. This addresses the limitations of related technologies, such as coarse granularity in user behavior feature mining, inability to distinguish the differences in the contribution of different interactive actions (e.g., clicks and views) to intent, and the lack of targeted reasoning due to the decoupling of user historical interest representation from the current target to be recommended. It achieves fine-grained capture of the complex evolution trajectory of user intent, enabling the generated intent features to dynamically adapt and perceive different candidate service items, thus improving the prediction accuracy and traffic distribution efficiency of the service recommendation system when facing dynamically changing user needs. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a service recommendation method provided in an embodiment of this specification;

[0023] Figure 2 This is a schematic diagram of a semantic structure unit acquisition process provided in the embodiments of this specification;

[0024] Figure 3 This is a schematic diagram of an intent reasoning process provided in an embodiment of this specification;

[0025] Figure 4 This is a schematic diagram of the structure of a service recommendation device provided in the embodiments of this specification;

[0026] Figure 5 This is a schematic diagram of the structure of an electronic device provided in the embodiments of this specification. Detailed Implementation

[0027] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0028] In the description of this specification, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this specification, it should be noted that, unless otherwise expressly specified and limited, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. Those skilled in the art can understand the specific meaning of the above terms in this specification based on the specific circumstances. Furthermore, in the description of this specification, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0029] The present specification will now be described in detail with reference to specific embodiments.

[0030] In one embodiment, such as Figure 1 As shown, a service recommendation method is proposed, which can be implemented using a computer program and run on a service recommendation device based on the von Neumann architecture. This computer program can be integrated into an application or run as a standalone utility application. The service recommendation device can be a terminal device, including but not limited to: personal computers, tablets, handheld devices, in-vehicle devices, wearable devices, computing devices, or other processing devices connected to a wireless modem. In different networks, the terminal device can be called by different names, such as: user equipment, access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent or user equipment, cellular phone, cordless phone, terminal device in 5G networks or future evolved networks, etc.

[0031] Specifically, the recommended methods for this service include:

[0032] S102: Obtain the original interaction behavior logs of the target user for the target platform service, and use a preset interaction data collection point identification system to collect semantic structure units including service item attribute features and interaction action features from the original behavior logs.

[0033] Raw interaction behavior logs: These refer to the complete transaction logs of user actions on the service platform, recorded by the server or client SDK. They typically include timestamps, unique user identifiers, device information, event types (such as impressions, clicks, and swipes), and event tracking parameter strings.

[0034] The pre-defined interactive data collection tracking system refers to a pre-defined structured tracking system used to map user interaction page-triggered behaviors within a service platform's services (such as apps and H5 pages) to business interaction logic semantics (e.g., a structured tracking system for financial services on a target platform). It can be viewed as a data collection tool for recording user behavior, a key collection encoding protocol that transforms physical user operations into logical semantic features. Optionally, a hierarchical structure (e.g., a four-level structure of site-page-block-location) can be adopted to uniquely identify the specific object of user interaction.

[0035] Optionally, the preset interactive data collection point identification system can consist of three parts: hierarchical topology, encoding generation rules, and a global semantic mapping table (also known as mapping rules).

[0036] Hierarchical topology: This accurately depicts the interaction context of users within the services provided on the service platform (such as APP clients, H5 web pages, and mini-programs). Hierarchical topology uses a multi-level tree structure to define the location of each interactive element. For example, a four-level structure of site-page-block-location can be used (this is not limited to this and can be customized for different services). Using a hierarchical topology, any interactive behavior (such as a click) is represented by the spatial context information of "which APP, which page, which block" where it occurs.

[0037] Encoding generation rules: A unique encoding generation rule (such as SPM encoding rule) is preset. When the user triggers an interaction, the location code is automatically generated according to the encoding generation rule.

[0038] Global Semantic Mapping Table: A bridge connecting location encoding and business interaction logic semantics. By maintaining a global mapping table, it supports the parsing and mapping process. The optional global mapping table is the mapping relationship between the tracking point identifier name, logical level coordinates, business semantic description, and associated action type. For example, a tracking point identifier name (Key) is xxx: the corresponding logical level coordinate is homepage-first focal image-first frame, the corresponding business semantic description is: Spring Festival shopping promotion, and the associated action type is exposure.

[0039] Semantic structural units refer to multi-data sets containing operation objects and operation properties generated after parsing and mapping. They form the basis for subsequent construction of behavioral sequences and may include service item attribute features (representing what the user saw or clicked on at this location) and interaction action features (representing what the user did with that service).

[0040] As an illustration, the system can read the full set of raw interaction behavior logs of a target user within a specified time window from a distributed file system or data warehouse according to a preset time period (e.g., early morning every day). After acquiring the logs, the system cleans the log text using preset regular expression parsing rules, extracting the user's unique identifier, event timestamp, and key tracking point location codes. A pre-built global tracking point semantic mapping table is loaded. This table, based on a preset interaction data collection tracking point identification system, records the correspondence between tracking point location codes at each level and specific business service items. Using the extracted tracking point location codes as indexes, the system performs a matching search in the global tracking point semantic mapping table, identifying and matching the service item attribute features (such as the specific service product ID, article topic) and interaction action features (such as click, exposure, or swipe) corresponding to each data collection tracking point. If a match is successful, the service item attribute features and interaction action features are combined and encapsulated to generate a semantic structure unit containing the operation object and operation nature; if a match fails, the unit is discarded or marked as an unknown type according to a preset default strategy.

[0041] S104: Extract service item attribute features and interaction action features from each semantic structural unit, and encode the service item attribute features and interaction action features alternately over time to construct an interleaved behavior sequence that represents the user's intention evolution trajectory;

[0042] Service item attribute features refer to the set of information characterizing the inherent attributes of specific service objects (such as goods, financial products, articles, videos, etc.) within a service platform. In this embodiment, these features typically include discrete features (such as service item ID, category ID, tag ID) and continuous features (such as price, rating, popularity value). Before entering the model, discrete features are usually transformed into dense vectors through an embedding layer.

[0043] Interactive action features refer to a set of information characterizing the specific operational behaviors performed by a user on the aforementioned service objects. These features may include explicit action types (such as click, exposure, favorite, share), and in some embodiments, may also include contextual attributes of the action (such as dwell time, swipe speed, click location coordinates).

[0044] Interleaved coding: refers to a specific method of constructing interleaved sequences. Unlike traditional recommendation models that treat user behavior as a single-dimensional sequence of items, this interleaved coding method treats items and actions as two independent tokens, arranged alternately according to causal logic. For example, the sequence structure is presented as: "Item 1 - Action 1 - Item 2 - Action 2...".

[0045] The user intent evolution trajectory refers to the dynamic migration process of user needs over time, as presented by the aforementioned interlaced sequence. For example, from browsing multiple similar products to clicking on a specific product and then to adding it to favorites, this trajectory implicitly expresses the evolution of user intent from vague to explicit, and from superficial to profound.

[0046] Indicatively, by decoupling features from each semantic structural unit, service item attribute features and interaction action features are extracted. A pre-defined feature dictionary is used to convert the discrete feature IDs into corresponding integer indices, which are then searched in pre-trained service item embedding matrices and action embedding matrices to obtain the corresponding low-dimensional dense embedding vectors. Based on this, interleaved coding is performed according to the chronological order of the user's actions. Specifically, for interaction events within the same time step, the service item embedding vector corresponding to that time step is first placed in the sequence, followed by the interaction action embedding vector, and so on, constructing an interleaved arrangement structure in the form of service item vector – action vector – service item vector – action vector…

[0047] Optionally, if the length of the generated interleaved behavior sequence exceeds a preset threshold (e.g., the most recent 50 interactions), a truncation operation is performed to retain the most recent behavioral features in time; if the length is insufficient, it is padded with zero vectors, and finally a fixed-length interleaved behavior sequence is output for subsequent calls.

[0048] S106: Inject the features of candidate service items to be recommended into the interleaved behavior sequence to obtain a target perception transduction sequence, and perform behavioral intention reasoning on the target perception transduction sequence to obtain target perception intention features;

[0049] Candidate service item features refer to the set of feature vectors of the specific object (such as a specific service product, platform service, news article, etc.) that is currently being recommended to the target user. In this embodiment, the feature is usually composed of the unique identifier and key attributes of the service item after vectorization mapping.

[0050] A target-aware transduction sequence is a novel data structure formed by appending candidate service item features to the end of a historical interleaved behavior sequence through specific positions. Unlike static sequences that only record the past, this target-aware transduction sequence explicitly establishes a causal temporal link between historical behavior and the current target, aiming to trigger generative models to transduce and infer the intent of a specific target.

[0051] Behavioral intent reasoning refers to the process of encoding and calculating the aforementioned target perception transduction sequence using deep learning models (such as hierarchical sequence transduction units based on attention mechanisms, HSTU). The core of this process lies in calculating the attention weight or gating dependency of each historical semantic unit in the sequence to the last candidate service item, thereby calculating the user's potential preference strength for the candidate target in the current context based on the attention weight or gating dependency.

[0052] Target-aware intent features refer to the high-dimensional hidden state vectors output by machine learning models (such as hierarchical sequence transduction units based on attention mechanisms) at the sequence position corresponding to the candidate service item (usually the last time step) after behavioral intent reasoning. These high-dimensional hidden state vectors can characterize the semantic correlation between the user's historical behavior trajectory and the current candidate service item.

[0053] To illustrate, a specific candidate service item (e.g., a particular video) is obtained for recommendation, and its feature vector is extracted. A last-position injection operation is then performed, appending the feature vector of the candidate service item directly to the end of the interleaved behavior sequence as a new sequence element. At this point, the sequence length changes from L to L+1, and the semantic role of this new element is defined as "the next interactive object to be predicted." To maintain the integrity of the sequence structure (i.e., content-action alternation), a pre-defined action placeholder vector is added to this position, representing an unknown future action.

[0054] Then, this target-aware transduction sequence, including candidate service items, is input into a pre-defined hierarchical sequence transduction model (such as a deep learning model, neural network model, or large language model, etc.). The model uses a self-attention mechanism to calculate the attention weights of all historical behavioral elements in the sequence for the last candidate service item, thereby capturing the deep dependency between the user's historical interests and the current candidate target. Finally, the hidden state vector output by the hierarchical sequence transduction model at the last position of the sequence is extracted and used as a target-aware intent feature representing the user's potential intent for the candidate service item.

[0055] S108: When the preset service recommendation model is invoked to recommend candidate services to the target user, the target perceived intent feature is inserted into the feature input encoding layer of the preset service recommendation model, and the candidate service item is evaluated by the preset service recommendation model to obtain a predicted service recommendation score.

[0056] A preset service recommendation model refers to a machine learning or deep learning model deployed on the server side for accurately scoring and ranking candidate service items. In this embodiment, the preset service recommendation model can be any pre-trained machine learning model used for service recommendation in any related technology, and is not limited here. Its input format is not limited and can cover multi-source features such as user profiles, item attributes, and contextual environment. Its output format is usually a specific probability value or ranking score for the candidate service recommendation item.

[0057] The feature input encoding layer can be understood as the first functional layer in the preset service recommendation model responsible for transforming the original features into computable vectors within the model.

[0058] Predictive service recommendation score refers to the numerical value output by the model that characterizes the probability of a target user generating a specific interactive behavior (such as clicking, purchasing, or applying) with candidate service items. This value is usually between 0 and 1, such as click-through rate (CTR), or it can be a real-number ranking score used to determine whether the service item is shown to the user and in what order.

[0059] In one feasible implementation, when the platform service recommendation coordination requires scoring and evaluating candidate service items, a preset service recommendation model is invoked. This model typically includes a feature input encoding layer to receive feature inputs from different domains. Then, the target-aware intent feature is obtained. Considering that this target-aware intent feature usually has a high semantic dimension (e.g., 256 or 512 dimensions), while other sparse features in the recommendation model (such as user age, address, and device type) may have lower dimensions after embedding layer mapping, the target-aware intent feature is usually first dimensionally aligned using a linear projection layer or fully connected layer of the preset service recommendation model to ensure consistency with the model's internal feature space scale. The preset service recommendation model then flattens and concatenates the aligned target-aware intent feature with the attribute feature vector of the current candidate service item, the profile feature vector of the target user, and the feature vector of the current context environment, constructing a dense input vector containing multi-source information. The dense input vector is transmitted to the multilayer perceptron component of the pre-defined service recommendation model. Utilizing the stacked hidden layers and non-linear activation functions (such as ReLU and GeLU) within the multilayer perceptron component, high-order cross-extraction is performed on user intent features and other basic features to uncover implicit association patterns such as "enhanced user preference for specific services under specific intent." Finally, through the logistic regression unit or Softmax layer of the pre-defined service recommendation model, a value between 0 and 1 is output, representing the predicted service recommendation score that characterizes the target user's probability of clicking or converting to the candidate service.

[0060] In the embodiments of this specification, the original unstructured logs are parsed into semantic structural units containing both service item content and interactive actions by adopting a preset interactive data collection and tracking system. A behavioral sequence with alternating content and actions is constructed along the time dimension. Furthermore, the features of candidate service items to be recommended are injected into the sequence to trigger target perception reasoning in a hierarchical sequence transduction model. Finally, the generated dynamic intent features are integrated into a preset recommendation model for evaluation. This addresses the limitations of related technologies, such as coarse granularity in user behavior feature mining, inability to distinguish the differences in the contribution of different interactive actions (e.g., clicks and views) to intent, and the lack of targeted reasoning due to the decoupling of user historical interest representation and the current target to be recommended. It achieves fine-grained capture of the complex evolution trajectory of user intent, enabling the generated intent features to dynamically adapt and perceive different candidate service items, thus improving the prediction accuracy and traffic distribution efficiency of the service recommendation system when facing dynamically changing user needs.

[0061] Optional, please see Figure 2 , Figure 2This is a flowchart illustrating a semantic structure unit collection process as described in this specification. Specifically, the collection of semantic structure units from the original behavior log using a preset interactive data collection point identification system can be performed as follows:

[0062] S202: Determine polymorphic raw behavior event data, including page monitoring events, exposure events, and click events, from the raw behavior log, and extract the tracking identifier name of each event interaction context from the polymorphic raw behavior event data;

[0063] Polymorphic event identification can be understood as the fact that raw behavior logs are often heterogeneous, including various types of user interaction signals. Using pre-defined event filters, at least three core event data categories are extracted from massive logs:

[0064] Page monitoring events indicate that a user has entered a new page environment, and usually carry the page ID and the source of the access.

[0065] An exposure event indicates that a specific service item (such as a product card or article title) has appeared in the user's visible area, and usually carries the service item ID and location index.

[0066] Click events indicate that a user has developed a clear interest in a particular service, typically carrying interaction coordinates and a trigger control ID. These three types of data constitute polymorphic raw behavioral event data, which may differ in data structure (for example, PV logs are wide table structures, while exposure logs are nested array structures), but can be standardized into a unified event object.

[0067] The extraction of event tracking identifier names involves using regular expressions or a JSON parser to extract the core field, the event tracking identifier name, from each standardized event object. This name is a key-value pair in the interactive data collection event tracking identifier system.

[0068] S204: Perform linear temporal reorganization of the embedded point identifier names according to the event occurrence time to construct an original behavior sequence representing the trajectory of user intent changes;

[0069] Linear time-series reordering can be understood as a data processing logic used to handle event data from different log sources (such as client-side real-time streams, server-side log files, and database change logs). Due to network latency or system architecture, the physical time when these logs arrive at the processing system may differ from their logical time of occurrence. This step rearranges the out-of-order data using a logical clock or event occurrence time, and serializes parallel events (such as simultaneously exposing multiple products) to form a unidirectional timeline.

[0070] The original behavior sequence refers to the intermediate data structure generated after reorganization but before semantic feature mapping. It consists of a list or queue of tracking point identifiers ordered by time. At this stage, the elements in the sequence are still engineered codes and have not yet been transformed into business semantics, but their order can already reflect the sequential dependencies of user operations.

[0071] User intent evolution trajectory refers to the psychological changes hidden behind the original behavioral sequence. For example, the intent evolution trajectory corresponding to the sequence [Homepage PV - List Page PV - Details Page - Order Click] is [Browsing without a specific purpose - Generating category interest - Generating specific product intent - Establishing a purchase decision]. The purpose of constructing this sequence is to capture user intent evolution.

[0072] Intuitively, a temporary sequence construction buffer is created in memory or distributed storage, using the target user's unique identifier as the aggregation key. All polymorphic raw behavioral event data is then imported into this buffer. Since page view (PV) events, exposure events, and click events may be reported by different server interfaces, their timestamp formats may differ by milliseconds or microseconds (e.g., some accurate to 13 digits, others to 10 digits). First, a timestamp normalization operation is performed, converting the occurrence time of all events into a standard long integer millisecond timestamp. A stable sorting algorithm (such as merge sort) is then called to sort all event identifiers in the buffer in ascending order according to the normalized timestamps. After sorting, the sequence is segmented into sessions, i.e., the time difference between two adjacent events is calculated. If the time difference exceeds a preset session threshold (e.g., 30 minutes), the current sequence is truncated and treated as an independent user intent transition trajectory, thus outputting multiple raw behavioral sequences.

[0073] In one feasible implementation, the linear temporal reorganization of the event tracking identifier names according to the event occurrence time to construct the original behavior sequence representing the trajectory of user intent changes can refer to the following steps:

[0074] Step A2: Determine the original interactive behaviors triggered by the target user at different logical locations on the platform, and extract the timestamps corresponding to each of the original interactive behaviors;

[0075] The different logical levels of the platform refer to the topological layering of the service platform in the front-end display architecture. For example, this may include: container layer, page layer, block layer, slot layer, etc.

[0076] Step A4: Based on the timestamp, aggregate, sort, and string together the tracking point identifiers of the page monitoring events, exposure events, and click events across log sources to obtain the original behavior sequence representing the trajectory of user intent changes.

[0077] Aggregation sorting and behavior chaining refers to aggregating data from the above heterogeneous data sources using user ID as the anchor point, sorting the aggregated results based on a unified timestamp, and connecting the sorted discrete points into a line to form a complete chain reflecting the user's psychological decision-making process.

[0078] To illustrate, a distributed computing framework is used to scan multiple raw fact tables stored in a distributed file system in parallel. These tables store page monitoring event logs recorded by the server, exposure stream logs recorded by the recommendation engine, and click event logs reported by the client SDK. Since the formats of the timestamps recorded by different log sources may differ (e.g., a mix of ISO8601 string timestamps and Unix timestamps), a timestamp normalization operation is performed while reading the data, converting the occurrence time of all events into a standard long integer millisecond timestamp. Then, using the unique identifier of the target user as the sharding key, all cross-source behavioral data of that user, scattered across different data nodes, are aggregated into an in-memory list on the same computing node. Based on this, a quicksort algorithm is called to strictly sort the events in the list in ascending order based on the normalized timestamps, and the unique request ID (Request_ID) in the logs is used to remove duplicate data caused by network retries. Finally, the sorted event list is traversed, and the corresponding tracking point identifier name for each event is extracted and linearly concatenated to generate a raw behavioral sequence in the form of identifier A – identifier B – identifier C.

[0079] S206: Based on the mapping rules in the interactive data collection and tracking point identification system, the tracking point identification names in the original behavior sequence are combined and mapped with the corresponding action behavior type identifications to generate the semantic structure unit composed of the service item attribute features and interactive action features.

[0080] Mapping rules refer to predefined logical protocols or lookup tables in the interactive data collection tracking point identification system, used to convert tracking point identifier names (such as SPM codes) into specific service item business attributes. These rules typically include multi-dimensional relationships such as "identifier-ID mapping", "identifier-category mapping", and "identifier-attribute mapping".

[0081] Action / behavior type identifiers refer to the classification labels after normalizing user operation events recorded in the original logs.

[0082] Combination mapping refers to a data fusion operation based on context-based feature binding. Specifically, it aligns and encapsulates service item attribute features and interaction action features at the same time step, forming an indivisible feature tuple.

[0083] In a schematic manner, a pre-defined global semantic mapping table is loaded into memory. This table stores all valid event tracking identifiers and their corresponding service item metadata in key-value pairs. The original behavior sequence is traversed, and for each event tracking identifier in the sequence, it is used as the query key to retrieve data from the global semantic mapping table. If a match is found, the static service item attribute features corresponding to that event tracking identifier are obtained (such as service item ID, first-level / second-level category ID, service item tag, etc.). Simultaneously, the original event type corresponding to that node is parsed, and standardized interaction action features are generated through a pre-defined action normalization logic (e.g., mapping all tap events to click type identifiers). A combination operation is then performed to encapsulate the retrieved static service item attribute features and the normalized interaction action features into a single data structure, serving as the semantic structural unit for that time step. If no corresponding key-value pair is found in the mapping table, the node is marked as an unknown service item or directly removed from the sequence according to a pre-defined fault tolerance strategy.

[0084] In the embodiments of this specification, the above-described method achieves accurate transformation from multi-source heterogeneous raw behavior logs to standardized semantic structure units. Specifically, polymorphic event extraction and linear temporal reorganization overcome data disorder and clock asynchrony caused by log reporting delays and differences between the client and device, ensuring the causal logical integrity of user interaction trajectories in the time dimension. Furthermore, based on a pre-defined identifier system, a combination mapping mechanism translates the instrumentation code, which only has engineering significance, into semantic features containing rich business attributes and action properties, reducing noise interference in the original data and improving the robustness of intent reasoning.

[0085] Optional, please see Figure 3 , Figure 3 This is a flowchart illustrating an intent reasoning process proposed in this specification. Specifically, the execution of behavioral intent reasoning on the target perception transduction sequence to obtain target perception intent features can be referenced as follows:

[0086] S302: Based on a preset hierarchical sequence transduction model, the target perception transduction sequence is subjected to point-by-point projection processing of multidimensional components to obtain the gating weight component, value vector component, query vector component and key vector component corresponding to the multidimensional components.

[0087] The pre-defined hierarchical sequence transduction model refers to a high-order sequence modeling network built on deep attention mechanisms (such as HSTU and Transformer variants). This pre-defined hierarchical sequence transduction model consists of multiple stacked nonlinear transduction units, used to process interleaved behavior sequences with long temporal dependencies. Its core mechanism is to extract features and spatially map the input sequence data layer by layer through internal multi-head self-attention sublayers and gated modulation sublayers. While capturing the deep semantic association between historical interaction behaviors and current candidate service items, it uses a gating mechanism to dynamically filter out random noise interference in the sequence, and finally transduces the discrete temporal input into a continuous high-dimensional intent feature vector to achieve accurate representation of user dynamic preferences.

[0088] The query vector component (Q) represents the active search signal of "what information is being sought" at the current moment.

[0089] The key vector component (K) represents the index label of "what information is possessed" at a historical moment.

[0090] The value vector component (V) represents the specific content carrier of a historical moment.

[0091] Gated weighted components It is used to calculate the information retention rate at each time step, acting similarly to the forget gate in LSTM.

[0092] Pointwise projection processing refers to using a linear transformation layer to perform matrix multiplication operations independently on each time step in the sequence, without involving the interaction between time steps.

[0093] The point-by-point projection processing of the multidimensional components of the target perception transduction sequence based on the preset hierarchical sequence transduction model can be expressed by the following formula:

[0094]

[0095] in, It is a linear layer. This is the SiLU nonlinear activation function. This operation simultaneously computes the weights used for gating using a fusion kernel. Value vector Query vector components and key vector components This improves computational efficiency.

[0096] This illustration demonstrates the use of a highly efficient parallel computing architecture based on a fusion kernel to perform feature space mapping. Specifically, the system receives the target-aware transduction sequence X generated by the preceding steps as input. Unlike traditional methods that require calling four independent linear layers to generate different components, this embodiment uses a unified linear layer. A one-time linear transformation is performed on the input sequence X, and the transformation formula is denoted as: Where W1 is the preset fusion weight matrix and b1 is the bias vector. A nonlinear activation function is applied to the intermediate result of the linear transformation output. For further processing, the SiLU (Sigmoid Linear Unit) activation function can be used to introduce smooth nonlinearity while preserving gradients. Then, a split operation is performed to decompose the activated high-dimensional mixture tensor along the feature channel dimension, thereby simultaneously separating and outputting four key multidimensional components in a single forward propagation: a gating weight component for dynamically adjusting the information retention ratio, a value vector component carrying the actual semantic content, a query vector component for capturing sequence dependencies, and a key vector component. This approach reduces memory accesses and computational latency, improving the overall computational efficiency of the hierarchical sequence transduction model while ensuring feature extraction accuracy.

[0097] S304: Based on a preset hierarchical sequence transduction model, construct a dot product correlation matrix for the query vector component and the key vector component. After injecting attention bias into the dot product correlation matrix, perform spatial feature aggregation to obtain an aggregated feature vector that characterizes the internal correlation strength of the sequence. Use the gated weight component to perform element-wise modulation processing on the aggregated feature vector to filter noise information in the interleaved behavior sequence, and output target perception intent features based on the preset hierarchical sequence transduction model.

[0098] The dot product correlation matrix refers to the original similarity matrix generated by calculating the product between the query vector component and the transpose of the key vector component;

[0099] (Relative) Attention bias refers to the additional correction term added to the dot product incidence matrix before Softmax normalization. .

[0100] Spatial feature aggregation can refer to the process of weighted summation of value vector components using normalized attention weights. Through this process, the model aggregates fragmented information scattered across different time steps in a long sequence (such as specific product features of historical clicks) into a dense vector containing global contextual information, i.e., the aggregated feature vector.

[0101] Element-wise modulation processing refers to performing a Hadamard product operation with the aforementioned aggregated feature vector using gated weight components. This operation is equivalent to applying a learnable "valve" to each dimension of the feature vector. When the gate value is close to 0, the aggregated features (i.e., noise) of the corresponding dimension are suppressed; when the gate value is close to 1, the effective features are preserved and passed to the next layer.

[0102] Schematic, S304 can be represented as spatial feature aggregation and point-by-point transformation process:

[0103] Spatial feature aggregation:

[0104]

[0105] This represents the dot product incidence matrix. Injection of relative attention bias as an additional correction term The subsequent corrected dot product incidence matrix; A(X) represents the result of applying the corrected dot product incidence matrix to a predefined activation function. The attention weight matrix obtained after (which can be viewed as a point-by-point aggregation attention operator) uses point-by-point aggregation attention instead of the traditional Softmax attention. This mechanism preserves the strong feature of the number of prior data points related to the target, and can better capture the intensity of user preferences, especially suitable for non-stationary dynamic vocabulary (such as continuously added tracking points). It is a relative attention bias that integrates position (p) and time (t) information to inject temporal structure information into the sequence.

[0106] Element-wise modulation processing corresponding to point-wise transformation:

[0107]

[0108] Y(x) represents the target-aware intent feature. For layer normalization, This represents element-wise multiplication, which is also the aggregated feature after layer normalization. A bit-aligned multiplication operation is performed with the gated weight component U(X). In this process, U(X) acts as a dynamic, learnable soft mask or valve: when certain dimensions in the feature vector are determined to be noise generated by user accidental touches or random browsing, the corresponding U(X) value will approach 0, thereby directly erasing (suppressing) these noisy features through multiplication, that is, filtering out noise information in the interleaved behavior sequence; conversely, if it is determined to be a strong intent feature, the corresponding value will approach 1, allowing it to pass.

[0109] After the above gated filtering is completed, the modulated feature vector is input into the function. , This is represented as a linear projection layer responsible for performing the final dimensional mapping and spatial transformation on the modulated features, thereby outputting the final target perception intent feature Y(x) of the sequence transduction unit at this level.

[0110] To illustrate, spatial aggregation is first performed to construct a basic association matrix using the transposes of the query vector components and the key vector components, and then a specific attention bias is injected into it. This bias deeply integrates the relative positional information (p) between sequence elements with the actual time interval information (t) of interaction occurrences, thus explicitly injecting temporal structure features into the association computation. Based on this, a point-by-point aggregation attention mechanism is adopted instead of the traditional Softmax normalized attention, i.e., through the activation function... After processing the correlation matrix, it is directly aggregated with the value vector component V(X). This mechanism can retain the strong feature of the number of prior data points related to the target, thus capturing the strength of user preferences more accurately. It is particularly suitable for processing non-stationary dynamic vocabulary containing continuously added embedding points. Subsequently, a point-by-point transformation operation is performed to perform layer normalization on the feature vector obtained by the above spatial aggregation. The gated weight component U(X) generated in step S302 is used to perform element-by-element multiplication modulation processing on the normalized features, and the modulation result is finally transformed through the linear layer f2 to obtain the output Y(X).

[0111] Optionally, the injection of attention bias into the dot product association matrix can be performed as follows: extract the relative position information of each semantic structural unit in the interleaved behavior sequence, as well as the real time interval information between interactive behaviors; construct an attention bias based on the relative position information and the real time interval information, and inject the attention bias into the dot product association matrix.

[0112] Optionally, the step of performing element-wise modulation processing on the aggregated feature vector using the gated weight components to implicitly filter noise information in the interleaved behavior sequence can be performed in the following manner:

[0113] The aggregated feature vector is normalized to obtain normalized features, and the normalized features are then multiplied element-wise using the gated weight components to implicitly filter behavioral noise.

[0114] This specification addresses the technical challenges of traditional self-attention mechanisms when handling long sequences and non-stationary dynamic vocabulary, such as high computational cost, susceptibility to random interactive noise like accidental touches, and difficulty in quantifying user preference strength. By employing the aforementioned methods, it significantly reduces model memory usage and inference latency while achieving accurate perception of complex temporal dependencies and true intent strength within user behavior sequences. The implicit feature filtering function implemented through a gating mechanism effectively filters redundant noise, improving the purity of target-perceived intent features and the robustness of the recommendation system in dynamic environments.

[0115] Optionally, the specific execution of injecting candidate service item features to be recommended into the interleaved behavior sequence to obtain the target-aware transduction sequence can refer to the following methods:

[0116] The features of candidate service items to be recommended are injected into the end of the interleaved behavior sequence to obtain the target perception transduction sequence.

[0117] Optionally, after obtaining the target-aware transduction sequence, the following methods can also be used:

[0118] Step B2: Construct a shadow perturbation sequence for the interleaved behavior sequence, wherein the shadow perturbation sequence is generated by performing a random masking operation on the interaction action features in the interleaved behavior sequence;

[0119] A shadow perturbation sequence refers to a control sample sequence generated based on a constructed original interleaved behavior sequence by applying controlled noise or minor structural variations. It can be viewed as a twin copy of the original sequence, but differs from the original sequence in certain local features (such as the type or presence of individual actions). During inference, it is not used to directly generate recommendation results, but rather serves as a test probe to detect whether the model is overfitting or overly sensitive to noise.

[0120] Indicatively, a controllable noise injection mechanism is used to construct a contrasting sample, the shadow perturbation sequence, that is highly semantically similar to the original user behavior sequence but differs in microscopic features. Specifically, a complete data copy of the generated original interleaved behavior sequence is cloned. According to a pre-defined robustness detection strategy, a random masking operation is performed: a small perturbation probability threshold (e.g., 10% or 15%) is set, and a mask matrix matching the sequence length and following a Bernoulli distribution is generated. This mask matrix is ​​then used to scan the feature vectors of interactive actions in the sequence bit by bit. For matched action positions, the original high-dimensional embedding vector is replaced with a pre-defined special mask vector. This operation aims to simulate data packet loss and missed detections that may occur in real-world network environments. In this way, a counterfactual user behavior trajectory, the shadow perturbation sequence, is constructed. This sequence retains the main content structure of the original sequence but fine-tunes the intensity and nature of the interactive behavior, thus providing an accurate contrasting input for subsequent evaluation of whether the model is overfitting to a specific operation.

[0121] Step B4: Synchronously infer the contrastive intent features generated by the shadow perturbation sequence using a preset hierarchical sequence transduction model, and calculate the cosine similarity between the target perception intent features and the contrastive intent features;

[0122] The contrastive intent feature refers to the high-dimensional hidden state vector output by the model after the aforementioned shadow perturbation sequence is input into a preset hierarchical sequence transduction model for synchronous inference. It serves as a control group relative to the target perception intent feature. By comparing the distance between these two feature vectors in the vector space, the degree of perturbation of the model output by small changes in the input can be quantified.

[0123] Indicatively, a synchronous inference mechanism for the shadow perturbation sequence is initiated using a pre-defined hierarchical sequence transduction model. During this process, the model keeps all weight parameters constant and performs point-by-point projection, attention bias injection, and gating modulation operations on the shadow perturbation sequence consistent with the original sequence. Since some interactive action features in the input sequence have been randomly masked, the hidden state vector of the output layer will undergo a certain degree of semantic drift after multiple layers of nonlinear transduction. The pre-defined hierarchical sequence transduction model captures this drifted output vector as a contrastive intent feature, and then calculates the cosine similarity between the target perceived intent feature and the contrastive intent feature.

[0124] Step B6: If the cosine similarity is lower than a preset robustness threshold, a penalty factor based on the cosine similarity is generated, and the penalty factor is used to reduce the weight of the target perceived intent feature to obtain the reduced-weight target perceived intent feature.

[0125] When the cosine similarity is detected to be below the robustness threshold, it means that the current user behavior sequence is highly sensitive to small input perturbations, indicating that the intent features extracted by the model may contain overfitting noise. Instead of directly discarding the intent feature, a penalty calculation logic is initiated, generating a scalar penalty factor between 0 and 1 based on the deviation between the current cosine similarity and the robustness threshold. To smoothly control the weighting magnitude, a linear decay function or an exponential decay function is used to map the similarity difference to the penalty factor. For example, the penalty factor can be set to equal the cosine similarity divided by the robustness threshold, or a variant of the sigmoid function can be used to map low similarity to a very small weight close to 0.

[0126] Furthermore, a weighting process is performed, applying the calculated scalar penalty factor to the original target-perceived intent feature vector and performing element-wise scalar multiplication. This operation proportionally compresses the vector's magnitude while maintaining the feature vector's orientation in the high-dimensional semantic space (i.e., the specific category of the user's intent, such as a preference for financial management). After weighting, the numerical strength of the target-perceived intent feature is suppressed. When this feature is subsequently input into the downstream recommendation scoring model (step S108) and subjected to dot product or concatenation operations with candidate service item features, it produces a lower activation value. This mechanism ensures that when the model's robustness is low, the final recommendation ranking score of the candidate service item can be automatically reduced, thereby preventing the model from pushing irrelevant services to the user due to overconfidence, and achieving the safety and robustness of the recommendation system in dynamic noise environments.

[0127] In one feasible implementation, specifically, the step of inserting the target perceived intent features into the feature input encoding layer of a preset service recommendation model, and then using the preset service recommendation model to evaluate the candidate service items to obtain a predicted service recommendation score, can be performed in the following manner:

[0128] Step C2: Determine the linear space dimension of the existing features in the input encoding layer of the preset service recommendation model, perform mapping alignment on the target perception intent features based on the linear space dimension so that the target perception intent features and the existing features have the same representation scale, and concatenate the mapped and aligned target perception intent features with the existing features in the input encoding layer to obtain the enhanced feature vector for the candidate service item.

[0129] The existing features in the input encoding layer refer to the set of basic features in the pre-defined service recommendation model that have not been enhanced with intentional features. These features typically include the user's static profile, the basic attributes of candidate services, and contextual features (such as time and network status). In the model, they are usually processed by the embedding layer and exist in the form of dense vectors.

[0130] Linear space dimension refers to the length of a uniform vector that a recommendation model forces to map sparse features from different sources to in order to perform feature interactions.

[0131] Indicatively, the configuration file or network structure parameters of the preset service recommendation model are read to determine the linear spatial dimension of the input encoding layer used to process common sparse features (such as user ID, geographic location, device type, etc.). Since the target-aware intent features output in step S304 are usually generated by deep networks and have high dimensionality, direct concatenation will lead to an imbalance in feature weights. Therefore, a learnable linear projection layer is constructed as an adapter with the linear spatial dimension of the existing features as the target dimension, performing matrix multiplication and bias addition on the high-dimensional target-aware intent features. This mapping alignment operation not only completes the compression or adaptation of dimensions, but more importantly, it transfers the intent features from their original generative semantic space to a numerical distribution space compatible with the existing features of the recommendation model, giving both the same representation scale. Next, a vector concatenation operation is performed. After feature alignment, the mapped intent feature vector is flattened and concatenated with the embedding vectors of all existing features in the input encoding layer along the feature channel dimension. This operation combines independent feature blocks into a dense, enhanced feature vector containing historical deep intent, current static profile, and real-time environmental context.

[0132] Step C4: Transmit the enhanced feature vector to the multilayer perceptron architecture of the preset service recommendation model, and use the hidden layer of the preset service recommendation model to perform nonlinear cross-extraction calculation on the target perception intent features and the existing features of the input encoding layer to extract the activation value of the last output layer of the multilayer perceptron architecture. The activation value is then mapped to a predicted service recommendation score representing the target user's click probability on the candidate service item through a logistic regression operator.

[0133] Intuitively, the enhanced feature vector is transmitted as an input tensor to a multilayer perceptron architecture within the pre-defined service recommendation model. This architecture typically consists of stacked fully connected layers with decreasing dimensions, coupled to each other via nonlinear activation functions. During the signal propagation layer by layer, the hidden layers utilize learnable weight matrices and bias vectors to perform intensive matrix multiplication on the enhanced feature vector. This process essentially involves nonlinear cross-extraction computation, driving the model to automatically combine and learn higher-order dependencies between the target perceived intent features and the existing features of the input encoding layer within the latent space (e.g., the model can learn complex logic such as significantly increasing the weight of the financial product when the user's intent is a large-amount investment (intent feature) and the user is over 40 years old (existing feature)). After multiple layers of abstraction and feature reorganization, the network outputs unnormalized real values, i.e., activation values ​​(Logits), in the last fully connected layer. Finally, a logistic regression operator (usually the Sigmoid function) is called to spatially map these activation values, compressing them to the (0, 1) interval, thus obtaining a value with probabilistic statistical significance. This value is the predicted service recommendation score, which accurately quantifies the estimated probability that a target user will click or convert on the candidate service under the current intent. The candidate service list can then be sorted and truncated based on the score to complete the final recommendation display.

[0134] The following will combine Figure 4 This specification provides a detailed description of the service recommendation device provided in the embodiments. It should be noted that... Figure 4 The service recommendation device shown is used to perform this instruction manual. Figures 1-3 The methods shown in the embodiments are illustrated for ease of explanation, showing only the parts related to the embodiments of this specification. For specific technical details not disclosed, please refer to this specification. Figures 1-4 The example shown.

[0135] Please see Figure 4 This diagram illustrates the structure of a service recommendation device according to an embodiment of this specification. The service recommendation device 1 can be implemented as all or part of a device through software, hardware, or a combination of both. According to some embodiments, the service recommendation device 1 includes a log processing module 11, a sequence construction module 12, an intent reasoning module 13, and a service recommendation module 14, specifically used for:

[0136] Log processing module 11 is used to obtain the original interaction behavior logs of the target user for the target platform service, and to collect semantic structure units including service item attribute features and interaction action features from the original behavior logs using a preset interaction data collection point identification system.

[0137] The sequence construction module 12 is used to extract service item attribute features and interaction action features from each semantic structure unit, and to alternately encode the service item attribute features and interaction action features over time to construct an interleaved behavior sequence that represents the user's intention evolution trajectory.

[0138] Intent reasoning module 13 is used to inject candidate service item features to be recommended into the interleaved behavior sequence to obtain a target perception transduction sequence, and to perform behavioral intent reasoning on the target perception transduction sequence to obtain target perception intent features;

[0139] The service recommendation module 14 is used to insert the target perceived intent features into the feature input encoding layer of the preset service recommendation model when the preset service recommendation model is invoked to recommend candidate services to the target user, and to obtain a predicted service recommendation score by recommending and evaluating the candidate service items through the preset service recommendation model.

[0140] It should be noted that the service recommendation device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the service recommendation method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the service recommendation device and the service recommendation method embodiments provided in the above embodiments belong to the same concept, and the implementation process can be found in the method embodiments, which will not be repeated here.

[0141] The example numbers in this specification are for descriptive purposes only and do not represent the superiority or inferiority of the examples.

[0142] This specification also provides a computer storage medium that can store multiple instructions adapted to be loaded and executed by a processor as described above. Figures 1-3 The service recommendation method described in the illustrated embodiment can be found in the following documentation for its specific execution process. Figures 1-3 The specific details of the illustrated embodiments will not be elaborated here.

[0143] This specification also provides a computer program product that stores at least one instruction, said at least one instruction being loaded and executed by the processor as described above. Figures 1-3 The service recommendation method described in the illustrated embodiment can be found in the following documentation for its specific execution process. Figures 1-3 The specific details of the illustrated embodiments will not be elaborated here.

[0144] Please refer to Figure 5This is a structural block diagram of an electronic device provided in an embodiment of this specification. The electronic device in this specification may include one or more of the following components: a processor 1010, a memory 1020, an input device 1030, an output device 1040, and a bus 1050. The processor 1010, memory 1020, input device 1030, and output device 1040 may be connected to each other via the bus 1050.

[0145] Processor 1010 may include one or more processing cores. Processor 1010 connects to various parts of the electronic device using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 1020, and by calling data stored in memory 1020. Optionally, processor 1010 may be implemented using at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), or programmable logic array (PLA). Processor 1010 may integrate one or more of a central processing unit (CPU), graphics processing unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into processor 1010 and may be implemented separately through a communication chip.

[0146] The memory 1020 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 1020 may include non-transitory computer-readable storage medium. The memory 1020 may be used to store instructions, programs, code, code sets, or instruction sets.

[0147] The input device 1030 is used to receive input instructions or data, and includes, but is not limited to, a keyboard, mouse, camera, microphone, or touch device. The output device 1040 is used to output instructions or data, and includes, but is not limited to, a display device and a speaker. In this embodiment, the input device 1030 can be a temperature sensor for acquiring the operating temperature of the electronic device. The output device 1040 can be a speaker for outputting audio signals.

[0148] In addition, those skilled in the art will understand that the structure of the electronic device shown in the above figures does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements. For example, the electronic device may also include radio frequency circuits, input units, sensors, audio circuits, wireless fidelity (WIFI) modules, power supplies, Bluetooth modules, etc., which will not be described in detail here.

[0149] In the embodiments of this specification, the executing entity for each step can be the electronic device described above. Optionally, the executing entity for each step can be the operating system of the electronic device. The operating system can be Android, iOS, or other operating systems; this specification does not limit this.

[0150] exist Figure 5 In the electronic device, the processor 1010 can be used to call a program stored in the memory 1020 and execute it to implement the service recommendation method as described in the various method embodiments of this specification.

[0151] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0152] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the features, logs, and information involved in this specification were all obtained under full authorization.

[0153] The above-disclosed embodiments are merely preferred embodiments of this specification and should not be construed as limiting the scope of this specification. Therefore, any equivalent variations made in accordance with the claims of this specification shall still fall within the scope of this specification.

Claims

1. A service recommendation method, characterized in that, Applied to a service platform, the method includes: Obtain the original interaction behavior logs of the target user for the target platform service, and use a preset interaction data collection point identification system to collect semantic structure units including service item attribute features and interaction action features from the original behavior logs; Extract service item attribute features and interaction action features from each semantic structural unit, and encode the service item attribute features and interaction action features alternately over time to construct an interleaved behavior sequence that represents the user's intention evolution trajectory; Injecting candidate service item features to be recommended into the interleaved behavior sequence yields a target perception transduction sequence, and performing behavioral intent reasoning on the target perception transduction sequence yields target perception intent features; When the preset service recommendation model is invoked to recommend candidate services to the target user, the target perceived intent features are inserted into the feature input encoding layer of the preset service recommendation model, and the predicted service recommendation score is obtained by evaluating the candidate service items through the preset service recommendation model. Wherein, the step of performing behavioral intent reasoning on the target perception transduction sequence to obtain target perception intent features includes: The target perception transduction sequence is subjected to point-by-point projection processing of multi-dimensional components based on a preset hierarchical sequence transduction model to obtain the gating weight component, value vector component, query vector component, and key vector component corresponding to the multi-dimensional components. A dot product correlation matrix is ​​constructed between the query vector component and the key vector component based on the preset hierarchical sequence transduction model. After injecting attention bias into the dot product correlation matrix, spatial feature aggregation is performed to obtain an aggregated feature vector that characterizes the internal correlation strength of the sequence. The aggregated feature vector is subjected to element-by-element modulation processing using the gating weight component to filter noise information in the interleaved behavior sequence. Finally, the target perception intent feature is output based on the preset hierarchical sequence transduction model.

2. The method according to claim 1, characterized in that, The method employs a pre-defined interactive data collection and identification system to collect semantic structure units, including service item attribute features and interactive action features, from the original behavior logs, including: The polymorphic raw behavior event data, including page monitoring events, exposure events, and click events, is determined from the raw behavior logs, and the tracking identifier name of each event interaction context is extracted from the polymorphic raw behavior event data; The event tracking identifiers are linearly reorganized according to the event occurrence time to construct an original behavior sequence that represents the trajectory of user intent changes. Based on the mapping rules in the interactive data collection and tracking system, the tracking name in the original behavior sequence is combined and mapped with the corresponding action behavior type identifier to generate the semantic structure unit composed of the service item attribute features and interactive action features.

3. The method according to claim 2, characterized in that, The step of linearly reorganizing the event tracking identifiers according to the event occurrence time to construct an original behavior sequence representing the trajectory of user intent changes includes: Identify the original interactive behaviors triggered by the target user at different logical locations on the platform, and extract the timestamps corresponding to each of the original interactive behaviors; Based on the timestamp, the names of the tracking points for page monitoring events, exposure events, and click events across log sources are aggregated, sorted, and linked together to obtain the original behavior sequence representing the trajectory of user intent changes.

4. The method according to claim 1, characterized in that, The step of injecting attention bias into the dot product association matrix includes: extracting the relative position information corresponding to each semantic structural unit in the interleaved behavior sequence, as well as the real time interval information between interactive behaviors; constructing an attention bias based on the relative position information and the real time interval information, and injecting the attention bias into the dot product association matrix; The step of performing element-wise modulation processing on the aggregated feature vector using the gated weight components to filter noise information in the interlaced behavior sequence includes: normalizing the aggregated feature vector to obtain normalized features, and performing element-wise multiplication processing on the normalized features using the gated weight components.

5. The method according to claim 1, characterized in that, The step of injecting candidate service item features to be recommended into the interleaved behavior sequence to obtain a target-aware transduction sequence includes: The features of candidate service items to be recommended are injected into the end of the interleaved behavior sequence to obtain the target perception transduction sequence.

6. The method according to claim 5, characterized in that, After obtaining the target perception transduction sequence, the process further includes: Construct a shadow perturbation sequence for the interleaved behavior sequence, wherein the shadow perturbation sequence is generated by performing a random masking operation on the interaction action features in the interleaved behavior sequence; The contrastive intent features generated by the shadow perturbation sequence are synchronously inferred using a preset hierarchical sequence transduction model, and the cosine similarity between the target perception intent features and the contrastive intent features is calculated. If the cosine similarity is lower than a preset robustness threshold, a penalty factor based on the cosine similarity is generated, and the penalty factor is used to reduce the weight of the target perceived intent feature to obtain the reduced-weight target perceived intent feature.

7. The method according to claim 1, characterized in that, The step of inserting the target-aware intent features into the feature input encoding layer of a preset service recommendation model, and then using the preset service recommendation model to evaluate the candidate service items to obtain a predicted service recommendation score, includes: The linear space dimension of the existing features in the input encoding layer of the preset service recommendation model is determined. The target perceived intent features are mapped and aligned based on the linear space dimension so that the target perceived intent features and the existing features have the same representation scale. The mapped and aligned target perceived intent features and the existing features in the input encoding layer are concatenated to obtain the enhanced feature vector for the candidate service item. The enhanced feature vector is transmitted to the multilayer perceptron architecture of the preset service recommendation model. The hidden layer of the preset service recommendation model is used to perform nonlinear cross-extraction calculation on the target perception intent features and the existing features of the input encoding layer to extract the activation value of the last output layer of the multilayer perceptron architecture. The activation value is then mapped to a predicted service recommendation score representing the target user's click probability on the candidate service item through a logistic regression operator.

8. A service recommendation device, characterized in that, The device, applied to a service platform, includes: The log processing module is used to obtain the original interaction behavior logs of the target user for the target platform service, and to collect semantic structure units including service item attribute features and interaction action features from the original behavior logs using a preset interaction data collection point identification system. The sequence construction module is used to extract service item attribute features and interaction action features from each of the semantic structural units, and to alternately encode the service item attribute features and interaction action features over time to construct an interleaved behavior sequence that represents the user's intention evolution trajectory. The intent reasoning module is used to inject candidate service item features to be recommended into the interleaved behavior sequence to obtain a target perception transduction sequence, and to perform behavioral intent reasoning on the target perception transduction sequence to obtain target perception intent features; The service recommendation module is used to insert the target perceived intent features into the feature input encoding layer of the preset service recommendation model when the preset service recommendation model is invoked to recommend candidate services to the target user, and to obtain a predicted service recommendation score by recommending and evaluating the candidate service items through the preset service recommendation model. Wherein, the step of performing behavioral intent reasoning on the target perception transduction sequence to obtain target perception intent features includes: The target perception transduction sequence is subjected to point-by-point projection processing of multi-dimensional components based on a preset hierarchical sequence transduction model to obtain the gating weight component, value vector component, query vector component, and key vector component corresponding to the multi-dimensional components. A dot product correlation matrix is ​​constructed between the query vector component and the key vector component based on the preset hierarchical sequence transduction model. After injecting attention bias into the dot product correlation matrix, spatial feature aggregation is performed to obtain an aggregated feature vector that characterizes the internal correlation strength of the sequence. The aggregated feature vector is subjected to element-by-element modulation processing using the gating weight component to filter noise information in the interleaved behavior sequence. Finally, the target perception intent feature is output based on the preset hierarchical sequence transduction model.

9. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions adapted for loading by a processor and executing the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product stores at least one instruction, which is loaded by a processor and executed according to any one of claims 1 to 7.

11. An electronic device, characterized in that, include: A processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the steps of the method as described in any one of claims 1 to 7.

Images