Personalized content recommendation method and platform based on user behavior track

Abstract

The invention discloses a personalized content recommendation method and platform based on a user behavior track, and relates to the technical field of personalized content recommendation, and the method comprises the steps: constructing a user behavior sequence through multi-source data; extracting a short-term interest vector and a long-term interest vector from the user behavior sequence, and weighting to obtain a user interest vector; performing dynamic fusion according to context data and the user interest vector, and performing recommendation calculation to obtain context-aware recommendation; recalling the context awareness recommendation according to the real-time interest vector, and obtaining a context recall awareness recommendation through online reasoning; and performing personalized content recommendation through the context recall perception recommendation. The technical problems that in the prior art, personalized content recommendation is difficult to give consideration to short-term interest changes and long-term preferences of users, contextual adaptability is poor, and recommendation accuracy is insufficient are solved, and the technical effect of improving accuracy, real-time performance and contextual adaptability of personalized content recommendation is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of personalized content recommendation, and in particular to a personalized content recommendation method and platform based on user behavior trajectories. Background Art

[0002] In the field of personalized content recommendation, existing technologies often struggle to effectively integrate short-term fluctuations in user interests and long-term preferences within user behavior trajectories. Furthermore, they lack dynamic adaptability to contextual environments (such as time and scene), resulting in recommendations with low accuracy, a lack of novelty, and a lack of diversity. Furthermore, for cold-start users, the lack of sufficient historical interaction data makes it difficult to quickly generate effective recommendations. Furthermore, recommendation systems lack robustness and self-recovery capabilities in the face of service anomalies, impacting user experience.

[0003] Existing technologies have the technical problem that personalized content recommendations are difficult to take into account users' short-term interest changes and long-term preferences, resulting in poor context adaptability and insufficient recommendation accuracy. Summary of the Invention

[0004] This application provides a personalized content recommendation method and platform based on user behavior trajectories, which is used to solve the technical problem in the existing technology that personalized content recommendation is difficult to take into account users' short-term interest changes and long-term preferences, resulting in poor context adaptability and insufficient recommendation accuracy.

[0005] In view of the above problems, this application provides a personalized content recommendation method and platform based on user behavior trajectories.

[0006] A first aspect of the present application provides a personalized content recommendation method based on user behavior trajectories, the method comprising:

[0007] A user behavior sequence is constructed using multi-source data; short-term interest vectors and long-term interest vectors are extracted from the user behavior sequence and weighted to obtain a user interest vector; context data and the user interest vector are dynamically integrated and recommendation calculations are performed to obtain context-aware recommendations; the context-aware recommendations are recalled based on the real-time interest vectors, and context-recall-aware recommendations are obtained through online reasoning; personalized content recommendations are performed using the context-recall-aware recommendations.

[0008] The second aspect of the present application provides a personalized content recommendation platform based on user behavior trajectories, the platform comprising:

[0009] The user behavior sequence construction module is used to construct a user behavior sequence through multi-source data; the user interest vector acquisition module is used to extract short-term interest vectors and long-term interest vectors from the user behavior sequence and obtain the user interest vector by weight; the recommendation calculation module is used to dynamically fuse the context data and the user interest vector and perform recommendation calculation to obtain context-aware recommendation; the context recall-aware recommendation module is used to recall the context-aware recommendation according to the real-time interest vector and obtain the context recall-aware recommendation through online reasoning; the personalized content recommendation module is used to perform personalized content recommendation through the context recall-aware recommendation.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] The system constructs user behavior sequences from multi-source data; extracts short-term and long-term interest vectors from these sequences and weights them to derive user interest vectors; dynamically integrates contextual data with these interest vectors and performs recommendation calculations to generate context-aware recommendations; recalls these context-aware recommendations based on real-time interest vectors, and generates context-recall-aware recommendations through online inference; and uses these context-recall-aware recommendations to make personalized content recommendations. This approach improves the accuracy, real-time nature, and contextual adaptability of personalized content recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0013] Figure 1 A flowchart of a personalized content recommendation method based on user behavior trajectory provided in an embodiment of the present application;

[0014] Figure 2 A schematic diagram of the structure of a personalized content recommendation platform based on user behavior trajectories provided in an embodiment of the present application.

[0015] Explanation of the accompanying symbols: user behavior sequence construction module 10, user interest vector acquisition module 20, recommendation calculation module 30, context recall perception recommendation module 40, personalized content recommendation module 50. DETAILED DESCRIPTION

[0016] This application provides a personalized content recommendation method and platform based on user behavior trajectories, which is used to solve the technical problem in the existing technology that personalized content recommendation is difficult to take into account users' short-term interest changes and long-term preferences, resulting in poor context adaptability and insufficient recommendation accuracy.

[0017] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, the present application provides a personalized content recommendation method based on user behavior trajectory, the method comprising:

[0019] Step S100: Constructing a user behavior sequence through multi-source data.

[0020] Specifically, by collecting the target user's explicit feedback data (such as ratings, likes, favorites, etc.) and implicit feedback data (such as browsing history, stay time, scroll depth, etc.), multi-source feedback data is obtained, and the context data corresponding to the multi-source feedback data (including time, device type, network environment, geographic location, etc.) is obtained at the same time; then, anomaly detection is performed on the multi-source feedback data and context data, abnormal data is filtered out, and the data is sequence-normalized to construct a triple data set of user-behavior sequence-context; finally, context encoding is performed based on the triple data set to output a user behavior sequence that can represent the user behavior trajectory.

[0021] Step S200: extracting a short-term interest vector and a long-term interest vector from the user behavior sequence, and weighting them to obtain a user interest vector.

[0022] Specifically, we first extract multiple interaction events within a preset sliding window from the user behavior sequence, and then use the gated recurrent unit (GRU) to capture the short-term dependencies of these interaction events to complete the modeling of short-term interest vectors; at the same time, we extract all interaction events in the user behavior sequence, and use the self-attention mechanism to capture the long-term dependencies of all interaction events to achieve the modeling of long-term interest vectors; finally, we weight the short-term interest vectors and long-term interest vectors according to the time decay mechanism to obtain a user interest vector that can comprehensively reflect the user's interest.

[0023] Step S300: dynamically integrating the context data and the user interest vector and performing recommendation calculation to obtain context-aware recommendations.

[0024] Specifically, the context data (including time, device type, network environment, usage scenario, etc.) is first encoded and processed to generate the corresponding context vector; the user interest vector and the context vector are spliced ​​together to form a user feature vector, and the importance score of the context in the feature vector is calculated through a gating mechanism, which is used as the context weight; the user interest vector is scaled using the weight to obtain the user dynamic interest vector; the feature information of the project (such as content category, tag, popularity, release time, etc.) is extracted, the user dynamic interest vector is matched with the project features, and the user-project matching degree is calculated; finally, based on the user-project matching degree, multi-objective optimization sorting is performed in combination with the project novelty and project diversity to obtain the context-aware recommendation result.

[0025] Step S400: recalling the context-aware recommendation according to the real-time interest vector, and obtaining the context-recall-aware recommendation through online reasoning.

[0026] Specifically, the short-term interest vector in the preset sliding window is updated according to real-time interaction events, and the long-term interest vector is incrementally learned to obtain the real-time interest vector through time decay weighting (if the target user is a cold start user, the group popular interest vector is generated based on its cold start information as the preset default interest vector); then the real-time interest vector is context-adapted, and the candidate set is recalled from the context-aware recommendation to obtain the recalled recommendation; then the item features of the recalled recommendation are extracted, and the user-recalled item matching degree is calculated by matching the real-time interest vector with the recalled item features; finally, based on the matching degree, multi-objective optimization sorting is performed in combination with the novelty and diversity of the recalled items, so as to obtain the context-recall-aware recommendation through online reasoning.

[0027] Step S500: Perform personalized content recommendation through the context recall-aware recommendation.

[0028] Specifically, the context-recall-aware recommendation results after multi-objective optimization and sorting are output to the user terminal to complete personalized content display; at the same time, the recommendation results are cached to the storage module, and the service response time, error rate and other availability indicators are monitored in real time. When the preset monitoring threshold is triggered (such as the number of response timeouts exceeds 5 times / minute), the hierarchical degradation mechanism is automatically started (giving priority to the core recommendation function), and the user-unaware service switching strategy is used to continuously try to restore the main service. After the fault is eliminated, it gradually switches back to the normal recommendation mode to ensure the stability and continuity of the recommendation service.

[0029] In one possible implementation, step S100 further includes:

[0030] Step S110: Obtain multi-source feedback data by collecting explicit feedback data and implicit feedback data of the target user.

[0031] Step S120: Acquire context data of the multi-source feedback data.

[0032] Step S130: performing anomaly detection and filtering on the multi-source feedback data and the context data, and performing sequence normalization to construct a triplet data set of user-behavior sequence-context.

[0033] Step S140: performing context encoding according to the triplet data set, and outputting the user behavior sequence.

[0034] Specifically, by collecting the target users' explicit feedback data on the platform (such as active ratings, likes, favorites, comments, and other behavioral data that directly express preferences) and implicit feedback data (such as browsing history, page dwell time, scrolling operations, click locations, and other behavioral data that indirectly reflect interests), we integrate and form multi-source feedback data to comprehensively capture users' behavioral characteristics and interest tendencies.

[0035] Obtain contextual data corresponding to multi-source feedback data, specifically including the time information when the target user generates explicit feedback data and implicit feedback data (such as the specific date and time period of the operation), device environment (such as the terminal type used is a mobile phone, tablet or computer, operating system version, etc.), network status (such as 4G, 5G or Wi-Fi network), geographic location (such as the city and region), and usage scenarios (such as work scenarios, leisure scenarios, etc.). By obtaining this contextual data, we can provide scenario-based background information for subsequent analysis of user behavior.

[0036] Statistical analysis methods (such as the 3σ principle) are used to detect anomalies in multi-source feedback data (including explicit feedback ratings and likes, and implicit feedback browsing time and click traces) and contextual data (timestamps, device IDs, network types, etc.). Numerical data thresholds (such as browsing time <500ms, ratings >5 points or <1 point) and logical data rules (such as device ID is empty, geographic location coordinates are outside the normal range) are set, and abnormal data is eliminated through filtering algorithms. Data cleaning tools are used to perform sequence standardization on valid data, including unified timestamp formatting, missing value interpolation filling (such as using mean / mode filling), and one-hot encoding of categorical data. Finally, user identification, time-ordered behavioral event sequence (such as interaction type, content ID, operation time), and context feature vector are combined into a triple data set, which is stored in a distributed database for subsequent call.

[0037] For the constructed user-behavior sequence-context triplet dataset, a deep learning encoding model (such as the Transformer encoder) is used to vectorize the context data (including scene features such as time, device, and network environment), mapping the unstructured context information into a low-dimensional dense context vector; at the same time, the interactive events in the behavior sequence (such as browsing, liking, and collecting operations) are temporally encoded, and the context vector is combined with the temporal features of the behavior event to generate a user behavior sequence containing user behavior trajectory and scene information, providing structured input for subsequent interest vector extraction.

[0038] In one possible implementation, step S200 further includes:

[0039] Step S210: extracting multiple interaction events of a preset sliding window according to the user behavior sequence.

[0040] Step S220: Capturing short-term dependencies of the multiple interaction events based on the gated recurrent unit to complete short-term interest vector modeling.

[0041] Step S230: extracting all interaction events according to the user behavior sequence.

[0042] Step S240: Capture the long-term dependencies of all the interaction events through the self-attention mechanism to complete the long-term interest vector modeling.

[0043] Step S250: weighting the short-term interest vector and the long-term interest vector by time attenuation to obtain the user interest vector.

[0044] Specifically, multiple interaction events of a preset sliding window are extracted based on the user behavior sequence. Specifically, according to the preset time window size (such as the last 1 hour, 24 hours, etc.) or event quantity window (such as the last 50 or 100 interaction records), a certain time range or a certain number of interaction events are intercepted from the user behavior sequence to obtain the user's recent behavior data, which provides input for the subsequent short-term interest vector modeling, thereby capturing the user's recent interest changes and behavior patterns.

[0045] The Gated Recurrent Unit (GRU) captures short-term dependencies between multiple interaction events extracted within a preset sliding window. Specifically, this involves chronologically inputting interaction event sequences (such as browsing, clicking, and favorites, along with their timestamps and content features) into the GRU network. The GRU's reset and update gate mechanisms dynamically adjust the degree of historical state retention. The reset gate determines the proportion of historical information to be ignored, while the update gate controls the influence of the previous state on the current state. By iteratively updating the hidden layer state, the GRU captures the temporal dependencies between recent interaction events (such as the behavior of users continuously browsing similar content). Ultimately, the hidden state at the last time step is used as a short-term interest vector, which encodes the user's interest preferences and temporal correlations in recent behavior.

[0046] Extract all interaction events based on the user behavior sequence. Specifically, it means completely extracting all interaction events generated by users on the platform from the user behavior sequence, including all explicit feedback behaviors (such as ratings, likes, favorites, comments, etc.) and implicit feedback behaviors (such as browsing history, page dwell time, scrolling operations, click tracks, etc.) from the first use of the platform to the present. No time window or event number limit is set, so as to obtain the user's full-cycle behavior data, provide comprehensive behavior trajectory information for the subsequent long-term interest vector modeling, and thus capture the user's long-term and stable interest characteristics.

[0047] The self-attention mechanism captures the long-term dependencies of all interaction events. All interaction events extracted from user behavior sequences (including information such as content features, operation types, and timestamps) are converted into vector sequences. The self-attention mechanism then calculates the attention weights between each interaction event and all other events. These weights reflect the degree of influence of different events on user interests. By aggregating information from all events through a weighted summation, the model focuses on historical behaviors that are distant from the current time but have a significant impact on user interests (such as the association between a certain type of content collected six months ago and current browsing behavior). This allows the model to capture the semantic dependencies and temporal associations between long-distance interaction events. Ultimately, the aggregated vector is used as the long-term interest vector, which represents the user's long-term stable interest preferences and behavioral patterns across time.

[0048] Determine the time decay function, such as using an exponential decay function to calculate the weight of the short-term interest vector, so that the closer the interaction event is to the current time, the greater the weight of the short-term interest vector, and as time goes by, the weight decreases exponentially. For the long-term interest vector, considering that it reflects the user's more stable interests, it can be given a relatively stable weight based on the time distribution or importance of historical interaction events. Then, the short-term interest vector and the long-term interest vector are multiplied by their respective weights, and then linearly combined to obtain a user interest vector that comprehensively considers the user's recent behavior and long-term preferences. For example, assuming that the weight of the short-term interest vector is the decay value calculated based on the time of the most recent interaction event, the weight of the long-term interest vector is 1 minus the decay value. Through such a weighting method, the user interest vector can reflect the current interest changes while retaining long-term interest characteristics.

[0049] In one possible implementation, step S300 further includes:

[0050] Step S310: Context-encoding the context data to obtain a context vector.

[0051] Step S320: using a gating mechanism to calculate the degree of adjustment of the context vector to the user interest vector, and performing scaling to obtain a user dynamic interest vector.

[0052] Step S330: extracting item features, matching the item features with the user dynamic interest vector, and obtaining a user-item matching degree.

[0053] Step S340: performing multi-objective optimization sorting based on the user-item matching degree, combined with the item novelty and item diversity, to obtain the context-aware recommendation.

[0054] Specifically, the context data is context-encoded to obtain a context vector. The context data of the multi-source feedback data obtained, such as time, device type, network environment, geographic location and other information, is converted into a structured, low-dimensional dense numerical vector representation through a specific encoding method, such as the encoding model in deep learning (such as the Transformer encoder), thereby obtaining a context vector that can represent the context features, providing a basis for subsequent fusion with the user interest vector and recommendation calculation.

[0055] The user interest vector and the context vector are concatenated into a user feature vector. The context importance score of the feature vector is calculated through a fully connected layer and used as the context weight, which reflects the influence of the current context on the user's interest. The user interest vector is then scaled according to the weight. If the context weight is high, the regulatory effect of the context on the user's interest is enhanced, otherwise it is weakened. Finally, a user dynamic interest vector that integrates context information is obtained, which enables dynamic adjustment of user interests in different scenarios.

[0056] Extract project features and match them with user dynamic interest vectors to obtain user-project matching. Extract target project features from the content library, such as keywords, category labels, and semantic vectors of text content, visual features and audio features of multimedia content, as well as structured and unstructured features such as project release time, popularity rating, and interaction data. Convert these features into project feature vectors of the same dimension as the user dynamic interest vectors. Then, calculate the matching degree between the user dynamic interest vectors and the project feature vectors through cosine similarity calculation to obtain a quantitative user-project matching degree, which reflects the user's potential interest level in a specific project in the current context.

[0057] Context-aware recommendations are achieved through a multi-objective optimization ranking based on user-item match, combined with item novelty and item diversity. Specifically, user-item match is used as the baseline metric, while item novelty (such as the number of days since the item was published and the inverse of the number of times it has been recommended) and item diversity (such as the coverage of the item's category and the difference between its feature vector and recommended items) are incorporated into a multi-objective optimization function. By linearly combining these metrics with weighting coefficients (e.g., 0.6 for match, 0.2 for novelty, and 0.2 for diversity), or by adopting a Pareto optimal strategy to address trade-offs between these objectives, the project collection is finally ranked, resulting in a context-aware recommendation list that not only meets the user's current interests but also exhibits a certain degree of novelty and diversity, thereby satisfying user preferences while expanding their content exposure.

[0058] In one possible implementation, step S320 further includes:

[0059] Step S321: concatenate the user interest vector and the context vector to form a user feature vector.

[0060] Step S322: Calculate a context importance score for the user feature vector as a context weight.

[0061] Step S323: scaling the user interest vector according to the context weight to obtain the user dynamic interest vector.

[0062] Specifically, the user interest vector and the context vector are concatenated to form a user feature vector. The user interest vector, extracted from the user behavior sequence and weighted by time decay, is concatenated with the context vector generated by encoding the context data in the feature dimension. For example, if the dimension of the user interest vector is D1 and the dimension of the context vector is D2, the dimension of the concatenated user feature vector is D1+D2. In this way, the user's interest features and the context features of the current scene are fused into a unified feature representation, providing comprehensive feature input for the subsequent calculation of the degree to which the context regulates the user's interest.

[0063] The concatenated user feature vector (which combines the user interest vector and the context vector) is input into a neural network model (such as a multi-layer perceptron). Through linear transformations of the hidden layer weight matrix and bias term, combined with an activation function (such as sigmoid or ReLU), nonlinear mapping is performed to output a value in the range [0, 1]. During this process, the model automatically learns the associated weights between contextual features and user interest features. For example, by optimizing parameters through a backpropagation algorithm, the output score accurately represents the degree to which the context influences user interest. A higher score indicates a more significant impact of the current context on user interest. Ultimately, this score is used as the context weight for subsequent scaling of the user interest vector.

[0064] The calculated context weight is element-wise multiplied by the user interest vector, and the weights are used to adjust the dimensions of the user interest vector. A large context weight indicates that the current context significantly influences the user's interest, and the user interest vector is enhanced to highlight the role of context. A small weight indicates that the user interest vector is less affected by the context and remains relatively stable. Ultimately, this scaling operation generates a dynamic user interest vector that dynamically reflects the current context, enabling the user interest model to adapt to different scenarios.

[0065] In one possible implementation, step S400 further includes:

[0066] Step S410: Read the real-time interest vector of the target user from the cache.

[0067] Step S420: performing context adaptation on the real-time interest vector, and recalling a candidate set for the context-aware recommendation to obtain a recalled recommendation.

[0068] Step S430: extracting the recalled item features for the recalled recommendation, matching the recalled item features with the real-time interest vector, and obtaining a user-recalled item matching degree.

[0069] Step S440: performing multi-objective optimization sorting based on the user-recalled item matching degree, combined with the recalled item novelty and recalled item diversity, to obtain the contextual recall-aware recommendation.

[0070] Specifically, the real-time interest vector of the target user is read from the cache, that is, the latest interest vector of the target user is obtained from the system cache. The short-term interest vector is updated with a preset sliding window through real-time interaction events, and the long-term interest vector is incrementally learned. It is then generated after time attenuation weighting, which can reflect the user's current interest status.

[0071] The attention mechanism is used to perform weighted fusion of the real-time interest vector and the context vector (such as time, device, scene tags, etc.). The weight distribution of the real-time interest vector is adjusted by calculating the correlation of the features of each dimension to achieve context adaptation. At the same time, an inverted index or vector retrieval technology (such as FAISS) is used to retrieve items with a cosine similarity higher than a threshold with the adapted real-time interest vector in the context-aware recommendation results. Combined with Bloom filter deduplication, a recall recommendation list containing the Top-K similar items is generated to ensure that the recalled items are both in line with the real-time interest and have recommendation efficiency.

[0072] A convolutional neural network is used to extract deep features such as text semantics and visual features from the recalled recommended items. The attention mechanism is used to weight the item's category label, keywords, user historical interaction weight and other features to generate an item feature vector. At the same time, the real-time interest vector is mapped to the same feature space through a multi-layer perceptron (MLP). The cosine similarity algorithm is used to calculate the semantic matching degree between the two, capturing the high-order interaction relationship between the features. Finally, the user-recalled item matching degree is output to quantify the degree of fit between the user's real-time interest and the recalled item features.

[0073] Ranking is achieved through a multi-objective optimization mechanism. A deep model is first used to calculate the degree of match between users and recalled items. Novelty is then assessed based on the item's historical exposure (fewer exposures, more novelty). Diversity is also measured based on the item's category coverage (the more diverse the category, the better). These three metrics are fed into the optimization model, which then integrates the requirements of each dimension through adaptive weight learning. Finally, recalled items are ranked by their combined scores, generating a contextual recall-aware recommendation list that balances interest alignment, fresh content, and rich categories.

[0074] In one possible implementation, step S410 further includes:

[0075] Step S411: updating the short-term interest vector of the preset sliding window according to the real-time interaction event, and performing incremental learning of the long-term interest vector.

[0076] Step S412: updating the short-term interest vector and the long-term interest vector by time-attenuated weighting to obtain the real-time interest vector.

[0077] Specifically, the short-term interest vector of the preset sliding window is updated according to real-time interaction events, and the long-term interest vector is incrementally learned. When the user generates a new interaction behavior, the latest interaction event in the window is captured based on the preset sliding window size, and these recent events are processed using models such as gated recurrent units, and the short-term interest vector is updated to capture the user's current interest changes. At the same time, for all historical interaction events, the long-term interest vector is incrementally learned through methods such as the self-attention mechanism, so that the long-term interest vector can be gradually optimized with the addition of new data, and the characterization of the user's long-term preferences is continuously improved.

[0078] The real-time interest vector is obtained by updating the short-term interest vector and the long-term interest vector through time decay weighted updating. The updated short-term interest vector and the long-term interest vector are weightedly fused based on the time decay mechanism, that is, different weights are assigned according to the time distance of the interaction event. Recent interaction events have a greater impact on the user's current interest, so a higher weight is assigned to the short-term interest vector. The long-term interest vector reflects the user's historical preference trend, and the influence weight of early data is reduced by the time decay formula. Finally, the two are linearly combined according to the attenuated weights to generate a real-time interest vector that can capture the user's recent interest fluctuations and take into account long-term preferences, so that it is more in line with the user's current real interest status.

[0079] In one possible implementation, step S410 further includes:

[0080] Step S413: If the target user is a cold start user, a preset default interest vector is used, wherein the preset default interest vector generates a group hot interest vector based on the cold start information of the cold start user.

[0081] Step S414: performing context adaptation and item feature matching based on the preset default interest vector, outputting cold start recommendations and performing cold start exploration, and adding the recommendations to the context-aware recommendations.

[0082] Specifically, when the target user is a cold-start user, by extracting their cold-start information (such as age, gender, preference tags in the registration information, or group classification characteristics), counting popular features such as content categories and keywords with high interaction frequency in similar groups, and vectorizing and aggregating these group popular features (for example, calculating the frequency of occurrence or weighted mean of popular features), a preset default interest vector is generated, which is used as the initial interest representation of the cold-start user to ensure that recommendations can be made based on group commonalities when individual historical behavior data is lacking.

[0083] The preset default interest vector is concatenated with the current contextual data (such as the recommended time period, device environment, and scene tags). This is then fed into a fully connected layer for feature transformation to achieve contextual adaptation, aligning the interest vector with the recommended scene semantics. Cosine similarity is then used to calculate the match between this vector and the item feature vector (derived from the item's category label, keywords, and popular attributes). A matching threshold is then applied to select popular items within the group for cold-start recommendations. Simultaneously, a certain percentage of items are randomly selected from a pool of novel items (such as recent, low-profile, but high-potential content) and merged with the cold-start recommendations to form a candidate set. Cold-start exploration recommendations are generated through online inference, and ultimately, the cold-start recommendations and exploration results are added to the candidate list for context-aware recommendations.

[0084] In one possible implementation, step S500 further includes:

[0085] Step S510: Cache context-recall-aware recommendations and monitor service availability in real time.

[0086] Step S520: When the monitoring indicator is triggered, continue to try to restore the main service through hierarchical degradation and user-unaware switching, and gradually switch back to normal mode.

[0087] Specifically, the generated context-recall-aware recommendation results are stored in a distributed cache system (such as Redis), and the response efficiency of the recommendation service is improved by setting a reasonable cache expiration strategy (such as time-based or user interaction update). At the same time, a real-time monitoring module is deployed to periodically collect and analyze key service indicators (including interface call time, service error rate, cache hit rate, server load, etc.), and ensure timely detection of service anomalies through alarm threshold configuration (such as triggering an early warning when the response time exceeds 500ms or the error rate is higher than 5%).

[0088] When the real-time monitoring module detects that a service metric (such as an interface response time exceeding a threshold, a spike in error rates, or resource utilization exceeding limits) triggers a preset alarm rule, it immediately initiates a hierarchical degradation mechanism, shutting down non-essential modules (such as the novelty calculation module in multi-objective optimization) in sequence according to pre-defined service priorities (e.g., core recommendation functions > non-core extension functions). At the same time, the load balancer seamlessly routes user requests to a pre-launched backup service cluster to ensure the continuous availability of the recommendation service. In the background, the fault diagnosis system automatically locates anomalies in the primary service and attempts to repair them. As the performance of the primary service gradually recovers, the degraded functional modules are restarted in stages in the order of function activation, health check, and traffic switching back. Traffic is gradually switched back to the primary service through a grayscale release mechanism, ultimately fully restoring normal operation. The entire process ensures that users do not noticeably experience service fluctuations through a service circuit breaker and retry mechanism.

[0089] Example 2 is based on the same inventive concept as the method for personalized content recommendation based on user behavior trajectory in the above embodiment. Figure 2 As shown, this application provides a personalized content recommendation platform based on user behavior trajectory. The platform and method embodiments in this application are based on the same inventive concept. The platform includes:

[0090] The user behavior sequence construction module 10 is used to construct a user behavior sequence through multi-source data.

[0091] The user interest vector acquisition module 20 is configured to extract a short-term interest vector and a long-term interest vector from the user behavior sequence, and weight them to obtain a user interest vector.

[0092] The recommendation calculation module 30 is used to dynamically integrate the context data and the user interest vector and perform recommendation calculation to obtain context-aware recommendations.

[0093] The context recall-aware recommendation module 40 is configured to recall the context recall-aware recommendation according to the real-time interest vector and obtain the context recall-aware recommendation through online reasoning.

[0094] The personalized content recommendation module 50 is configured to perform personalized content recommendation through the context recall-aware recommendation.

[0095] Furthermore, the platform is also used to implement the following functions:

[0096] Multi-source feedback data is obtained by collecting explicit feedback data and implicit feedback data of target users; context data of the multi-source feedback data is obtained; anomaly detection and anomaly filtering are performed on the multi-source feedback data and the context data, and sequence normalization is performed to construct a triple data set of user-behavior sequence-context; context encoding is performed based on the triple data set to output the user behavior sequence.

[0097] Furthermore, the platform is also used to implement the following functions:

[0098] Extract multiple interaction events of a preset sliding window based on the user behavior sequence; capture short-term dependencies of the multiple interaction events based on a gated recurrent unit to complete short-term interest vector modeling; extract all interaction events based on the user behavior sequence; capture long-term dependencies of all interaction events through a self-attention mechanism to complete long-term interest vector modeling; weight the short-term interest vector and the long-term interest vector through time decay to obtain the user interest vector.

[0099] Furthermore, the platform is also used to implement the following functions:

[0100] The context data is context-encoded to obtain a context vector; a gating mechanism is used to calculate the degree of adjustment of the context vector on the user interest vector, and scaling is performed to obtain a user dynamic interest vector; item features are extracted, and the item features are matched with the user dynamic interest vector to obtain a user-item matching degree; based on the user-item matching degree, multi-objective optimization sorting is performed in combination with item novelty and item diversity to obtain the context-aware recommendation.

[0101] Furthermore, the platform is also used to implement the following functions:

[0102] The user interest vector and the context vector are concatenated to form a user feature vector; a context importance score is calculated for the user feature vector as a context weight; and the user interest vector is scaled according to the context weight to obtain the user dynamic interest vector.

[0103] Furthermore, the platform is also used to implement the following functions:

[0104] The real-time interest vector of the target user is read from the cache; the real-time interest vector is context-adapted, and a candidate set is recalled for the context-aware recommendation to obtain a recalled recommendation; recalled item features are extracted for the recalled recommendation, and the recalled item features are matched with the real-time interest vector to obtain a user-recalled item matching degree; based on the user-recalled item matching degree, a multi-objective optimization sorting is performed in combination with the novelty and diversity of the recalled items to obtain the context-recalled-aware recommendation.

[0105] Furthermore, the platform is also used to implement the following functions:

[0106] The short-term interest vector of the preset sliding window is updated according to the real-time interactive event, and the long-term interest vector is incrementally learned; the short-term interest vector and the long-term interest vector are updated by time-attenuated weighted updating to obtain the real-time interest vector.

[0107] Furthermore, the platform is also used to implement the following functions:

[0108] If the target user is a cold start user, a preset default interest vector is used, wherein the preset default interest vector generates a group popular interest vector based on the cold start information of the cold start user; context adaptation and item feature matching are performed based on the preset default interest vector, cold start recommendations are output, cold start exploration is performed, and the recommendations are added to the context-aware recommendations.

[0109] Furthermore, the platform is also used to implement the following functions:

[0110] Cache context recall-aware recommendations and real-time monitoring of service availability; when monitoring indicators are triggered, continuous attempts are made to restore the primary service through graded degradation and user-unaware switching, gradually switching back to normal mode.

[0111] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0112] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

[0113] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A personalized content recommendation method based on user behavior trajectory, characterized in that: include: Construct user behavior sequences through multi-source data; Extracting a short-term interest vector and a long-term interest vector from the user behavior sequence, and weighting them to obtain a user interest vector; Dynamically integrating the context data and the user interest vector and performing recommendation calculations to obtain context-aware recommendations; Recalling the context-aware recommendation according to the real-time interest vector, and obtaining the context-recall-aware recommendation through online reasoning; Personalized content recommendation is performed through the context recall-aware recommendation.

2. The personalized content recommendation method based on user behavior trajectory according to claim 1, characterized in that: Construct user behavior sequences through multi-source data, including: By collecting explicit feedback data and implicit feedback data from target users, multi-source feedback data is obtained; Acquiring context data of the multi-source feedback data; Performing anomaly detection and filtering on the multi-source feedback data and the context data, and performing sequence normalization to construct a triplet dataset of user-behavior sequence-context; Context encoding is performed according to the triplet data set, and the user behavior sequence is output.

3. The personalized content recommendation method based on user behavior trajectory according to claim 1, characterized in that: Extracting a short-term interest vector and a long-term interest vector from the user behavior sequence and weighting them to obtain a user interest vector includes: Extracting multiple interaction events of a preset sliding window according to the user behavior sequence; Capturing short-term dependencies of the multiple interaction events based on a gated recurrent unit to complete short-term interest vector modeling; Extract all interaction events according to the user behavior sequence; The long-term dependencies of all interaction events are captured through the self-attention mechanism to complete the long-term interest vector modeling; The short-term interest vector and the long-term interest vector are weighted by time decay to obtain the user interest vector.

4. The personalized content recommendation method based on user behavior trajectory according to claim 1, characterized in that: Dynamically integrating the context data and the user interest vector and performing recommendation calculations to obtain context-aware recommendations includes: Performing context encoding on the context data to obtain a context vector; Using a gating mechanism to calculate the degree to which the context vector adjusts the user interest vector, and scaling the result to obtain a user dynamic interest vector; Extracting item features, matching the item features with the user's dynamic interest vector to obtain a user-item matching degree; According to the user-item matching degree, a multi-objective optimization sorting is performed in combination with the item novelty and the item diversity to obtain the context-aware recommendation.

5. The personalized content recommendation method based on user behavior trajectory according to claim 4, characterized in that: The gating mechanism is used to calculate the degree of adjustment of the context vector to the user interest vector, and scaling is performed to obtain the user dynamic interest vector, including: splicing the user interest vector and the context vector to form a user feature vector; Calculating a context importance score for the user feature vector as a context weight; The user interest vector is scaled according to the context weight to obtain the user dynamic interest vector.

6. The personalized content recommendation method based on user behavior trajectory according to claim 1, characterized in that: Recalling the context-aware recommendation based on the real-time interest vector and obtaining the context-recall-aware recommendation through online reasoning includes: Read the target user's real-time interest vector from the cache; Performing context adaptation on the real-time interest vector and recalling a candidate set for the context-aware recommendation to obtain a recalled recommendation; Extracting recall item features from the recall recommendation, matching the recall item features with the real-time interest vector, and obtaining a user-recall item matching degree; According to the user-recalled item matching degree, combined with the recalled item novelty and recalled item diversity, a multi-objective optimization sorting is performed to obtain the context recall-aware recommendation.

7. The personalized content recommendation method based on user behavior trajectory according to claim 6, characterized in that: Read the target user's real-time interest vector from the cache, including: Update the short-term interest vector of the preset sliding window based on real-time interaction events, and perform incremental learning of the long-term interest vector; The real-time interest vector is obtained by updating the short-term interest vector and the long-term interest vector through time-attenuated weighted updating.

8. The personalized content recommendation method based on user behavior trajectory according to claim 6, characterized in that: If the target user is a cold start user, a preset default interest vector is used, wherein the preset default interest vector generates a group hot interest vector based on the cold start information of the cold start user; Context adaptation and item feature matching are performed based on the preset default interest vector, cold start recommendation is output and cold start exploration is performed, and the recommendation is added to the context-aware recommendation.

9. The personalized content recommendation method based on user behavior trajectory according to claim 1, characterized in that: After personalized content recommendation is performed through the context recall-aware recommendation, it includes: Cache context-recall-aware recommendations and monitor service availability in real time; When monitoring indicators are triggered, we continuously try to restore the primary service through graded degradation and user-unaware switching, and gradually switch back to normal mode.

10. A personalized content recommendation platform based on user behavior trajectory, characterized by: The platform is used to implement the personalized content recommendation method based on user behavior trajectory according to any one of claims 1 to 9, and the platform includes: User behavior sequence construction module, used to construct user behavior sequences through multi-source data; A user interest vector acquisition module, configured to extract a short-term interest vector and a long-term interest vector from the user behavior sequence, and weight them to obtain a user interest vector; A recommendation calculation module, configured to dynamically integrate the context data and the user interest vector and perform recommendation calculations to obtain context-aware recommendations; A context-recall-aware recommendation module, configured to recall the context-aware recommendation according to the real-time interest vector and obtain the context-recall-aware recommendation through online reasoning; The personalized content recommendation module is used to perform personalized content recommendation through the context recall-aware recommendation.

Images