Information pushing method and device based on user behavior analysis, equipment and medium

Abstract

The application discloses an information pushing method and device based on user behavior analysis, equipment and medium. Wherein, the method comprises: acquiring a multi-modal data source for user behavior analysis, inputting the multi-modal data source into a target hierarchical space-time model, and outputting a user state vector; fusing user behavior instant features and user life cycle value features to construct a target DQN model, and inputting the user state vector into the target DQN model; based on user activity and market intensity, the dynamic weights of the user behavior instant features and the user life cycle value features are adjusted in real time, wherein the dynamic weights of the user instant features and the user life cycle value features are used to determine the output of the target DQN model; acquiring the output target function of the target DQN model, determining at least one target information pushing action from the candidate information pushing actions based on the target function, and the application solves the technical problems of low information pushing efficiency and low flexibility in the prior art.

Description

Technical Field

[0001] This application relates to the field of computers, and more specifically, to an information push method, apparatus, device, and medium based on user behavior analysis. Background Technology

[0002] In existing apps, users receive push notifications when interacting with the application. Most current technologies build user profiles based on historical data such as purchase history and browsing time, and use collaborative filtering algorithms to generate recommendation lists. For example, they might push similar coupons to users who frequently purchase baby products. Alternatively, they use a pre-defined user group (e.g., "users who haven't made a purchase in the last 30 days") to periodically push information. However, this approach ignores the impact of device performance and environmental characteristics on push efficiency, and the long model update cycle makes it unable to respond in real-time to market changes or fluctuations in user behavior, resulting in a poor user experience.

[0003] In summary, existing technologies suffer from low information delivery efficiency and limited flexibility. Currently, no effective solution has been proposed to improve information delivery efficiency amidst complex and ever-changing information flows. Summary of the Invention

[0004] This application provides an information push method, apparatus, device, and medium based on user behavior analysis, to at least solve the technical problem of improving information push efficiency during user interaction with applications.

[0005] According to one aspect of the embodiments of this application, an information push method based on user behavior analysis is provided, including: Responding to user behavior commands, the system acquires multimodal data sources for user behavior analysis, inputs these multimodal data sources into the target hierarchical spatiotemporal model, and outputs user state vectors. The target hierarchical spatiotemporal model is divided into multiple layers by different time scales and modalities to fuse multimodal data sources and process various types of data. A target DQN model is constructed by integrating real-time user behavior features and user lifetime value features. The aforementioned user state vectors are input into the target DQN model, where the aforementioned user feature vectors serve as the state space of the target DQN model, and the information push parameters serve as the action space of the target DQN model. The dynamic weights of immediate user behavior features and user lifetime value features are dynamically adjusted in real time based on user activity and market strength. The dynamic weights of the aforementioned immediate user features and user lifetime value features are used to determine the output of the target DQN model. Obtain the output objective function of the aforementioned target DQN model, and determine at least one target information push action from the candidate information push actions based on the aforementioned objective function.

[0006] According to another aspect of the embodiments of this application, an information push device based on user behavior analysis is also provided, comprising: The acquisition unit is used to respond to user behavior commands, acquire multimodal data sources for user behavior analysis, input the multimodal data sources into the target hierarchical spatiotemporal model, and output user state vectors. The target hierarchical spatiotemporal model is divided into multiple layers by different time scales and modalities to fuse multimodal data sources and process various types of data. The fusion unit is used to fuse real-time user behavior features and user lifetime value features to construct a target DQN model. The user state vector is input into the target DQN model, where the user feature vector serves as the state space of the target DQN model and the information push parameters serve as the action space of the target DQN model. The adjustment unit is used to adjust the dynamic weights of the immediate user behavior features and the user lifetime value features in real time based on user activity and market strength. The dynamic weights of the aforementioned immediate user features and user lifetime value features are used to determine the output of the target DQN model. The determining unit is used to obtain the output objective function of the aforementioned target DQN model, and determine at least one target information push action from the candidate information push actions based on the aforementioned objective function.

[0007] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer-readable storage medium, and the computer program is configured to execute the above-described information push method based on user behavior analysis when running.

[0008] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described information push method based on user behavior analysis through the computer program.

[0009] In this embodiment, in response to user behavior commands, a multimodal data source for user behavior analysis is obtained. This multimodal data source is then input into a target hierarchical spatiotemporal model, outputting a user state vector. The target hierarchical spatiotemporal model, through a multi-layered architecture with different time scales and modalities, is used to fuse multimodal data sources and process various data types. A target DQN model is constructed by fusing real-time user behavior features and user lifetime value features. The user state vector is then input into the target DQN model, where the user feature vector serves as the state space of the target DQN model, and the information push parameters serve as the action space of the target DQN model. Based on user activity... The dynamic weights of real-time user behavior features and customer lifetime value features are dynamically adjusted in real time based on user activity and market intensity. These dynamic weights are used to determine the output of the target DQN model. The target objective function of the target DQN model is obtained, and at least one target information push action is determined from the candidate information push actions based on the target objective function. This achieves the purpose of multi-source data fusion, improves the diversity of information push, and dynamically adjusts the weight allocation based on user activity and market activity in real time, achieving a balance between CTR (click-through rate) and CLV (customer lifetime value), thus solving the technical problem of low information push efficiency in existing technologies. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of a flowchart of an optional information push method based on user behavior analysis according to an embodiment of this application; Figure 2 This is a schematic diagram of an optional information push device based on user behavior analysis according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application. Detailed Implementation

[0011] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0012] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] Most applications on the market have a large user base. During user interaction, such as viewing, clicking, and purchasing, data is generated, and users also receive push notifications. Current technologies either build user profiles based on historical data like purchase records and browsing time, then use collaborative filtering algorithms to generate recommendation lists (e.g., pushing similar coupons to users who frequently buy baby products), or trigger coupon pushes based on immediate actions, such as page dwell time exceeding 30 seconds, but these rely on only a single behavioral dimension. Furthermore, current push mechanisms typically either pre-select user groups and periodically send fixed messages, or selectively push information to high-gain users, leading to reliance on historical data and low update frequency. In data transmission, current technologies generally use a single data modality, only integrating user behavior logs such as clicks and add-to-cart actions, ignoring device performance factors like page loading speed and environmental characteristics like network latency affecting decision-making; or a centralized computing architecture, requiring user behavior data to be transmitted back to a central server for processing, which poses risks of privacy leaks and high response latency. Furthermore, traditional deep learning models cannot explain the reasons for push notifications, resulting in low user trust and a poor user experience.

[0014] This embodiment achieves multi-source data fusion through a hierarchical spatiotemporal attention network and designs a dual-objective deep learning approach. By constructing an objective DQN model and dynamically adjusting the weights of immediate conversion rate (CTR) and long-term user value (LTV) in real time based on user activity and market strength, it can respond to market changes or fluctuations in user behavior in real time, thereby improving information push efficiency and user experience.

[0015] Alternatively, as an optional implementation, such as Figure 1 As shown, the information push methods include: S102, responding to user behavior commands, acquires multimodal data sources for user behavior analysis, inputs the multimodal data sources into the target hierarchical spatiotemporal model, and outputs user state vectors. The target hierarchical spatiotemporal model is divided into multiple layers by different time scales and modalities to fuse multimodal data sources and process various types of data. S104, integrate real-time user behavior features and user lifetime value features to construct a target DQN model, and input the user state vector into the target DQN model, where the user feature vector serves as the state space of the target DQN model and the information push parameters serve as the action space of the target DQN model; S106, dynamically adjust the weights of immediate user behavior features and user lifetime value features in real time based on user activity and market strength, wherein the dynamic weights of immediate user features and user lifetime value features are used to determine the output of the target DQN model. S108, obtain the output objective function Q of the target DQN model, and determine the target information push action from the candidate information push actions based on the objective function Q, wherein the target information push action is at least one of the candidate information push actions.

[0016] Optionally, in this embodiment, the user behavior instruction in S102 can be understood as detecting the real-time click stream of the user terminal, the duration of the user's page stay, etc.; the acquisition of multimodal data sources for user behavior analysis can be, but is not limited to, real-time behavior collected by the user terminal through the SDK. The collected real-time behavior can include, but is not limited to, page scrolling speed, click hot zone coordinates, GPU utilization, GPS positioning, etc. Based on the above user behavior and environmental information, the data can be classified into multimodal data sources such as user behavior data, device and environment data, and historical preference data. Among them, user behavior data can include, but is not limited to, real-time click stream, page stay duration, purchase path, etc., device and environment data can include, but is not limited to, GPU utilization, page loading speed, geographical location, etc., and historical preference data can include, but is not limited to, data such as purchase cycle and category preference aggregated across clients through a federated learning framework.

[0017] Optionally, in this embodiment, after determining the multimodal data source, considering that any modal data source may have modal gaps due to factors such as device acquisition, leading to incomplete data, such as some clients lacking GPS data resulting in missing device and environmental data, thus lowering the accuracy of the final result, this embodiment supplements or prioritizes the acquisition of missing modal data through Shapley value quantization and time feature fusion methods, thereby achieving the technical effect of improving data comprehensiveness.

[0018] Optionally, in this embodiment, a multimodal data source is input into the target hierarchical spatiotemporal model, which outputs a user state vector. The target hierarchical spatiotemporal model is structured into multiple layers based on different time scales and modalities to fuse multimodal data sources and process various data types. The target hierarchical spatiotemporal model can be, but is not limited to, a local end-user model based on the MHSTAN neural network. Specifically, it includes a short-term behavior layer for processing second-level operation sequences, a long-term periodic layer for analyzing historical periodic features, and cross-modal attention for fusing device performance. The outputs of each layer interact through cross-layer attention or feature concatenation. For example, long-term periodic features serve as contextual input to the short-term behavior layer, and all layers are trained jointly rather than independently and then concatenated. The target hierarchical spatiotemporal model outputs a user state vector. ,in In mathematics, this represents an n-dimensional real vector space, where each component of each vector is a real number. Let represent the user's state vector at time t. It should be noted that the choice of 512 dimensions is an empirical one, achieving a balance between expressive power and computational costs, such as memory and communication overhead. In addition, higher-dimensional spaces can encode complex features more precisely, such as user behavior and image semantics.

[0019] Optionally, in this embodiment, in S104, a target DQN model is constructed based on a dual-objective machine learning strategy of user behavior instantaneous features (CTR) and user lifetime value features (LTV), and the user feature vector is... As the state space of the model, information is pushed to parameters. This serves as the action space for the model, and the reward function is dynamically adjusted accordingly. The formula for dynamically adjusting the reward function is as follows: ; in, This represents the user activity weight, used to control the impact of user activity on the overall weight. The intensity of the impact, for example, the business focus on long-term retention. The value will be increased, causing highly active users to trigger the CTR (Click-Through Rate) feature more quickly, while the weight of the CTR feature will decrease. The weight representing market strength is used to reflect the degree of intervention by external marketing activities in strategy adjustments, such as during the Double 11 shopping festival. The value is temporarily increased, and even highly active users need to maintain a high weighting of the real-time characteristics of user behavior (CTR) in order to boost short-term gross merchandise volume (GMV).

[0020] Optionally, in this embodiment, in S106, the dynamic weights of the immediate user behavior features and the user lifetime value features are adjusted in real time based on user activity and market strength. The dynamic weights of the immediate user behavior features and the user lifetime value features are used to determine the output of the target DQN model. This may include, but is not limited to, using an MLP neural network to analyze the immediate user behavior features and an LSTM neural network to analyze the user lifetime value features. By adjusting the weights, the output weights of the MLP neural network and the LSTM neural network are adjusted, ultimately affecting the calculated value of the objective function, i.e., the output of the target DQN model.

[0021] Optionally, in this embodiment, in S108, the target information push action is determined from the candidate information push actions based on the objective function. For example, if the candidate actions are pushing clothing coupons and pushing daily necessities discount coupons, and the objective function for pushing clothing coupons is greater than that for pushing daily necessities discount coupons, then the clothing coupon push action is selected as the target information push action, and the information of clothing coupons is pushed to the user with a high probability, while the information of daily necessities discount coupons is attempted with a low probability.

[0022] In this embodiment, responding to user behavior commands, a multimodal data source for user behavior analysis is acquired. This multimodal data source is then input into a target hierarchical spatiotemporal model, outputting a user state vector. The target hierarchical spatiotemporal model, with its multi-layered architecture based on different time scales and modalities, is used to fuse multimodal data sources and process various data types. A target DQN model is constructed by fusing real-time user behavior features and user lifetime value features. The user state vector is then input into the target DQN model, where the user feature vector serves as the state space of the target DQN model, and the information push parameters serve as the action space. Based on user activity and... The market strength dynamically adjusts the weights of real-time user behavior features and customer lifetime value features. The dynamic weights of real-time user features and customer lifetime value features are used to determine the output of the target DQN model. The target objective function of the target DQN model is obtained, and at least one target information push action is determined from the candidate information push actions based on the objective function. This achieves the purpose of multi-source data fusion and improves the diversity of information push. Furthermore, the weight allocation is dynamically adjusted in real time based on user activity and market activity, which achieves the goal of balancing CTR (click-through rate) and CLV (customer lifetime value), thus improving the technical effect of information push efficiency.

[0023] Optionally, as an alternative implementation, the dynamic weights of real-time user behavior characteristics and user lifetime value characteristics are adjusted based on user activity and market strength, including: S1. Determine the user's activity level based on user behavior information, which includes user behavior frequency, number of deep user interactions, and number of core user actions. S2, to obtain user activity levels and market strength; S3, when the user activity level is greater than the target high activity threshold, triggers the weight formula to adjust the dynamic weights of the user behavior instant features and the user lifetime value features. S4, when the market activity level is greater than the target high activity level, triggers the weight formula to adjust the dynamic weights of the user behavior instant features and user lifetime value features.

[0024] Optionally, in this embodiment, the user's activity level can be determined based on user behavior information such as user behavior frequency, number of deep user interactions, and number of core user actions. For example, users who meet at least two of the following criteria are identified as highly active users: (1) Behavior frequency threshold: Daily active users: ≥5 days of login in the past 7 days, or ≥10 interactions per day (such as clicks, favorites); Weekly active users: ≥4 weeks of the past 30 days have reached the above daily active user standard; (2) Deep participation indicators: Content production: ≥3 UGC content posts per week (such as comments, notes); Social interaction: ≥5 private messages / shares per day; (3) Core business actions: ≥8 times per week, or ≥40% monthly repurchase rate.

[0025] Optionally, in this embodiment, market activity can also be determined by the frequency of market activities and the level of incentives. If the market activity or user activity exceeds the target high activity threshold, a weighting formula is triggered to adjust the dynamic weights of the user's immediate behavior characteristics and user lifetime value characteristics. The weighting formula is the reward function:

[0026] As an optional implementation, for example, each user device's local model receives global parameters and calculates the local user behavior's immediate characteristics CTR and user lifetime value (LTV) prediction based on user behaviors such as clicks and adding items to cart, and long-term values ​​such as historical ARPU and repurchase cycle. It acquires user activity levels and market strength, collecting data that a user has added items to their cart 10 times in the past week, indicating a user activity level greater than the target high activity threshold, triggering the weighting formula: ; Among them, parameter settings (High activity level) (rapid descent) (Weak impact of market activities); The final result is (t) = 0.3 (originally 0.7), CTR weight decreased from 70% to 30%, while LTV weight increased to 70%. This achieves the goal of dynamically adjusting the weights of immediate user behavior characteristics and user lifetime value characteristics in real time based on user activity and market strength. The final effect is that the server aggregates the low CTR weight tendency of highly active users, generates a global model, and pushes "member-exclusive benefits" to these users instead of "limited-time discounts," focusing on long-term retention. For users, the short-term click-through rate decreased by 20%, but the 90-day repurchase rate increased by 35%. For the market's cumulative LTV value increased by 18%, and the customer lifetime value was extended. In other words, a balance was achieved between immediate user behavior characteristics and user lifetime value characteristics, improving the efficiency of information push.

[0027] Optionally, as an optional implementation, obtaining the output objective function Q of the target DON model includes: Calculate the objective function Q: ; in, It is the action value function of choosing action a from the current state s. It is a dynamic weight of the user behavior instant features and the user lifetime value features. MLP is a neural network used to analyze the user behavior instant features. MLP(S) is the output parameter of the MLP neural network when processing the current state S. LSTM is a neural network used to analyze the user lifetime value features. LSTM(CLV) is the output parameter of the LSTM neural network when processing the current state S.

[0028] Optionally, in this embodiment, the local DQN is based on Generate candidate action set Calculate the Q-value and dynamic weights. The action selection is adaptively adjusted based on user activity, and can be, but is not limited to, using the ε-greedy strategy: High-activity users: ε=0.1 (focus on utilization); Low-activity users: ε=0.5 (focus on exploring new strategies).

[0029] As a specific implementation, the Q-value is calculated for highly active users with ε=0.1 as the primary focus: First, a user profile is obtained, specifically user ID A. The user's activity level is assessed, indicating daily login, average monthly spending of 5000 yuan, and a historical preference for "maternal and infant products." Based on the current user's action commands, a candidate action set is generated: candidate coupon types include {maternal and infant product discount coupons, daily necessities discount coupons, clothing coupons, and electronics coupons}. The Q-score is calculated to determine if the user is highly active, and then the dynamic weights of the user's immediate behavioral characteristics and lifetime value characteristics are determined. The Q-value is 0.9, at which point the output of the neural network MLP used to analyze real-time characteristics of user behavior dominates the weights. MLP input: The user is currently browsing a baby stroller page → the predicted "baby and maternity discount coupon" has the highest Q-value; LSTM input shows that historical CLV data indicates a high repurchase rate for baby and maternity products → reinforcing the long-term value of the "baby and maternity discount coupon".

[0030] Final Q-value ranking: Mother and baby product discount coupons > Daily necessities coupons > Apparel coupons > Electronics product coupons.

[0031] Targeted information push action selection: 90% probability of selecting the action with the highest Q value, namely the mother and baby product discount coupon; 10% random attempt at other coupons such as electronics coupons. The final result accurately matches known user preferences, improving short-term conversion rates; limited exploration avoids strategy rigidity. This improves information push efficiency and user experience.

[0032] As a specific example, Q-score calculation is performed for highly active users with ε=0.5 who focus on exploring new strategies: First, we obtain the user profile, i.e., user ID B, and the user's activity level, i.e., a new user who has only purchased low-priced daily necessities once, resulting in a low CLV prediction. Based on the action instructions in the current state, we generate a candidate action set: candidate coupon types: {daily necessities coupon, beauty coupon, sports equipment coupon, book coupon}. We calculate the Q-score, determining the user to be a low-activity user. In this case, the dynamic weights of the user's immediate behavior features and user lifetime value features are set to 0.2, and the LSTM neural network used to analyze user lifetime value features dominates the weights. MLP input: The user recently searched for "yoga mat" → predicting a higher Q-score for "sports equipment coupon"; LSTM input: The CLV prediction model finds that "beauty users" have higher long-term value → increasing the weight of "beauty coupon".

[0033] Final Q-value ranking: Beauty coupons > Sports equipment coupons > Daily necessities coupons > Book coupons.

[0034] Targeted information push action selection: 50% probability of selecting the action with the highest Q value to push beauty coupons, 50% random attempt to push other coupons, such as sports equipment coupons, book coupons, etc. By proactively exploring users' potential interests such as beauty and sports through the final results, behavioral data is quickly accumulated, optimizing long-term CLV (Customer Lifetime Value). This improves information push efficiency and user experience.

[0035] Optionally, as an alternative implementation, before inputting the multimodal data source into the target hierarchical spatiotemporal model, a target hierarchical spatiotemporal model is constructed, including: S1 constructs a short-term behavior layer by processing second-level operation sequences through a bidirectional gated recurrent neural network; S2, a long-term periodic layer is constructed by analyzing historical purchase intervals using fast Fourier transform and neural networks; S3 dynamically allocates weights by fusing device and environmental features through cross-modal attention fusion.

[0036] Optionally, in this embodiment, a lightweight hierarchical spatiotemporal model is constructed and executed on the device: specifically, it includes a short-term behavior layer, a long-term cycle layer, and cross-modal attention. The short-term behavior layer processes second-level operation sequences through a bidirectional gated recurrent neural network (BiGRU), such as processing the feature of continuous clicks on the same category of goods; the long-term cycle layer analyzes historical purchase intervals through fast Fourier transform and neural network (FFT+MLP), such as extracting weekly and monthly purchase frequency cycle features; cross-modal attention (MHA) fuses device and environmental features and dynamically allocates weights, such as reducing the weight of real-time features when memory is low to reduce memory usage.

[0037] Optionally, in this embodiment, the hierarchical spatiotemporal model features hierarchical joint optimization, with the outputs of each layer interacting through cross-layer attention or feature concatenation. For example, long-term periodic features serve as contextual input for the short-term behavior layer. The hierarchical spatiotemporal model also features end-to-end training: all layers are trained jointly, rather than being trained independently and then concatenated. For instance, the output of the short-term behavior layer is input to the attention layer along with Fast Fourier Transform features.

[0038] It should be noted that the hierarchical spatiotemporal model features topology-aware supernetworks and conditional fusion. It can dynamically generate the attention weight matrix based on device performance, such as memory usage, rather than statically adjusting it. For example, when GPU usage exceeds 80%, the update frequency of real-time behavioral features is reduced, while the weight of long-term periodic features is increased. Considering the impact of device performance on information delivery, the weights and update frequency are adjusted in real-time based on GPU usage data.

[0039] Optionally, in this embodiment, the hierarchical spatiotemporal model considers both temporal proximity (such as the timing of operation sequences) and spatial correlation (such as the similar behaviors of users in geographically close locations) in its attention mechanism. The model employs hierarchical spatiotemporal embedding, encoding timestamps and spatial coordinates into a joint vector, which is then input into each layer of the model.

[0040] Optionally, as an alternative implementation, acquiring a multimodal data source for user behavior analysis includes: S1, collect multimodal data sources, including user behavior data, device and environment data, and historical preference data; S2, in the case of missing modal data sources, the impact weight of missing modal data sources is quantified by the Shapley value; S3 prioritizes the collection or compensation of missing modal data sources with higher weights.

[0041] Optionally, in this embodiment, the user terminal collects real-time behavior through an SDK to obtain a multimodal data source. Based on the collected real-time user behavior commands and environmental information, the multimodal data source can be classified into user behavior data, device and environment data, and historical preference data. To improve the comprehensiveness of data collection, in the event that any modal data source is missing, the impact weight of the missing modal data source is quantified using the Shapley value.

[0042] It should be noted that the Shapley value, derived from game theory, is used to fairly allocate cooperative payoffs among participating modalities, contributing to the final model performance. In this embodiment, the Shapley value is used to quantify the contribution of each modality to the model's prediction results; for example, device performance data contributes 20% to the push decision.

[0043] Optionally, in this embodiment, the influence weight of the true mode on the task can be quantified by Shapley value to determine the true mode with high weight for priority data collection, or the missing mode can be imputed by time feature fusion method to ensure the comprehensiveness of the data.

[0044] Optionally, in this embodiment, when the target model data is sent from the local end to the server, the server performs federated model aggregation. The target global model parameters are the model parameters after federated model aggregation, and then the server sends the target global model parameters to the client. The server uses a modal contribution quantization aggregation algorithm to dynamically adjust the client weights. Optionally, in this embodiment, This represents the weight value of each client, calculated using the formula above. The local model parameters on the client have a greater impact on the global model. Therefore, in the model data transmission from the server to the client, a federated resource allocation is designed to prioritize the allocation of computing resources to high-weight clients, giving them more bandwidth or training time, thereby improving the utilization rate of data transmission.

[0045] Optionally, in this embodiment, after each round of training, global knowledge is transferred to the local model through mutual knowledge distillation (MKD) to prevent forgetting, such as high-value user features that need to be retained long-term. Mutual knowledge distillation is used to distill high-value features in the global model, such as long-term user preferences, into the local model, avoiding the loss of important information due to dynamic weight adjustments. Even if the weight of a client decreases, the knowledge it has contributed is still retained in the global knowledge base through MKD, improving data integrity and security.

[0046] Optionally, as an alternative implementation, a target DQN model is constructed by integrating real-time user behavior features and user lifetime value features, followed by: S1, add noise to the gradient of the target DQN model parameter data to ensure that a single parameter data cannot be inferred in reverse; S2, use an encryption algorithm to encrypt the perturbed gradient, determine the target ciphertext, and send the target ciphertext to the server; S3: The server aggregates all target ciphertext and updates the global model parameters after decryption. S4, the server sends the global model parameters to the client.

[0047] Optionally, in this embodiment, it can be understood as periodically uploading encryption gradients. The data is then sent to the parameter server, where the global model is aggregated and distributed. Specifically, each device adds noise, such as Laplacian noise, to the gradient of the target DQN model parameter data. This ensures that individual parameter data cannot be reverse-engineered, controls the privacy budget, and sets privacy parameters. To limit the risk of information leakage in each training round, encryption algorithms such as Paillier additive homomorphic public-key encryption are used to encrypt the gradients instead of the original data, generating the target ciphertext. The server directly sums the encryption gradients to obtain the aggregated result. The server decrypts and aggregates gradients, updates the global model, and distributes it to each device. By employing differential privacy federated learning and localizing user data processing, only gradients are encrypted during transmission, rather than directly uploading the raw data to the server, thus overcoming the risk of data leakage. Furthermore, feature binning reduces edge-side inference latency to 50ms, improving transmission efficiency.

[0048] As an optional implementation, to avoid pushing high-value coupons to users with low customer lifetime value (CLV) (an abnormal action), the steps are as follows: First, user A's device trains a lightweight DQN model using local data (purchase history, click behavior) to generate gradients. Then, differential privacy processing is performed: Laplacian noise is added to the gradient. ,in It is random noise added to meet differential privacy requirements. This indicates that the noise scaling parameter is 0.5, which is related to data sensitivity and privacy budget. Indicate a privacy budget and control the level of privacy protection. The smaller the size, the stronger the privacy protection, but the louder the noise.

[0049] Then use the Paillier public key to encrypt the perturbed gradient E( Uploaded to the server; the server performs aggregation and anomaly detection, aggregating all encryption gradients. After decryption, update the global model.

[0050] Optionally, this embodiment implements asynchronous updates and offline support. Local caching can be understood as, but is not limited to, caching decision logs on the device when the network is down, such as user B's push notification records, which are then encrypted and uploaded after the network is restored. Delayed synchronization can be understood as, for example, the server aggregating delayed updates from offline devices every 6 hours in a time window to ensure eventual consistency. This embodiment achieves the technical effect of protecting user-level privacy and not losing critical behavioral data even in weak network environments.

[0051] Optionally, as an alternative implementation, the server sends global model parameters to the client, followed by: S1, calculate the federal target value and determine the target attribution report, whereby the target attribution report is used to explain the push decision in real time; S2, determine the target interception value based on the global model parameters issued by the server; S3 intercepts the target information push action if the client's federated target value is less than the target interception value.

[0052] Optionally, the federated DeepSHAP target value is calculated, a target attribution report is determined, and a target interception value is determined based on global model parameters. Interception is triggered when the client's federated target value is less than the target interception value; that is, if the client's DeepSHAP target value is detected to be less than the target interception value, the rule engine intercepts the abnormal push, such as pushing high-value coupons to low-value users. The target attribution report is used to explain push decisions in real time. The real-time attribution formula for federated DeepSHAP is as follows: Local computation feature contribution ; in, For feature sensitivity, The activation strength of local features is calculated by multiplying two parts to determine the feature contribution. The sensitivity of the function to features and the activation strength of local features. The global value function in federated learning evaluates the expected benefits of taking action a in state s, such as model prediction accuracy and collaboration efficiency.

[0053] As an optional implementation, a high-risk example: the action is to push a 500 yuan high-priced product coupon to user B whose historical purchasing power is ≤100 yuan; the feature contribution analysis is the user's purchasing power contribution. The abnormal characteristics, where the global mean is 0.6, The blocking rules are: The value is less than 0.6 − 2 × 0.15 = 0.3, meaning the push notification is blocked. This implements server-side aggregation of attribution results, triggering the rule engine to block high-risk actions.

[0054] Optionally, in this embodiment, each device periodically calculates the DeepSHAP target value and generates an attribution report, such as "This push is due to the user frequently browsing this category in the past 3 days," to explain the reason for the push. Abnormal decisions are intercepted based on the global mean and interception rules, such as pushing high-value coupons to low-value users.

[0055] Alternatively, as an alternative implementation, training can be carried out in stages during the model building phase. Simulated behavioral data can be used for pre-training during the cold start phase, and real interaction data can be gradually introduced during the maturity phase to avoid distribution bias.

[0056] Optionally, as an alternative implementation, during the online federated learning process, if user behavior changes, interest drift can be detected in the local model using KL divergence. Specifically, the client monitors the changes in KL divergence of the behavior distribution in real time.

[0057] ; in, The Kullback-Leibler divergence (KL divergence) is defined as a measure of the divergence between two probability distributions. and The degree of difference between them The definition is the distribution of new behaviors currently monitored by the client, such as user clicks, purchases, and other interaction data, used to reflect the latest state of system / user behavior; The definition is the historical baseline distribution, the behavior distribution during the training phase or the last fine-tuning, used as a "normal state" reference for comparison; The definition is a preset threshold, a critical value used to determine whether a change in distribution is significant, and needs to be determined through experiments or business requirements.

[0058] As an optional implementation, when the edge inference latency is, for example, <50ms, and the federated aggregation convergence speed is, for example, <1% error within 5 rounds, performance testing is performed. Optionally, this can be understood as verifying whether the new scheme of the interest drift detection model is superior to the traditional collaborative filtering scheme, reducing the risk of full deployment. For example, if the user feedback rating of the old scheme is greater than that of the new scheme for three consecutive days, the old scheme is automatically rolled back to the old scheme. Or, if the performance monitoring triggers an alarm if the single-point latency is greater than 50ms, and the backup server node is switched after 3 consecutive rounds of communication failure.

[0059] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0060] According to another aspect of the embodiments of this application, an information push device for implementing the above-described information push method is also provided. For example... Figure 2 As shown, the device includes: The fusion unit 202 is used to respond to user behavior commands and fuse target multi-source data to construct a target hierarchical spatiotemporal model. The target hierarchical spatiotemporal model is divided into multiple layers through different time scales and modalities to process various types of data. The adjustment unit 204 is used to determine the values ​​of the first indicator and the second indicator based on the target global model parameters, and to dynamically adjust the weights of the first indicator and the second indicator based on user activity and market activity. The target global model parameters are the model parameters after aggregation of the federated model. The first indicator is used to represent the immediate conversion rate, and the second indicator is used to represent the long-term user value. The determining unit 206 is used to determine the target execution action according to the target dynamic scheduling algorithm, wherein the target execution action is at least one of the candidate execution actions.

[0061] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described information push method is also provided, such as... Figure 3 As shown, the electronic device includes a memory 302 and a processor 304. The memory 302 stores a computer program, and the processor 304 is configured to execute the steps in any of the above method embodiments via the computer program.

[0062] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0063] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program: S1 responds to user behavior commands, acquires multimodal data sources for user behavior analysis, inputs the multimodal data sources into the target hierarchical spatiotemporal model, and outputs user state vectors. The target hierarchical spatiotemporal model is divided into multiple layers by different time scales and modalities to fuse multimodal data sources and process various types of data. S2, integrate real-time user behavior features and user lifetime value features to construct the target DQN model, and input the user state vector into the target DQN model. The user feature vector serves as the state space of the target DQN model, and the information push parameters serve as the action space of the target DQN model. S3 dynamically adjusts the weights of immediate user behavior features and user lifetime value features in real time based on user activity and market strength. The dynamic weights of immediate user features and user lifetime value features are used to determine the output of the target DQN model. S4, obtain the output objective function of the target DQN model, and determine at least one target information push action from the candidate information push actions based on the objective function.

[0064] Alternatively, as those skilled in the art will understand, Figure 3 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 3 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 3 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 3 The different configurations shown.

[0065] The memory 302 can be used to store software programs and modules, such as the program instructions / modules corresponding to the information push method and apparatus in this embodiment. The processor 304 executes various functional applications and data processing by running the software programs and modules stored in the memory 302, thereby realizing the aforementioned information push method based on user behavior analysis. The memory 302 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 302 may further include memory remotely located relative to the processor 304, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 302 may be used, but is not limited to, to store information such as candidate information push actions. As an example, such as... Figure 3 As shown, the memory 302 may include, but is not limited to, the acquisition unit 202, fusion unit 204, adjustment unit 206, and determination unit 208 of the information push device. Furthermore, it may include, but is not limited to, other module units of the information push device based on user behavior analysis, which will not be elaborated upon in this example.

[0066] Optionally, the transmission device 306 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 306 includes a Network Interface Controller (NIC), which can be connected to other network devices and routers via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 306 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0067] In addition, the aforementioned electronic device also includes: a display 308 for displaying information such as target information push actions; and a connection bus 310 for connecting various module components in the aforementioned electronic device.

[0068] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored therein, wherein the computer program is configured to perform the steps in any of the above method embodiments when it is run.

[0069] Optionally, in this embodiment, the computer-readable storage medium described above may be configured to store a computer program for performing the following steps: S1 responds to user behavior commands, acquires multimodal data sources for user behavior analysis, inputs the multimodal data sources into the target hierarchical spatiotemporal model, and outputs user state vectors. The target hierarchical spatiotemporal model is divided into multiple layers by different time scales and modalities to fuse multimodal data sources and process various types of data. S2, integrate real-time user behavior features and user lifetime value features to construct the target DQN model, and input the user state vector into the target DQN model. The user feature vector serves as the state space of the target DQN model, and the information push parameters serve as the action space of the target DQN model. S3 dynamically adjusts the weights of immediate user behavior features and user lifetime value features in real time based on user activity and market strength. The dynamic weights of immediate user features and user lifetime value features are used to determine the output of the target DQN model. S4, obtain the output objective function of the target DQN model, and determine at least one target information push action from the candidate information push actions based on the objective function.

[0070] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0071] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0072] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.

[0073] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0074] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0075] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0077] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An information push method based on user behavior analysis, characterized in that, The method includes: Responding to user behavior commands, the system acquires multimodal data sources for user behavior analysis, inputs these multimodal data sources into a target hierarchical spatiotemporal model, and outputs a user state vector. The target hierarchical spatiotemporal model is a multi-layered architecture divided by different time scales and modalities, used to fuse multimodal data sources and process various types of data. A target DQN model is constructed by integrating real-time user behavior features and user lifetime value features. The user state vector is input into the target DQN model, wherein the user feature vector serves as the state space of the target DQN model, and the information push parameters serve as the action space of the target DQN model. The dynamic weights of immediate user behavior features and user lifetime value features are dynamically adjusted in real time based on user activity and market strength, wherein the dynamic weights of the immediate user behavior features and user lifetime value features are used to determine the output of the target DQN model. Obtain the output objective function of the target DQN model, and determine at least one target information push action from the candidate information push actions based on the objective function; The dynamic weighting of user behavior instant features and user lifetime value features based on user activity and market strength includes: The user's activity level is determined based on the user behavior information, wherein the user behavior information includes user behavior frequency, number of deep user interactions, and number of core user actions; The user activity level and market strength are obtained, wherein market strength is used to reflect the degree of intervention of external marketing activities on strategy adjustments; If the user activity level is greater than the target high activity threshold, a weighting formula is triggered to adjust the dynamic weights of the user behavior instant features and user lifetime value features. If the market intensity exceeds the target high activity threshold, a weighting formula is triggered to adjust the dynamic weights of the user behavior instant features and user lifetime value features. Obtaining the output objective function Q of the target DON model includes: Calculate the objective function Q: ， Among them, the It is the action value function for choosing action a from the current state s. It is a dynamic weight of the user behavior instant features and the user lifetime value features. The MLP is a neural network used to analyze the user behavior instant features. The MLP(S) is the output parameter of the MLP neural network when processing the current state S. The LSTM is a neural network used to analyze the user lifetime value features. The LSTM(CLV) is the output parameter of the LSTM neural network when processing the current state S.

2. The information push method based on user behavior analysis according to claim 1, characterized in that, Before inputting the multimodal data source into the target hierarchical spatiotemporal model, the method is characterized in that constructing the target hierarchical spatiotemporal model includes: A short-term behavior layer is constructed by processing second-level operation sequences through a bidirectional gated recurrent neural network; A long-term periodic layer is constructed by analyzing historical purchase intervals using Fast Fourier Transform and neural networks; Weights are dynamically allocated by fusing device and environmental features across modal attention.

3. The information push method based on user behavior analysis according to claim 1, characterized in that, The acquisition of multimodal data sources for user behavior analysis includes: Collect multimodal data sources, including user behavior data, device and environment data, and historical preference data; In the event that any modal data source is missing, the influence weight of the missing modal data source is quantified by the Shapley value; Prioritize the collection or compensation of data sources with higher weights for the missing modalities.

4. The information push method based on user behavior analysis according to any one of claims 1-3, characterized in that, The target DQN model is constructed by integrating real-time user behavior features and user lifetime value features, followed by: Noise is added to the gradient of the target DQN model parameter data to ensure that a single parameter data cannot be inferred in reverse. The perturbated gradient is encrypted using an encryption algorithm, the target ciphertext is determined, and the target ciphertext is sent to the server. The server aggregates all the target ciphertext and updates the global model parameters after decryption; The server sends the global model parameters to the client.

5. The information push method based on user behavior analysis according to claim 4, characterized in that, The server sends the global model parameters to the client, followed by: Calculate federal target values ​​and determine target attribution reports, wherein the target attribution reports are used to explain and push decisions in real time; The target interception value is determined based on the global model parameters issued by the server; If the client's federal target value is less than the target interception value, the target information push action will be intercepted.

6. An information push device based on user behavior analysis, characterized in that, include: The acquisition unit is used to respond to user behavior commands, acquire multimodal data sources for user behavior analysis, input the multimodal data sources into the target hierarchical spatiotemporal model, and output user state vectors. The target hierarchical spatiotemporal model is divided into multiple layers by different time scales and modalities to fuse multimodal data sources and process various types of data. The fusion unit is used to fuse real-time user behavior features and user lifetime value features to construct a target DQN model. The user state vector is input into the target DQN model, wherein the user feature vector serves as the state space of the target DQN model, and the information push parameters serve as the action space of the target DQN model. An adjustment unit is used to adjust the dynamic weights of immediate user behavior features and user lifetime value features in real time based on user activity and market strength, wherein the dynamic weights of the immediate user behavior features and user lifetime value features are used to determine the output of the target DQN model. The determining unit is used to obtain the output objective function of the target DQN model and determine at least one target information push action from the candidate information push actions based on the objective function. The dynamic weighting of user behavior instant features and user lifetime value features based on user activity and market strength includes: The user's activity level is determined based on the user behavior information, wherein the user behavior information includes user behavior frequency, number of deep user interactions, and number of core user actions; The user activity level and market strength are obtained, wherein market strength is used to reflect the degree of intervention of external marketing activities on strategy adjustments; If the user activity level is greater than the target high activity threshold, a weighting formula is triggered to adjust the dynamic weights of the user behavior instant features and user lifetime value features. If the market intensity exceeds the target high activity threshold, a weighting formula is triggered to adjust the dynamic weights of the user behavior instant features and user lifetime value features. Obtaining the output objective function Q of the target DON model includes: Calculate the objective function Q: ， Among them, the It is the action value function for choosing action a from the current state s. It is a dynamic weight of the user behavior instant features and the user lifetime value features. The MLP is a neural network used to analyze the user behavior instant features. The MLP(S) is the output parameter of the MLP neural network when processing the current state S. The LSTM is a neural network used to analyze the user lifetime value features. The LSTM(CLV) is the output parameter of the LSTM neural network when processing the current state S.

7. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 5 through the computer program.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed, performs the method described in any one of claims 1 to 5.

Images