Information pushing method and device

Abstract

This disclosure provides an information push method and apparatus, relating to the field of computer technology, and particularly to the field of information push technology. One specific implementation of the method includes: pushing information to a user using a target push strategy that matches the user's current environmental state; collecting multi-dimensional target feedback information from the user regarding the pushed information; determining the fitness of the target push strategy based on the multi-dimensional target feedback information; iteratively optimizing the target push strategy using a covariance matrix adaptive evolution strategy (CMA-ES) based on the fitness of the target push strategy to generate an optimal push strategy; replacing at least a portion of the target push strategy with the optimal push strategy, and continuing to push information to the user.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the field of information push technology. Background Technology

[0002] With the continuous development of mobile internet and digital businesses, push notifications have become an important technical means for platforms to reach users, improve activity, and convert business. As business scenarios expand, user scale grows, and interests and preferences dynamically change, push notification systems need to continuously adapt to diverse scenarios such as new regions, new user groups, and new functions, while responding to market trends and changes in the competitive environment to ensure the accuracy of push notifications and the effectiveness of user engagement.

[0003] Existing information push strategy systems mostly employ machine learning models or predefined rule sets based on fixed feature engineering to implement push decisions and optimizations. These systems rely on user interaction data accumulated in business scenarios to learn user feedback behaviors such as clicks and conversions to adjust strategies. Strategy generation and updates are typically executed through static configuration or periodic triggering. Summary of the Invention

[0004] This disclosure provides an information push method, apparatus, device, storage medium, and program product.

[0005] In a first aspect, embodiments of this disclosure propose an information push method, comprising: pushing information to a user using a target push strategy that matches the user's current environmental state; collecting multi-dimensional target feedback information from the user regarding the pushed information; determining the fitness of the target push strategy based on the multi-dimensional target feedback information; iteratively optimizing the target push strategy using a covariance matrix adaptive evolution strategy (CMA-ES) based on the fitness of the target push strategy to generate an optimal push strategy; replacing at least a portion of the target push strategy with the optimal push strategy, and continuing to push information to the user.

[0006] Secondly, embodiments of this disclosure propose an information push device, comprising: a push module configured to push information to a user using a target push strategy that matches the user's current environmental state; a collection module configured to collect multi-dimensional target feedback information from the user regarding the pushed information; a determination module configured to determine the fitness of the target push strategy based on the multi-dimensional target feedback information; an iteration module configured to perform iterative optimization using a covariance matrix adaptive evolution strategy (CMA-ES) based on the fitness of the target push strategy to generate an optimal push strategy; and a replacement module configured to replace at least a portion of the target push strategy with the optimal push strategy and continue pushing information to the user.

[0007] Thirdly, embodiments of this disclosure provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0008] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the methods described in the first aspect.

[0009] Fifthly, embodiments of this disclosure provide a computer program product including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0010] The key or essential features of the embodiments disclosed herein are not intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. Wherein: Figure 1 This is a flowchart of an embodiment of the information push method according to this disclosure; Figure 2 This is a flowchart of yet another embodiment of the information push method according to this disclosure; Figure 3 This is a diagram illustrating the overall architecture of the push-and-evolution system based on the AlphaEvolve mechanism. Figure 4 This is a schematic diagram of the policy evolution cycle based on CMA-ES; Figure 5 This is a schematic diagram of the structure of an embodiment of the information push device according to the present disclosure; Figure 6 This is a block diagram of an electronic device used to implement the information push method or video retrieval method of the embodiments of this disclosure. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0014] Figure 1 A flowchart 100 is shown as an embodiment of an information push method according to the present disclosure. The information push method includes the following steps: Step 101: Push information to the user using a target push strategy that matches the user's current environmental state.

[0015] In this embodiment, the entity executing the information push method can use a target push strategy that matches the user's current environmental state to push information to the user.

[0016] A heterogeneous initial strategy library is pre-built and initialized. This library contains numerous initial push strategies covering push timing, channel selection, content matching, and frequency control rules. Each push strategy is encoded and represented using a strategy descriptor vector. The strategy descriptor is a combination of parameters obtained by vectorizing the push strategy, encompassing multiple dimensions such as push timing, channel selection, content matching, and frequency control rules; it is the basic unit of CMA-ES evolutionary operations. For example, the strategy descriptor vector has 64 dimensions, with 16 dimensions for push timing, 8 for channel selection, 32 for content matching, and 8 for frequency control rules. Vectorization is achieved using a combination of one-hot encoding and normalization.

[0017] The system integrates three types of information: real-time context signals from the user's client, business status signals from the cloud, and historical interaction feedback, to generate a unified current environment state vector. The real-time context signals from the user's client include real-time behaviors such as browsing, clicking, dwelling, page switching, and session status within the application. The business status signals from the cloud include user tags, user types, and business status information stored and maintained in the cloud. Historical interaction feedback includes historical behavioral data such as clicks, conversions, negative feedback, and usage duration generated by the user in response to historical push notifications. The fusion process of the environment state vector employs an attention mechanism model, assigning higher attention weights to the user's context signals, which have high real-time requirements. The fused vector has, for example, 128 dimensions, which maps to the vector dimensions of the policy descriptor.

[0018] Based on the current environment state vector, the push strategy with the highest matching degree is selected from the dynamically maintained strategy library as the target push strategy, and the information is pushed according to the timing, channel, content, and frequency control rules specified by the target push strategy. The matching degree is calculated using the cosine similarity algorithm, and the similarity threshold is set, for example, to 0.8. When no strategy meets the similarity requirement, the default baseline strategy in the strategy library will be called.

[0019] In some embodiments, the information push process may be as follows: First, the current environment state is generated based on the user's behavior information.

[0020] User behavior information includes, but is not limited to: real-time session behavior, device status information, and historical interaction data. Device status information includes, but is not limited to, hardware and network characteristics such as device model, operating system version, network type, battery level, and screen on / off status. The above information is then fused and vector-encoded to generate the current environment state that can be used for policy matching.

[0021] Then, a push strategy that matches the current environment state is selected from the strategy library as the target push strategy.

[0022] The strategy library is a dynamically maintained and evolving strategy library. By calculating the similarity between the environment state and the strategy descriptor, the push strategy that best suits the current scenario is selected as the target push strategy.

[0023] Finally, a targeted push strategy is used to push information to users.

[0024] According to the push timing, push channel, content matching rules, and frequency control rules corresponding to the target push strategy, information push operations are performed to the user's device. Among them, push channels include, but are not limited to, in-app push, push notification bar, SMS, and email. The priority and activation conditions of each type of channel will be clearly specified in the strategy descriptor. For example, when the user opens an in-app conversation, the in-app push channel will be used first.

[0025] Step 102: Collect multi-dimensional target feedback information from users regarding the pushed information.

[0026] In this embodiment, the aforementioned execution entity can collect multi-dimensional target feedback information from users regarding the pushed information.

[0027] After the push notification is executed, comprehensive multi-dimensional feedback information from users regarding the push notification is collected. These feedback dimensions include, but are not limited to: short-term CTR (Click-Through Rate), CVR (Conversion Rate), user engagement time, DAU (Daily Active Users), mid-term user retention rate, long-term LTV (Lifetime Value), user negative feedback rate, and impression-click-conversion rate, etc., to comprehensively and objectively quantify the effectiveness of the push notification strategy.

[0028] Step 103: Determine the suitability of the target push strategy based on multi-dimensional target feedback information.

[0029] In this embodiment, the aforementioned execution entity can determine the suitability of the target push strategy based on multi-dimensional target feedback information.

[0030] Based on the collected multi-dimensional target feedback information, the fitness of the target push strategy is calculated. Fitness characterizes the degree to which the push strategy adapts to the current environment, user needs, and business goals, and is the core basis for CMA-ES iterative optimization. Fitness calculation adopts an incremental calculation method. When new feedback data is generated, only the fitness value of the corresponding dimension is updated, without recalculating all dimensions, thereby improving computational efficiency.

[0031] In some embodiments, the fitness of the target push strategy is calculated using a weighted summation or Pareto front method based on multi-dimensional target feedback information.

[0032] The weighted summation method assigns corresponding weights to different feedback metrics based on business priorities, and then sums the weighted feedback metrics across all dimensions to obtain the fitness value. The weights are dynamically adjusted according to the business phase; for example, during the launch of a new business, the weight of retention rate increases while the weight of CTR decreases.

[0033] The Pareto front method, in multi-objective optimization scenarios, selects the optimal strategy that cannot improve another metric without sacrificing any other metric, eliminating inefficient strategies that perform worse across all dimensions, thus ensuring the overall optimality of the strategy. The Pareto front method also incorporates crowding distance calculations to select the optimal strategy with a uniform distribution, preventing the strategy population from getting trapped in local optima.

[0034] Step 104: Based on the adaptability of the target push strategy, use CMA-ES for iterative optimization to generate the optimal push strategy.

[0035] In this embodiment, the aforementioned execution entity can use CMA-ES to iteratively optimize based on the fitness of the target push strategy to generate the optimal push strategy.

[0036] Using strategy descriptors as evolutionary individuals and the fitness of the target push strategy as the optimization objective, an iterative optimization strategy based on covariance matrix adaptive evolution (CMA-ES) is employed. Through a cyclical process of sampling to generate a new strategy population, evaluating population fitness, selecting the optimal individual, updating search distribution parameters, and adjusting the search step size, the system continuously converges towards a high-fitness region, ultimately generating the optimal push strategy that best suits the current environment, user needs, and business objectives.

[0037] The CMA-ES algorithm can be, for example, an improved version of IPOP-CMA-ES (Increasing Population Covariance Matrix Adaptation Evolution Strategy). When the algorithm gets stuck in a local optimum, it will automatically restart and reinitialize the search distribution parameters, with a maximum number of restarts, for example, 3. The criteria for determining the optimal push strategy are, for example, that the fitness ranks in the top 5% of the population and that the fitness does not decrease significantly for 3 consecutive iterations.

[0038] Step 105: Replace at least part of the target push strategy with the optimal push strategy, and continue to push information to the user.

[0039] In this embodiment, the aforementioned execution entity can replace at least part of the target push strategy with the optimal push strategy and continue to push information to the user.

[0040] The optimal push strategy generated through iterations undergoes compliance, stability, user experience, and security assessments before being officially added to the strategy library. The security assessment combines automated testing with manual sampling. Automated testing focuses on detecting issues such as exceeding frequency limits, content violations, and channel abuse. Manual sampling involves full manual review of strategies for high-risk business scenarios. Simultaneously, inefficient strategies with adaptability below a preset threshold, poor push performance, or negative user feedback are removed from the strategy library, resulting in a dynamic update of the strategy library. Information is then pushed to users based on the updated strategy library. The strategy library update employs a dual-library switching mechanism: a production library and a backup library are maintained in parallel. New strategies are added to the backup library and verified before seamlessly switching to the production library, preventing push service interruptions caused by strategy updates.

[0041] In some embodiments, the push policy replacement step may be as follows: First, the optimal push strategy that has passed the security assessment will be added to the strategy library.

[0042] The security assessment includes, but is not limited to, policy compliance verification, push notification frequency control verification, and user disturbance risk verification to ensure that newly added policies comply with platform specifications and user experience requirements. The user disturbance risk verification will combine users' historical negative feedback records to set a push notification tolerance score for each user. When the push notification frequency corresponding to a policy exceeds the user's tolerance threshold, the policy will be prohibited from pushing to that user.

[0043] Then, target push strategies with fitness values ​​below a preset threshold are removed from the strategy library.

[0044] The preset threshold is set based on business metrics and historical data. A fitness level below this threshold indicates poor strategy delivery performance, high resource consumption, and numerous negative user feedback, classifying it as an inefficient strategy and removing it. For deleted inefficient strategies, the system automatically generates a removal report, recording its lifecycle, fitness changes in each iteration, and reasons for removal, for use in subsequent algorithm optimization and strategy review.

[0045] This disclosure provides an information push method that effectively solves problems such as scenario cold start, strategy aging, and adaptive lag in traditional push systems by constructing an evolvable strategy library, real-time environment awareness, multi-objective effect evaluation, CMA-ES automatic iterative optimization, and dynamic updating of the strategy library, thereby significantly improving push accuracy and user end-to-end efficiency.

[0046] Continue to refer to Figure 2 This illustrates a flow 200 of another embodiment of the information push method according to the present disclosure. The information push method includes the following steps: Step 201: Push information to the user using a target push strategy that matches the user's current environmental state.

[0047] Step 202: Collect multi-dimensional target feedback information from users regarding the pushed information.

[0048] Step 203: Determine the suitability of the target push strategy based on multi-dimensional target feedback information.

[0049] In this embodiment, the specific operations of steps 201-203 have been described. Figure 1 The steps 101-103 in the illustrated embodiments are described in detail and will not be repeated here.

[0050] Step 204: Based on the current mean vector and covariance matrix, sample in the push strategy parameter space to generate a new push strategy.

[0051] In this embodiment, the aforementioned execution entity can sample in the push strategy parameter space based on the current mean vector and covariance matrix to generate a new push strategy.

[0052] The current mean vector and covariance matrix represent the distribution parameters of the current policy descriptor population in the push policy parameter space. Based on these distribution parameters, probability sampling is performed in a continuous push policy parameter space to generate a new generation of policy descriptor population, i.e., a new push policy. Each evolutionary iteration generates a new batch of policy individuals.

[0053] Step 205: Evaluate the adaptability of the new push strategy and select the new push strategy with the highest adaptability.

[0054] In this embodiment, the aforementioned execution entity can evaluate the adaptability of the new push strategy and select the new push strategy with the highest adaptability.

[0055] The newly generated strategy descriptor is decoded into an executable push strategy, and after sandbox security testing, it is executed for online push. The sandbox security test is a simulation test based on real online user data; for example, test users represent 1% of the total users, and stratified sampling is used to ensure the representativeness of the test results. The sandbox test lasts for, for example, 30 minutes, during which only feedback data is collected, and no real push content is shown to users. Corresponding multi-dimensional target feedback information is collected, and the fitness of the new push strategy is calculated using weighted summation or Pareto front methods. The strategies are then sorted from high to low fitness, and the individual new push strategy with the highest fitness in this iteration is selected as the basis for subsequent updates to the search distribution.

[0056] Step 206: Based on the selection results, adaptively update the mean and covariance matrix of the search distribution so that the search direction evolves towards a higher fitness region.

[0057] In this embodiment, the aforementioned execution entity can adaptively update the mean and covariance matrix of the search distribution based on the selection result, so that the search direction evolves towards a higher fitness region.

[0058] Based on the distribution characteristics of the policy individual with the highest fitness, the mean vector and covariance matrix of the search distribution are adaptively updated. The search distribution represents the search range and direction of the CMA-ES algorithm in the policy parameter space. After the update, the algorithm will continue to converge towards regions with higher fitness, improving the efficiency and accuracy of policy optimization.

[0059] Step 207: Adaptively adjust the search step size based on the evolutionary path information.

[0060] In this embodiment, the aforementioned execution entity can adaptively adjust the search step size based on the evolution path information.

[0061] The search step size is intelligently adjusted based on the convergence effect and iteration status of the evolutionary path. The evolutionary path includes successful paths and failed paths. A successful path is an iterative process of continuous fitness improvement over multiple rounds, while a failed path is an iterative process of continuous fitness decline over multiple rounds. For example, when the length of a successful path exceeds 5 rounds, the step size increases by 10%; when the length of a failed path exceeds 3 rounds, the step size decreases by 20%. The search step size represents the optimization magnitude and aggressiveness of each round of evolution in the CMA-ES algorithm. An excessively large step size can easily overshoot the optimal solution, while an excessively small step size results in slow convergence. Adaptive adjustment effectively balances the breadth of exploration and the accuracy of utilization, improving evolutionary stability.

[0062] Step 208: In response to the detection of a preset emergency event in the current environmental state, increase the exploration weight and search step size of CMA-ES.

[0063] In this embodiment, the aforementioned execution entity may increase the exploration weight and search step size of CMA-ES in response to the detection of a preset sudden event in the current environmental state.

[0064] Pre-defined contingency events include, but are not limited to, significant environmental changes such as new business launches, new market expansions, the outbreak of trending events, significant shifts in user interests, and the launch of new features. Monitoring of these pre-defined contingency events employs feature matching and threshold judgment. The system pre-stores feature vectors for various contingency events. When the similarity between the environmental state vector and the feature vector exceeds a preset similarity threshold, and the key indicator (such as the search volume for trending events) exceeds a preset indicator threshold, a contingency event is determined to have occurred. Upon detecting the above situations, the exploration weight and search step size of CMA-ES are automatically increased. The exploration weight is a weighting coefficient of key business indicators in multi-objective fitness calculation; increasing the step size accelerates evolutionary convergence, shortens adaptation time to new environments, and speeds up strategy evolution.

[0065] Step 209: In response to the abnormal breakage signal of the user's chain being identified by the change point detection algorithm, information is pushed to the user using a compensation push strategy for the abnormal breakage signal.

[0066] In this embodiment, the aforementioned execution entity can respond to the abnormal breakage signal of the user's receiving chain being identified through the change point detection algorithm, and push information to the user using a compensation push strategy for the abnormal breakage signal.

[0067] The user engagement chain refers to the entire user behavior process from receiving push notifications, viewing them, clicking to interact, to completing the conversion. Abnormal breaks in the user engagement chain include, but are not limited to, sudden drops in click-through rates for specific category push notifications, sharp decreases in conversion rates, a surge in negative user feedback, and interruptions in the push engagement chain. This application employs a Bayesian Online Changepoint Detection (BOCD) algorithm. When the change in key metrics within the algorithm's detection window exceeds a preset threshold, it is identified as a change point, i.e., an abnormal break signal. After identifying the break signal through the change point detection algorithm, a high-quality backup strategy is immediately invoked from the historically evolved strategy library, or a targeted patch-style compensation push strategy is quickly generated to rapidly repair the break in the engagement chain and ensure stable push notification performance.

[0068] Step 210: Increase the weight of abnormal breakage signals and compensation push strategies.

[0069] In this embodiment, the aforementioned execution entity can increase the weight of abnormal break signals and compensation push strategies.

[0070] Using the abnormal break signals and the execution results of the compensation push strategy as strong feedback signals, their weights are significantly increased in the multi-objective fitness calculation of CMA-ES, guiding the strategy evolution towards a more robust direction. For abnormal categories that are difficult to process through weighting, these categories can be isolated and independently controlled from the overall evolutionary logic to prevent local anomalies from lowering the overall push effect. The isolation process employs an independent population mechanism, i.e., a separate strategy population is built for abnormal categories, and an independent CMA-ES iterative optimization process is used. Once the category returns to normal, its strategy population is then merged into the global population.

[0071] Step 211: After multiple rounds of iterative optimization, the optimal push strategy is generated.

[0072] In this embodiment, the aforementioned execution entity can generate the optimal push strategy through multiple rounds of iterative optimization.

[0073] The entire evolutionary process of repeatedly sampling new populations, evaluating fitness, selecting the best individuals, updating search distribution, adjusting search step size, responding to environmental changes, and compensating for breaks is executed. After multiple rounds of iterative optimization, the optimal push strategy that adapts to the current environment, user needs, and business goals is finally generated.

[0074] Step 212: Replace at least part of the target push strategy with the optimal push strategy, and continue to push information to the user.

[0075] In this embodiment, the specific operation of step 212 has been described. Figure 1 Step 105 in the illustrated embodiment is described in detail and will not be repeated here.

[0076] This disclosure provides an information push method that further refines the CMA-ES iterative optimization process, adds an environmental mutation evolution acceleration mechanism and a user connection chain breakage adaptive compensation mechanism, forming a complete closed loop of perception-decision-execution-evaluation-evolution-compensation, realizing unsupervised, adaptive, and continuously evolving push strategy optimization, and fundamentally solving the problems of push strategy cold start and dynamic adaptation.

[0077] Further reference Figure 3The diagram illustrates the overall architecture of the push-evolution system based on the AlphaEvolve mechanism. It fully demonstrates the four-layer architecture and closed-loop data flow of this disclosed push-evolution system: the system is divided into an input layer, a core layer, an output layer, and a feedback layer. The input layer is responsible for real-time sensing of user terminal signals, cloud business signals, in-application behavior, device status, activity configuration, and content library information, and generates an environmental state vector through the signal fusion module. The core layer includes the AlphaEvolve policy meta-controller, push decision generator, and push executor. Based on the environmental state vector, it selects an appropriate policy from the policy library to complete the decision generation and push execution for push timing, channel, content, and frequency control. The output layer delivers information to user devices. The feedback layer acquires multi-objective reward signals such as CTR, CVR, retention, LTV, and negative feedback through a multi-dimensional feedback collector. On the one hand, it identifies abnormal breaks and triggers adaptive compensation strategies by receiving a breakpoint detector. On the other hand, it inputs the feedback into the CMA-ES evolution core to complete multi-objective fitness calculation, covariance matrix, and mean vector updates. The evolution acceleration controller adjusts the CMA-ES hyperparameters and search step size according to the environmental mutation signal, and finally updates the policy library and guides the next round of population sampling to achieve closed-loop adaptive evolution optimization throughout the entire process.

[0078] Further reference Figure 4 The diagram illustrates a policy evolution cycle based on CMA-ES. It details the internal process of CMA-ES-driven policy iterative optimization: the process begins with policy population initialization, generating N policy descriptor vectors; then, the population fitness is evaluated, and it is determined whether significant environmental mutations are detected. If mutations are detected, the search step size and exploration weight are increased to accelerate evolution; a new generation of policy population is generated based on the mean vector μ, covariance matrix C, and search step size σ of the current search distribution; the new policies undergo sandbox environment safety testing to assess potential risks and benefits; excellent new policies are adopted and updated to the production policy library, and the new policies participate in online push decisions; after completing one round of evolution, the mean vector of the search distribution is updated to move towards excellent individuals, the covariance matrix is ​​updated to adjust the direction and shape of the search ellipsoid, and the evolutionary path and step size are updated to balance exploration and utilization; the system continuously monitors environmental change indicators and repeatedly executes the above process, enabling the policy population to continuously evolve towards high-fitness regions, achieving dynamic, efficient, and adaptive updates to the policy library.

[0079] Furthermore, the information push method provided in this disclosure can be widely applied to various business scenarios, with typical application areas as follows: Application Area 1: Intelligent marketing push on e-commerce platforms.

[0080] Key challenges in the e-commerce sector: User interests shift rapidly, product lifecycles are short, and promotional activities are frequent. Fixed-rule push notifications (such as "browse but don't buy" reminders) can easily become overly intrusive and cause user resentment, while models based on historical sales / clicks cannot respond promptly to new product launches, price changes, or sudden trending topics (such as products popularized on social media).

[0081] The application process of this algorithm is as follows: Strategy Definition and Encoding: Encode a push strategy into a multi-dimensional vector. Dimensions include, but are not limited to: push timing (e.g., N hours after the user's last activity), triggering conditions (e.g., X% price reduction of items in the shopping cart), product category preference weight, copy template ID, coupon amount, etc.

[0082] Environmental status input: Real-time system input includes: real-time user profile (recent browsing / search / cart history), current time period (weekday / holiday), platform activity status (such as major promotion period), and real-time product inventory and price.

[0083] Evolutionary decision-making and execution: The system selects or combines the optimal policy vector for the user from the current policy population based on the environmental state assessment.

[0084] Push notifications: For example, during a user's lunch break (timing), push a "limited-time discount" promotion (text + offer) for the "coffee machine" category (conditions) that they have recently viewed.

[0085] Feedback Evaluation and Evolution: Track whether users click, add to cart, or place an order. Transform these behaviors into multi-objective rewards (click-through rate, conversion rate).

[0086] CMA-ES Driven Evolution: The CMA-ES algorithm treats policy vectors as continuous parameters to be optimized. The algorithm continuously evaluates the average performance (fitness) of different policy vectors across a user group. It performs an efficient search in the policy space by adaptively adjusting the mean of the sampling distribution (moving towards better-performing policies) and the covariance matrix (exploring other policy dimensions related to well-performing policies). For example, the algorithm might discover that during a major promotional period, the strategy combination of "high coupon amount + urgency copywriting" generally performs well for "high-value users," and then quickly adjust the policy population to ensure that this type of strategy is generated and adopted more frequently.

[0087] Application Area 2: Personalized content distribution on content information and short video platforms.

[0088] Key challenges in this field: Content consumption is highly time-sensitive and driven by exploratory interests. Users demand freshness from content, and their interests may shift rapidly with trending events. Traditional collaborative filtering or deep learning models fall short in capturing sudden interest and addressing the cold start of long-tail content.

[0089] The application process of this algorithm is as follows: Strategy Definition and Coding: Strategy vectors cover content type weight (news, entertainment, technology), timeliness coefficient (preference for the latest or classic content), exploration factor (how much unexposed long-tail content to inject), push frequency, cover image style preference, etc.

[0090] Environmental status inputs: real-time user conversation behavior (swiping speed, dwell time), social trending topics list, and real-time submissions from the content production end.

[0091] Evolutionary decision-making and execution: For the user's current session, the system selects a strategy vector, filters and sorts the content pool to generate a push list.

[0092] Feedback Evaluation and Evolution: Multi-objective rewards are based on user completion rate, likes, comments, shares, and retention in subsequent conversations.

[0093] CMA-ES-Driven Evolution: When sudden social hot topics emerge, some users may experience a sudden surge in interest in related news content. Traditional models require time to retrain. This system, through CMA-ES, can quickly evaluate the adaptability of a strategy variant—"temporarily increasing the weight of news items and significantly improving the timeliness coefficient"—in the current environment. If this strategy performs well, CMA-ES will rapidly adjust its search distribution, generating more strategies with similar characteristics in the strategy population. This allows the system as a whole to "perceive" and "adapt" to hot topics, achieving rapid response and accurate distribution of hot content.

[0094] Application Area 3: Improving player retention and activity in mobile games and applications.

[0095] Key pain points: Players require differentiated outreach strategies at different stages of their lifecycle (new player, growth, and stagnation). Inappropriate timing (such as pushing notifications while players are at work) or mismatched incentives (pushing highly paid content to free players) can lead to ineffective push notifications or even player churn.

[0096] The application process of this algorithm is as follows: Strategy definition and coding: The strategy vector includes the logic for judging the player's life cycle stage, the best time window for reaching, the type of incentive (stamina gift, item reward, level hint), and the emotional tone of the recall copy (encouraging, informing, benefit-oriented).

[0097] Environmental status input: player's last online time, current level, payment history, recent activity decline trend, device type / time.

[0098] Evolutionary Decision-Making and Execution: For a player at risk of churn, the system selects a strategy, such as sending a "item pack" (incentive) that the player urgently needs for the current level via in-app push notification (channel) half an hour before their usual online time (timing), accompanied by the message "We have prepared a boost for you to pass the level!"

[0099] Feedback Evaluation and Evolution: The core evaluation objective is whether players return within a specified time, their active duration after returning, and their spending behavior.

[0100] CMA-ES Drives Evolution: The CMA-ES algorithm continuously explores the high-dimensional space of "player reactivation strategies." For example, it simultaneously tests various strategy variations such as "pushing nostalgic content late at night," "pushing new events during weekday lunch breaks," and "pushing login rewards on weekend mornings." By analyzing the feedback of different player segments to different strategies, CMA-ES can adaptively learn the optimal strategy distribution. For example, for "moderately paying players," pushing "exclusive benefits for the new version" through "system notifications" within "24 hours of the version update" is the most effective reactivation strategy, and this knowledge is solidified at the center of the strategy population.

[0101] Further reference Figure 5 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of an information push device, which is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0102] like Figure 5 As shown, the information push device 500 of this embodiment may include: a push module 501, a collection module 502, a determination module 503, an iteration module 504, and a replacement module 505. The push module 501 is configured to push information to the user using a target push strategy that matches the user's current environmental state; the collection module 502 is configured to collect multi-dimensional target feedback information from the user regarding the pushed information; the determination module 503 is configured to determine the fitness of the target push strategy based on the multi-dimensional target feedback information; the iteration module 504 is configured to perform iterative optimization using a covariance matrix adaptive evolution strategy (CMA-ES) based on the fitness of the target push strategy to generate an optimal push strategy; and the replacement module 505 is configured to replace at least part of the target push strategy with the optimal push strategy and continue pushing information to the user.

[0103] In this embodiment, the specific processing of the push module 501, the acquisition module 502, the determination module 503, the iteration module 504, and the replacement module 505 in the information push device 500, and the resulting technical effects, can be found in reference to [reference needed]. Figure 1The relevant descriptions of steps 101-105 in the corresponding embodiments will not be repeated here.

[0104] In some optional implementations of this embodiment, the iteration module 504 is further configured to: sample in the push strategy parameter space based on the current mean vector and covariance matrix to generate a new push strategy; evaluate the fitness of the new push strategy and select the new push strategy with the highest fitness; adaptively update the mean and covariance matrix of the search distribution according to the selection result, so that the search direction evolves towards a higher fitness region; adaptively adjust the search step size according to the evolution path information; and generate the optimal push strategy after multiple rounds of iterative optimization.

[0105] In some optional implementations of this embodiment, the iteration module 504 is further configured to: increase the exploration weight and search step size of CMA-ES in response to detecting a preset sudden event in the current environmental state.

[0106] In some optional implementations of this embodiment, the iteration module 504 is further configured to: in response to the detection of an abnormal breakage signal in the user's support chain by the change point detection algorithm, push information to the user using a compensation push strategy for the abnormal breakage signal; Increase the weight of abnormal break signals and compensation push strategies.

[0107] In some optional implementations of this embodiment, the push module 501 is further configured to: generate the current environment state based on the user's behavior information; select a push strategy that matches the current environment state from the strategy library as the target push strategy; and push information to the user using the target push strategy.

[0108] In some optional implementations of this embodiment, the determining module 503 is further configured to: calculate the fitness of the target push strategy based on multi-dimensional target feedback information using a weighted summation or Pareto front method.

[0109] In some optional implementations of this embodiment, the replacement module 505 is further configured to: add the optimal push strategy that has passed the security assessment to the strategy library; and delete the target push strategy with a fitness level lower than a preset threshold from the strategy library.

[0110] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0111] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0112] Figure 6A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0113] like Figure 6 As shown, device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or a computer program loaded into random access memory (RAM) 603 from storage unit 608. RAM 603 may also store various programs and data required for the operation of device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0114] Multiple components in device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0115] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as the information push method. For example, in some embodiments, the information push method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program may be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the information push method described above may be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform the information push method by any other suitable means (e.g., by means of firmware).

[0116] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0117] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0118] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0119] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0120] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0121] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0122] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution provided in this disclosure can be achieved, and this is not limited herein.

[0123] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An information push method, comprising: Information is pushed to the user using a target push strategy that matches the user's current environmental state; Collect multi-dimensional target feedback information from users regarding the pushed information; Based on the multi-dimensional target feedback information, the adaptability of the target push strategy is determined; Based on the fitness of the target push strategy, the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) is used for iterative optimization to generate the optimal push strategy. The optimal push strategy is used to replace at least part of the target push strategy, and information continues to be pushed to the user.

2. The method according to claim 1, wherein, The optimal push strategy is generated by iterative optimization using the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) based on the fitness of the target push strategy, including: Based on the current mean vector and covariance matrix, a new push strategy is generated by sampling in the parameter space of the push strategy. Evaluate the adaptability of the new push strategy and select the new push strategy with the highest adaptability; Based on the selection results, the mean and covariance matrix of the search distribution are adaptively updated, causing the search direction to evolve towards a higher fitness region; The search step size is adaptively adjusted based on evolutionary path information. After multiple rounds of iterative optimization, the optimal push strategy was generated.

3. The method according to claim 2, wherein, The step of iteratively optimizing the optimal push strategy based on the fitness of the target push strategy using the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) to generate the optimal push strategy also includes: In response to the detection of a preset sudden event in the current environmental state, the exploration weight and search step size of CMA-ES are increased.

4. The method according to claim 2, wherein, The step of iteratively optimizing the optimal push strategy based on the fitness of the target push strategy using the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) to generate the optimal push strategy also includes: In response to the detection of an abnormal breakage signal in the user's chain by the change point detection algorithm, information is pushed to the user using a compensation push strategy for the abnormal breakage signal; Increase the weight of the abnormal breakage signal and the compensation push strategy.

5. The method according to claim 1, wherein, The method of pushing information to the user using a target push strategy that matches the user's current environmental state includes: The current environment state is generated based on the user's behavior information; Select a push strategy from the strategy library that matches the current environment state as the target push strategy; The target push strategy is used to push information to the user.

6. The method according to claim 1, wherein, The step of determining the fitness of the target push strategy based on the multi-dimensional target feedback information includes: Based on the multi-dimensional target feedback information, the fitness of the target push strategy is calculated using a weighted summation or Pareto front method.

7. The method according to claim 5, wherein, The step of replacing at least a portion of the target push strategy with the optimal push strategy includes: The optimal push strategy that has passed the security assessment will be added to the strategy library; Remove target push strategies with fitness values ​​below a preset threshold from the strategy library.

8. An information push device, comprising: The push module is configured to push information to the user using a target push strategy that matches the user's current environmental state; The data collection module is configured to collect multi-dimensional target feedback information from the user regarding the pushed information; The determination module is configured to determine the fitness of the target push strategy based on the multi-dimensional target feedback information. The iterative module is configured to perform iterative optimization using the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) based on the fitness of the target push strategy to generate the optimal push strategy. The replacement module is configured to replace at least a portion of the target push strategy with the optimal push strategy and continue to push information to the user.

9. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1-7.

11. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-7.

Images