Training of a push user screening model and information stream pushing method and device, equipment, medium and product

Abstract

The present disclosure provides a training method and device of a push user screening model and an information flow pushing method and device, equipment, medium and product, relates to the technical field of computers, and particularly relates to the technical fields of artificial intelligence, big data, information flow, deep learning, neural networks, intelligent pushing and cloud computing. The specific implementation scheme is as follows: sample information flow containing target information content is pushed to the first screen of the information flow application corresponding to the sample information flow brushing times of a sample user; whether the target information content promotes the sample user to become an active user is determined according to the interaction behavior data of the sample user in the sample information flow brushing times, and a standard pushing label corresponding to the target information content is generated according to the determination result; sample user features are input into a to-be-trained model to obtain a predicted pushing probability corresponding to the target information content, and a push user screening model is trained according to the predicted pushing probability and the standard pushing label. The present disclosure realizes the effect of improving the number of active users in the information flow application.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the fields of artificial intelligence, big data, information flow, deep learning, neural networks, intelligent push and cloud computing, and especially to a method, apparatus, device, medium and product for training a push user screening model and pushing information flow. Background Technology

[0002] In information flow recommendation systems, platforms continuously introduce new formats or styles of content, such as short videos, interactive videos, live stream clips, and information cards. The decision to prioritize these new contents on the user's first screen during the current refresh is crucial to increasing user engagement.

[0003] The attractiveness of the first screen content directly affects users' willingness to use the platform, thus influencing the conversion rate of active users. Therefore, how to scientifically formulate a first-screen penetration strategy for new content while ensuring user experience is a common technical focus in platform operations. Summary of the Invention

[0004] This disclosure provides a method, apparatus, device, and storage medium for increasing the number of active users in a news feed application.

[0005] According to one aspect of this disclosure, a method for training a push user filtering model is provided, comprising: Obtain a sample information stream containing target information content, and push the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user swipes the sample information stream; Based on the interaction behavior data of the sample users in the sample information stream, determine whether the target information content has led the sample users to become active users, and generate standard push tags corresponding to the target information content in the sample users' sample information stream based on the determination result. The sample user features corresponding to the sample user's refresh rate in the sample information stream are input into the model to be trained to obtain the predicted push probability of the target information content corresponding to the sample user's refresh rate in the sample information stream. Based on the predicted push probability and the standard push tag, the model to be trained is trained to obtain the push user screening model corresponding to the target information content.

[0006] According to another aspect of this disclosure, an information feed push method is provided, comprising: Obtain the target user characteristics corresponding to the current information flow refresh time; The target user features are input into the push user filtering model to obtain the predicted push probability of the target information content corresponding to the current information stream refresh of the target user; wherein, the push user filtering model is trained using any of the push user filtering model training methods described above; If the predicted push probability is greater than or equal to the probability threshold, then the target information stream containing the target information content is obtained, and the target information stream is pushed to the target user's first screen of the information stream application corresponding to the current information stream refresh.

[0007] According to another aspect of this disclosure, a training apparatus for a push user selection model is provided, comprising: The sample information stream acquisition module is used to acquire a sample information stream containing target information content and push the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user swipes the sample information stream; The standard push tag generation module is used to determine whether the target information content has led the sample user to become an active user based on the interaction behavior data of the sample user in the sample information stream, and generate a standard push tag corresponding to the target information content in the sample user's sample information stream based on the determination result. The model training module is used to input the sample user features corresponding to the sample user when the sample information stream is refreshed into the model to be trained, to obtain the predicted push probability of the target information content corresponding to the sample user's refresh rate in the sample information stream, and to train the model to be trained based on the predicted push probability and the standard push tag to obtain the push user screening model corresponding to the target information content.

[0008] According to another aspect of this disclosure, an information stream push device is provided, comprising: The target user feature acquisition module is used to acquire the target user features corresponding to the current information stream refresh. The predictive push probability determination module is used to input the target user features into the push user filtering model to obtain the predicted push probability of the target information content corresponding to the current information stream refresh of the target user; wherein, the push user filtering model is trained using any of the push user filtering model training methods described above. The target information stream push module is used to obtain a target information stream containing the target information content if the predicted push probability is greater than or equal to a probability threshold, and push the target information stream to the target user's first screen of the information stream application corresponding to the current information stream refresh.

[0009] According to another aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor to enable the at least one processor to perform any of the methods of this disclosure.

[0010] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform any of the methods of this disclosure.

[0011] According to another aspect of this disclosure, a computer program product is provided, including a computer program and a method for the computer program to be executed by a processor according to any of the methods disclosed herein.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a flowchart of a training method for a push user screening model provided according to an embodiment of this disclosure; Figure 2 This is a flowchart of another method for training a push user screening model according to an embodiment of this disclosure; Figure 3 This is a flowchart of another method for training a push user screening model according to an embodiment of this disclosure; Figure 4 This is a flowchart of an information stream push method provided according to an embodiment of the present disclosure; Figure 5 This is a schematic diagram of an overall information flow push process provided according to an embodiment of this disclosure; Figure 6 This is a schematic diagram illustrating an information flow push decision based on an embodiment of this disclosure; Figure 7 This is a schematic diagram of the structure of a training device for a push user screening model provided according to an embodiment of the present disclosure; Figure 8 This is a schematic diagram of the structure of an information stream push device according to an embodiment of the present disclosure; Figure 9 This is a block diagram of an electronic device used to implement the training method of the push user screening model and / or the information flow push method of the embodiments of this disclosure. Detailed Implementation

[0014] Existing technologies primarily use mapping transformations to make the features of new information content approximate the distribution of existing popular information content, thus allowing existing recommendation models to understand and rank them based on users' historical interests and preferences. However, this approach has the following drawbacks: First, most existing technologies optimize intermediate metrics such as click-through rate and duration of consumption, or business objectives such as advertising revenue, but fail to use the ultimate business objective of "whether users will be converted into active users" as a direct monitoring signal for the model, resulting in strategies that cannot accurately serve the growth of active users.

[0015] Second, existing technologies do not combine the special position of "first screen of information flow application" with the special scenario of "cold start of new information content" for joint modeling, which has extremely high user attention and retention value, resulting in insufficient utilization efficiency of first screen exposure resources.

[0016] 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.

[0017] Figure 1 This is a flowchart illustrating a training method for a push user filtering model according to an embodiment of this disclosure. This embodiment is applicable to situations where intelligent first-screen penetration decisions are made to increase the number of active users when new information content is introduced into an information flow recommendation platform. The method can be executed by a training device for a push user filtering model, which can be implemented in hardware and / or software and can be configured in an electronic device. (Reference) Figure 1 The method specifically includes the following: S101. Obtain a sample information stream containing the target information content, and push the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user swipes the sample information stream.

[0018] S102. Based on the interaction data of sample users in the sample information stream, determine whether the target information content prompts the sample users to become active users, and generate standard push tags corresponding to the target information content in the sample information stream of the sample users based on the determination results.

[0019] S103. Input the sample user features corresponding to the sample user's refresh rate in the sample information stream into the model to be trained, obtain the predicted push probability corresponding to the target information content refresh rate in the sample user's sample information stream, and train the model to be trained based on the predicted push probability and standard push tags to obtain the push user selection model corresponding to the target information content.

[0020] In S101, the target information content refers to a specific content unit that needs to be evaluated for push effectiveness and for a decision to push it to the user. Target information content can be newly introduced content formats in the information flow recommendation platform, such as short videos, interactive videos, live stream clips, and information cards, or it can be existing information content that needs to be prioritized for push or for which push effectiveness needs to be evaluated, such as specific types of products, articles, and advertisements. During the model training phase, the target information content is inserted into the sample information flow to trigger user interaction and collect training data; during the model application phase, the target information content is used as the object to be pushed, and the decision to push it to the target user is based on the predicted push probability output by the model.

[0021] A sample information stream refers to the collection of information streams pushed to sample users during the model training phase to collect training data. The sample information stream is generated by inserting target information content into the original content sequence intended for push to sample users. The sample information stream includes the target information content and other regular content, and is pushed as a whole to the sample user's first screen of the information stream application to trigger interactive behavior data from sample users, including but not limited to clicks, refreshes, and satisfied purchases, thereby providing basic sample data for model training.

[0022] Sample users refer to users selected as test subjects and participating in data collection during the model training phase. Sample users are typically randomly selected from all users on the platform, such as 5% of users. Their interaction behavior data across different feed refreshes is used to generate labels for training samples. Sample users do not refer to a specific, fixed group of users; the same user may be labeled with different sample tags at different times and in different scenarios, thus providing diverse training data for the model.

[0023] A sample feed refresh refers to a single exposure opportunity corresponding to a user's content request after a feed application is launched, refreshed, or pulled down. Each feed refresh is treated as an independent processing unit, corresponding to a training sample, used to determine whether pushing target information content in that request effectively encourages the user to become an active user. The same user may be labeled with different sample tags in different feed refreshes due to differences in user status and context.

[0024] The first screen of a news feed app refers to the first screen area that users see directly after launching the app, refreshing, or pulling down, without needing to swipe. The first screen of a news feed app has high user attention and significant retention value, making it a key exposure position on news feed recommendation platforms.

[0025] In one implementation of S101, during the model training phase, a predetermined proportion of users are randomly selected from all users of the information flow recommendation platform (hereinafter referred to as the platform) as sample users. When a sample user's information flow refresh occurs, the initial information flow originally to be pushed by the platform for that refresh is obtained. This initial information flow contains multiple regular information contents sorted according to a conventional recommendation strategy. Then, according to a preset insertion strategy, the target information content is inserted into a specified position in the initial information flow to generate a sample information flow containing the target information content. Finally, the sample information flow is pushed to the sample user's information flow application homepage for display in this refresh. The insertion position of the target information content can be dynamically adjusted according to experimental needs, such as inserting it into the first, third, or random position on the homepage to collect user interaction data under different exposure positions.

[0026] In another implementation of S101, during the model training phase, a predetermined proportion of users are randomly selected from all platform users as sample users. When a sample user's feed refresh occurs, a set of candidate content to be pushed is directly obtained from the content pool. This set of candidate content includes target information content and other regular information content. Then, all content in the candidate content set is uniformly sorted according to a preset sorting strategy to generate a sample feed containing target information content. The target information content participates in the sorting together with the regular information content, and its final sorting position depends on the calculation result of the sorting strategy. Finally, the sample feed is pushed to the sample user's feed application homepage corresponding to this refresh for display. In this implementation, the target information content no longer relies on external insertion but participates in the sorting as one of the regular candidate contents, thereby evaluating its push effect in a more natural recommendation process.

[0027] In section S102, interactive behavior data refers to various behavioral records generated by sample users during and after browsing the sample information stream. Interactive behavior data includes, but is not limited to, immediate interactive behaviors, such as clicking and refreshing information content in the information stream, as well as subsequent consumption behaviors, such as liking, commenting, sharing, and watching beyond the allotted time. Interactive behavior data is the primary basis for determining whether a push notification effectively helps sample users become active users, and it is also the foundation for generating standard push notification tags.

[0028] Active users refer to users whose behavior meets preset activity criteria after a push notification event. These criteria are directly related to core platform metrics, such as the number of active users, and are used to measure whether the push notification effectively increased user activity. In this solution, "active user" is a result-based judgment based on a specific number of times the user browses the feed, rather than a profile label of overall user attributes. The specific judgment rules can be configured according to platform operational goals and business needs, such as combining immediate interactive behavior with subsequent consumption behavior for comprehensive evaluation.

[0029] Standard push notification labels are supervisory signal labels generated based on the interaction behavior data of sample users during sample information stream refreshes, determining whether the push notification effectively helped a user become an active user. These labels are binary; if the target information content is determined to have helped a sample user become an active user, a positive sample label with a value of one is generated; otherwise, a negative sample label with a value of zero is generated. Standard push notification labels serve as the standard answer for model training, guiding the model to learn the mapping relationship between user characteristics and push notification effectiveness.

[0030] In one implementation of S102, after a sample user's sample information stream refresh occurs, the interactive behavior data generated by the sample user during that refresh is collected in real time. According to a preset judgment rule, if the interactive behavior data contains instant interactive behavior, it is determined that this push has helped the sample user become an active user, and a standard push label with a tag content of one is generated; if the interactive behavior data does not contain instant interactive behavior, it is determined that this push has not helped the sample user become an active user, and a standard push label with a tag content of zero is generated.

[0031] In another implementation of S102, after a sample user's sample information stream refresh occurs, the interactive behavior data generated by the sample user during that refresh is collected in real time, and the subsequent interactive behavior data within a preset time window after the refresh ends is monitored. If the interactive behavior data includes immediate interactive behavior, and the subsequent interactive behavior data includes satisfactory consumption behavior, then it is determined that this push has helped the sample user become an active user, and a standard push label with a tag content of one is generated; otherwise, it is determined that this push has not helped the sample user become an active user, and a standard push label with a tag content of zero is generated.

[0032] In section S103, "sample user features" refers to multi-dimensional feature data used to characterize the current state, historical behavioral patterns, and interaction preferences with information feed content of a sample user when the sample user refreshes the information feed multiple times. These features depict the real-time state and long-term habits of the sample user during interaction with the platform from different dimensions, collectively constituting the input basis for the model to judge "whether pushing target information content during the current information feed refresh is likely to make the user an active user." The selection of sample user features aims to comprehensively capture various factors that influence the user's potential to accept new content, including both the user's static attributes and dynamic behavioral patterns that change with time and context.

[0033] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this technical solution comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0034] The model to be trained refers to the initial deep learning model that has not yet completed parameter optimization. It is used to receive sample user features and output predicted push probabilities. This model typically adopts a deep neural network structure. During training, the model to be trained performs forward computation based on the input feature vector, outputs predicted push probabilities, and then compares the prediction results with standard push labels. Through the backpropagation algorithm, the model parameters are continuously adjusted, and the mapping relationship from user features to push effects is gradually learned.

[0035] The predicted push probability is a numerical value calculated by the model under training based on the characteristics of the input sample users. It measures the likelihood that pushing target information content in a sample information stream will lead to the sample users becoming active users. This probability value ranges from 0 to 1, with a higher value indicating that the model believes the push is more likely to result in active user conversion. The predicted push probability and the standard push label together constitute the supervision signal pair for model training. By continuously narrowing the gap between the predicted push probability and the standard push label, the model under training gradually acquires the ability to accurately predict push effects.

[0036] The push notification user selection model refers to a target model that can be used for online decision-making, obtained by optimizing parameters through a training process. The push notification user selection model is trained on a per-target information content basis; that is, each target information content corresponds to a dedicated push notification user selection model. During the online application phase, when a target user initiates a refresh of the information stream, the target user's characteristics are input into the push notification user selection model. The model outputs a predicted push probability for that target information content. The system then decides whether to push that target information content to the target user based on a comparison of this probability with a preset threshold.

[0037] In one implementation of S103, all sample data collected in the above process are divided into a training set and a validation set according to a preset ratio. During training, the sample user features corresponding to each sample information stream refresh in the training set are input into the model to be trained. The model performs forward calculation based on the input features and outputs the predicted push probability for the target information content in the current sample information stream refresh. The predicted push probability is compared with the standard push label corresponding to the sample information stream refresh, and the binary cross-entropy loss function is calculated. The Adam optimizer is used to iteratively update the model parameters through the backpropagation algorithm. After each training round, the model is evaluated using the validation set, and classification indicators such as the area under the ROC curve and accuracy are monitored. An early stopping method is used to prevent overfitting, that is, training is stopped when the validation set indicators no longer improve for several consecutive rounds, the optimal model parameters are saved, and finally the trained push user selection model is obtained.

[0038] After the model is trained, its actual effectiveness needs to be further verified through online A / B testing. The experimental group using this model to determine the penetration of target information content will be compared with a control group using a conventional recommendation strategy, focusing on improvements in core business metrics such as active users, average usage time per user, and next-day retention rate. If the experimental group significantly outperforms the control group in these metrics, it proves that this model can effectively achieve intelligent penetration of new content on the first screen and improve platform user activity.

[0039] This embodiment of the disclosure acquires a sample information stream containing target information content and pushes the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user browses the sample information stream; based on the interactive behavior data of the sample user during the number of times the sample user browses the sample information stream, it determines whether the target information content contributes to the sample user becoming an active user, and generates a standard push label corresponding to the number of times the target information content browses the sample user's sample information stream based on the determination result; the sample user features corresponding to the number of times the sample user browses the sample information stream are input into the model to be trained to obtain the predicted push probability corresponding to the number of times the target information content browses the sample user's sample information stream, and trains the model to be trained based on the predicted push probability and the standard push label to obtain a push user selection model corresponding to the target information content. The beneficial effects are: Firstly, this solution, through its core feature of "determining whether target information content leads to active users based on the interactive behavior data of sample users during sample information feed refreshes," directly transforms the model's learning objective from traditional intermediate indicators such as click-through rate (CTR) and consumption duration to the ultimate business objective of "whether it leads to users becoming active users," which is strongly correlated with the number of active users. The model is trained using standard push tags as supervisory signals, learning the mapping relationship between "whether pushing target information content during the current information feed refresh can lead to active user conversion," rather than simply fitting click probability or consumption depth. This aligns the model's optimization direction with user growth goals, resolving the potential deviation between CTR and active user numbers in traditional solutions. It ensures that push decisions accurately serve core business indicators, effectively increasing the number of active users in information feed applications.

[0040] Secondly, by learning the decision-making logic of "whether pushing target information content on the first screen of an information feed application with a high probability of user engagement can encourage user engagement to become an active user," the model can accurately identify high-value push scenarios. In practical applications, the predicted push probability output by the model is used to filter out information feeds with a higher probability of user engagement for first-screen penetration, while excluding low-probability information feeds from the penetration scope.

[0041] This mechanism brings two beneficial effects: 1) It concentrates limited and valuable first-screen exposure resources on high-value scenarios, avoiding ineffective exposure when users are not in a good state or the scenario is not suitable, significantly improving the distribution efficiency of target information content and the utilization value of first-screen resources. 2) By reducing forced pushes to uninterested users, it effectively reduces interference with the user's normal usage process, maintaining a good user experience while exploring target information content, and achieving a dual optimization of resource efficiency and user experience.

[0042] Figure 2 This is a flowchart illustrating another training method for a push user screening model provided according to an embodiment of this disclosure. It can be used to further optimize and extend the above technical solution and can be combined with the various optional implementation methods described above. (Reference) Figure 2 The method specifically includes the following: S201. Obtain the initial information stream to be pushed to the first screen of the information stream application when the sample information stream is refreshed by the sample user; insert the target information content into the initial information stream to obtain the sample information stream.

[0043] S202. Push the sample information stream to the home screen of the information stream application corresponding to the number of times the sample user refreshes the sample information stream.

[0044] S203. When the interaction behavior data includes first behavior data and second behavior data, determine the target information content that will encourage the sample user to become an active user.

[0045] S204. Based on the determined results, generate standard push tags corresponding to the number of times the target information content is refreshed in the sample information stream of the sample users.

[0046] S205. Input the sample user features corresponding to the sample user's refresh rate in the sample information stream into the model to be trained, obtain the predicted push probability corresponding to the target information content refresh rate in the sample user's sample information stream, and train the model to be trained based on the predicted push probability and standard push tags to obtain the push user selection model corresponding to the target information content.

[0047] In S201, the initial information stream refers to the sequence of information stream content that was originally intended to be pushed to the sample user according to the regular recommendation strategy when the sample user's information stream refresh occurs. This information stream is dynamically generated by the platform's recommendation system based on factors such as user profile, historical behavior, and scene characteristics. It contains multiple regular information contents arranged according to a specific sorting logic, but does not contain target information contents that need to be evaluated for push effect.

[0048] In one implementation of S201, when a sample user's feed is refreshed, the regular recommendation engine is first invoked to retrieve the initial feed originally to be pushed to that sample user during that refresh. Then, according to a preset insertion strategy, the target information content is inserted into a specified position in the initial feed, such as the first or third position on the first screen, or a random position, thereby generating a sample feed containing the target information content. Finally, this sample feed is pushed to the first screen of the feed application corresponding to that sample user's refresh for display.

[0049] In another implementation of S201, a predetermined proportion of users are randomly selected from all platform users as sample users. These sample users are then randomly divided into multiple experimental groups, each corresponding to a different target information content exposure strategy. When a sample user's sample information stream refreshes, the target information content and a sequence of regular information content generated according to the conventional recommendation strategy are directly retrieved from the content pool, based on the preset exposure strategy of the experimental group to which the sample user belongs. Subsequently, according to the fusion rules corresponding to the experimental group, the target information content and the regular information content are fused. For example, the target information content may replace content at a specific position in the regular information sequence, or the target information content may be inserted at a specific position in the regular information sequence, thereby generating a sample information stream that conforms to the strategy of that experimental group. Finally, the sample information stream is pushed to the first screen of the information stream application corresponding to the sample user in this refresh for display.

[0050] By obtaining the initial information stream to be pushed to the first screen of the information stream application when the sample user's information stream is refreshed, and inserting the target information content into the initial information stream to obtain the sample information stream, the beneficial effects are: Firstly, by first acquiring the initial information stream generated according to conventional recommendation strategies, the content distribution and ranking logic under natural recommendation conditions are preserved. Then, controlled intervention is carried out by inserting target information content. This "natural first, intervention later" generation method enables the subsequently collected interaction behavior data to clearly distinguish between the "effect of conventional content" and the "incremental impact of target content insertion," providing clean and attributable sample data for subsequent model training.

[0051] Secondly, since the initial information stream itself has already been generated according to conventional recommendation strategies and can basically meet the user's content needs in the current scenario, the insertion of target information content is a moderate intervention based on this. This generation method, while introducing new content, maintains the stability of the original recommendation experience to the greatest extent possible, avoiding fluctuations in user experience that may be caused by reconstructing the information stream.

[0052] In S203, the first line of data consists of the information click or refresh behavior generated by the sample user on the first screen of the information flow application.

[0053] Information click behavior refers to the active click action performed by sample users on any information content (including target information content and other regular information content) displayed on the first screen of the information feed application. This behavior indicates that the sample users have generated initial interest in specific information content and intend to view further details or interact with it, and is one of the core indicators for measuring users' willingness to interact in real time.

[0054] The sample users actively updated the content of their feed on the first screen of the app through various means, including pulling down the screen, clicking the refresh button, or other preset interactions. This behavior indicates that the user is dissatisfied with the content displayed on the current first screen or expects to access more new content, demonstrating the user's continued intention to use the app and seek better content during the current feed refresh.

[0055] The second behavioral data refers to the satisfactory consumption behavior of sample users towards any information content in the information flow application within a preset time period after the occurrence of the first behavioral data.

[0056] The preset time period refers to the time window used to observe whether sample users generate subsequent satisfactory consumption behavior after the first action data occurs. The length of this time window can be preset according to business characteristics and content consumption cycle, such as the current hour window, 30-minute window, or 2-hour window. The purpose of setting the preset time period is to measure whether users have generated in-depth consumption behavior within a reasonable duration after an instant interaction, thereby determining whether this push truly promoted user activity.

[0057] Satisfactory consumption behavior refers to the deep positive interaction behavior of sample users with any information content in a news feed application. It is used to measure the quality of users' information content consumption and the intensity of their interest. Satisfactory consumption behavior requires users to invest substantial attention in the information content or express a clear positive attitude, and can more accurately reflect whether users have gained satisfaction from this information content consumption.

[0058] Satisfactory consumer behaviors include liking, commenting, sharing, and watching for extended periods.

[0059] Liking refers to users expressing positive feedback by liking information content; commenting refers to users expressing subjective opinions such as text or emoticons on information content; sharing refers to users actively spreading information content by forwarding it to other platforms or other users.

[0060] The extended viewing time behavior indicates that the sample users' viewing time for the information content exceeds a preset multiple of the average viewing time for the information content, reflecting the user's deep immersion in the content and investment of interest.

[0061] Viewing duration refers to the length of time a user spends consuming a particular piece of information content, from the start to the end. Average viewing duration refers to the average viewing time of all users across all historical information content consumption, based on the content type of that information. The preset multiplier can be set according to needs, such as setting it to 1.5x.

[0062] In one implementation of S203, it is first detected whether the sample user's interaction behavior data includes first behavior data, i.e., the information click behavior or information refresh behavior generated by the sample user on the first screen of the information flow application. If the first behavior data is detected, it is further monitored whether the sample user has generated satisfactory consumption behavior for any information content in the information flow application within a preset time period after the occurrence of the first behavior data.

[0063] When the interaction behavior data of a sample user contains both first behavior data and second behavior data (i.e., satisfactory consumption behavior), it is determined that the target information content helped the sample user become an active user; conversely, if the interaction behavior data lacks first behavior data, or if there is first behavior data but no satisfactory consumption behavior is generated within the preset time period, it is determined that the target information content failed to help the sample user become an active user.

[0064] By analyzing interactive behavior data, including both first and second behavior data, the target information content is identified to encourage sample users to become active users. The first behavior data refers to the information clicks or refreshes made by sample users on the first screen of the information feed application. The second behavior data refers to the satisfied consumption behaviors of sample users towards any information content in the information feed application within a preset time period after the occurrence of the first behavior data. Satisfactory consumption behaviors include liking, commenting, sharing, and extended viewing time. Extended viewing time indicates that the sample user's viewing time for the information content exceeds a preset multiple of the average viewing time for that information content. The beneficial effects are: Firstly, using click behavior alone as the criterion for positive samples easily introduces a lot of noise such as accidental clicks and clickbait titles; using consumption behavior alone as the criterion for positive samples fails to capture the active value generated by users simply through clicks or refreshes. This solution combines immediate interactive behavior (first-order data) with subsequent satisfactory consumption behavior (second-order data), which not only retains the contribution of immediate interactive behavior to the number of active users, but also filters out invalid interactions through satisfactory consumption behavior, making the definition of positive samples purer and more accurate.

[0065] Secondly, the first set of behavioral data reflects the short-term activity value brought by the push notification; the second set of behavioral data reflects the sustained value brought by the push notification. The combination of the two allows the complete value chain of a push notification to be captured, and the learning objective of the model is upgraded from "whether the user clicks" to "whether the user truly becomes an active user because of this push notification".

[0066] Optionally, S204 includes: If the target information content is determined to have helped a sample user become an active user, a standard push label with a tag content of one is generated; if the target information content is determined not to have helped a sample user become an active user, a standard push label with a tag content of zero is generated.

[0067] In one implementation, if the interaction behavior data determines that the target information content helped a sample user become an active user, a standard push label with a content of one is generated for that sample information stream refresh, marking it as a positive sample. If the interaction behavior data determines that the target information content failed to help a sample user become an active user, a standard push label with a content of zero is generated for that sample information stream refresh, marking it as a negative sample. Through this method, each sample information stream refresh is assigned a binary standard push label. This label serves as a standard supervision signal for subsequent model training, guiding the model to learn the mapping relationship from sample user features to push effects.

[0068] If the target information content is determined to have helped sample users become active users, a standard push label with a tag of one is generated; if the target information content is determined not to have helped sample users become active users, a standard push label with a tag of zero is generated. The beneficial effects are: Firstly, the determination results of active users are uniformly converted into simple binary labels of "one" or "zero", so that subsequent model training can receive supervision signals in a standardized way without having to repeatedly execute complex determination logic during model training.

[0069] Secondly, the standard push label is strictly generated based on the business judgment result of "whether it leads to sample users becoming active users," ensuring that the model's learning objective is consistent with the business metric of the number of active users to be optimized. During the training process, the model continuously adjusts its parameters to accurately predict this label, essentially learning how to infer from user characteristics whether "the content of the push target information can lead to the conversion of active users," thus avoiding deviation between the optimization objective and the business objective.

[0070] Optionally, the "sample user characteristics" in S205 include at least one of the following: user profile characteristics, user scenario characteristics, information content consumption statistics, satisfied consumption information content statistics, and first-click information content statistics.

[0071] User profile features refer to static feature data that describes the basic attributes and long-term preferences of sample users. These features include, but are not limited to, demographic attributes, device information (such as mobile phone model, operating system, etc.), and long-term activity indicators (such as the number of active days in the past 30 days, average daily usage time, etc.). User profile features reflect the stable attributes of users and provide basic user background information for the model.

[0072] User scenario features refer to dynamic feature data describing the real-time state and environment of a sample user when the sample information stream is refreshed. These features include, but are not limited to, time information (such as current hour, weekday or weekend), network environment (such as WiFi or mobile network), geographic location context (such as city, business district), and device status (such as remaining battery power, screen brightness). User scenario features capture the specific context of the user's current refresh, helping the model understand the differences in user behavior patterns in different scenarios.

[0073] Information content consumption statistics refer to the quantitative characteristics formed by the content consumption behavior of statistical sample users within a historical time window for various types of information content in information flow applications. This characteristic is divided according to the type of information content, and quantitative indicators of user consumption behavior for each type of content are statistically analyzed under different time windows to characterize the degree of user preference and consumption quality for different types of content.

[0074] The statistical characteristics of satisfied consumption information content are the set of information content in which sample users generate satisfied consumption behaviors in information flow applications. Satisfactory consumption behaviors include liking, commenting, sharing, and extended viewing time. Extended viewing time means that the sample user's viewing time for any information content exceeds a preset multiple of the average viewing time of that information content.

[0075] The statistical characteristics of the first click information content are the set of information content that a sample user clicks for the first time in the historical information stream.

[0076] Historical feed refreshes refer to the historical exposure opportunities that a sample user had in the feed application before the current feed refresh. Each historical feed refresh corresponds to a content request made by a sample user when they launched or refreshed the feed application in the past, and records the user's behavioral data during that request.

[0077] The first click refers to the first click performed by a user on content within a single historical feed refresh. This click typically occurs shortly after the user launches the app or refreshes the feed, reflecting the user's most pressing content need or most interesting content direction during that refresh.

[0078] By setting sample user characteristics including at least one of user profile characteristics, user scenario characteristics, information content consumption statistics characteristics, satisfied consumption information content statistics characteristics, and first-click information content statistics characteristics, the beneficial effects are: Firstly, user profile features provide the user's basic attribute background; user scenario features capture the user's real-time status during the current browsing session; information content consumption statistics quantify the intensity and quality of the user's preference for different types of content; satisfied consumption information content statistics focus on the specific content that leads to deep satisfied consumption; and first-click information content statistics capture the core content that can stimulate the user's intention to open the application. These five categories of features together construct a multi-layered, multi-dimensional panoramic view of user characteristics, providing rich decision-making basis for the model.

[0079] Secondly, by integrating the five categories of features mentioned above, the model can comprehensively evaluate from multiple dimensions whether pushing target information content in the current information stream can encourage users to become active users. This comprehensive feature input enables the model to learn complex user behavior patterns, significantly improving the prediction accuracy of the target information content push effect.

[0080] Optionally, the statistical characteristics of information content consumption include at least one of the following: 1) Percentage of time spent by sample users on different types of information content in information flow applications.

[0081] Here, "content types" refers to content categories categorized according to different dimensions within an information flow application. These categories can be based on content format, content domain, content source, or other business dimensions. By statistically analyzing content by type, it is possible to precisely depict users' differentiated preferences for different categories of content.

[0082] The consumption time percentage refers to the proportion of a sample user's cumulative consumption time for a specific type of information content to their total cumulative consumption time for all types of information content during the same period. This indicator reflects the relative strength of users' preference for this type of content.

[0083] 2) Effective click-through rate of sample users for various types of information content in the information flow application.

[0084] The effective click-through rate (CTR) is determined based on the number of effective clicks for each type of information content and the total number of push notifications. For example, it can be determined by the ratio between the number of effective clicks and the total number of push notifications. The total number of push notifications refers to the total number of times a sample user is exposed to a certain type of information content within a specific time window; that is, the number of times that type of information content is pushed to the user's information stream and seen by the user.

[0085] Valid clicks refer to the number of clicks made by sample users within a specific time window for a certain type of information content that meet the valid criteria. Whether a click is valid depends on the duration of information content consumption after the click. If the user consumes the information content for more than a preset valid duration threshold after clicking, it is counted as a valid click; if the user exits immediately after clicking or the consumption time is too short, it is not included in the valid click count.

[0086] 3) The average consumption time of sample users for each type of information content in the information flow application.

[0087] The average consumption time refers to the average duration of a single consumption of a certain type of information content by sample users within a specific time window. This metric is obtained by dividing the cumulative consumption time of this type of information content by the number of consumptions, reflecting the average consumption depth of users for this type of information content. For example, the average consumption time for short videos is 45 seconds, and the average consumption time for long videos is 280 seconds.

[0088] 4) Completion rate of sample users for each type of information content in the information flow application.

[0089] Completion rate refers to the proportion of times a sample user completes a full consumption of a certain type of information content within a specific time window, out of the total number of consumptions. For video content, completion rate means the user watches until the end of the video; for text and image content, completion rate can be defined as the user scrolling to the bottom of the content or reaching a preset threshold for complete reading. Completion rate reflects the degree to which users fully accept the information content.

[0090] By setting statistical characteristics of information content consumption, including at least one of the following: the percentage of time sample users spend consuming each type of information content in the information feed application; the effective click-through rate of sample users for each type of information content in the information feed application; wherein the effective click-through rate is determined based on the number of effective clicks for each type of information content and the total number of pushes, and the number of effective clicks is determined based on the consumption time of the information content after the click; the average consumption time of sample users for each type of information content in the information feed application; and the completion rate of sample users for each type of information content in the information feed application, the beneficial effects are: Firstly, by statistically analyzing the percentage of consumption time for each type of information content, we can accurately depict the relative preference intensity of sample users for different types of information content.

[0091] Secondly, traditional click-through rate (CTR) statistics only count the click behavior itself and cannot distinguish between accidental clicks, clickbait clicks, and genuinely intentional clicks. This embodiment introduces "the duration of information content consumption after the click" as a criterion for determining effective CTR; only clicks that result in sufficient consumption time are counted as valid clicks. This design effectively filters out invalid click noise, making the click signals learned by the model more representative of users' true interests.

[0092] Thirdly, this embodiment comprehensively measures the quality of user consumption from multiple perspectives: the percentage of consumption time reflects the intensity of preference, the effective click-through rate reflects the authenticity of the click intention, the average consumption time reflects the depth of a single consumption, and the completion rate reflects the degree of complete acceptance of the content. The integration of multi-dimensional indicators enables the model to fully understand the user's consumption behavior patterns, thereby more accurately assessing the user's potential to accept the target information content.

[0093] Figure 3 This is a flowchart illustrating another training method for a push user screening model provided according to an embodiment of this disclosure. It can be used to further optimize and extend the above technical solution and can be combined with the various optional implementation methods described above. (Reference) Figure 3 The method specifically includes the following: S301. Obtain a sample information stream containing the target information content, and push the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user swipes the sample information stream.

[0094] S302. Based on the interaction data of sample users in the sample information stream, determine whether the target information content prompts the sample users to become active users, and generate standard push tags corresponding to the target information content in the sample information stream of the sample users based on the determination results.

[0095] S303. The first feature vector corresponding to the user profile features, the second feature vector corresponding to the user scenario features, the third feature vector corresponding to the information content consumption statistics features, the fourth feature vector corresponding to the satisfied consumption information content statistics features, and the fifth feature vector corresponding to the first click information content statistics features are input into the model to be trained through the model input layer of the model to be trained.

[0096] S304. The first feature vector, second feature vector, third feature vector, fourth feature vector and fifth feature vector are concatenated through the feature concatenation layer of the model to be trained to obtain the concatenated feature vector.

[0097] S305. The concatenated feature vector is extracted by the feature extraction layer of the model to be trained to obtain the feature extraction vector.

[0098] S306. The predicted push probability is obtained from the model output layer of the model to be trained based on the feature extraction vector.

[0099] S307. Based on the predicted push probability and standard push tags, train the model to be trained to obtain the push user selection model corresponding to the target information content.

[0100] In S303, the model input layer refers to the hierarchical structure in the model to be trained that is responsible for receiving the original feature data input.

[0101] The first feature vector refers to the numerical vector representation generated after the user profile features have been vectorized. This vector converts the user's basic attribute information into a dense vector of fixed dimensions, serving as the first type of feature input received by the model's input layer, and is used to represent the user's static background information.

[0102] The second feature vector refers to the numerical vector representation generated after vectorization of user scene features. This vector converts the real-time context information of the user during the current information stream refresh into a dense vector of fixed dimensions, which serves as the second type of feature input received by the model input layer to characterize the user's dynamic scene information.

[0103] The third feature vector refers to the numerical vector representation generated after vectorization of the statistical features of information content consumption. This vector converts statistical indicators such as the proportion of time users spend consuming various types of information content, effective click-through rate, average consumption time, and completion rate into a dense vector of fixed dimensions. This vector serves as the third type of feature input received by the model's input layer, representing the user's content consumption preferences and consumption quality.

[0104] The fourth feature vector refers to the numerical vector representation generated after vectorization of the statistical features of the satisfactory consumption information content. This vector transforms the set of information content that generates user satisfactory consumption behavior into a dense vector of fixed dimensions through a pre-trained embedding model, serving as the fourth type of feature input received by the model's input layer, and is used to characterize the user's deep interest direction.

[0105] The fifth feature vector refers to the numerical vector representation generated after vectorization of the statistical features of the information content clicked during the first click. This vector transforms the set of information content clicked by the user during the first click in the historical information stream into a dense vector of fixed dimensions through a pre-trained embedding model. It serves as the fifth type of feature input received by the model's input layer, representing the common core resources that can stimulate the user's intention to open the application.

[0106] Embedding models include, but are not limited to, embedding models.

[0107] In S304, the feature concatenation layer refers to the hierarchical structure in the model to be trained that is responsible for merging multiple feature vectors.

[0108] Feature concatenation refers to the process of merging multiple independent feature vectors along a specific dimension. This operation connects feature vectors from different sources and with different dimensions, combining them into a new vector with a higher dimension.

[0109] A concatenated feature vector is a combined vector containing all the original feature information generated after processing by a feature concatenation layer. This vector is formed by connecting the first, second, third, fourth, and fifth feature vectors in a preset order, and is a complete vectorized representation of user profile features, user scenario features, information content consumption statistics, satisfied consumption information content statistics, and first-click information content statistics.

[0110] In S305, the feature extraction layer refers to the deep neural network hierarchy located after the feature concatenation layer and before the model output layer in the model to be trained. This layer is specifically implemented as a multilayer perceptron, typically containing 2 to 4 hidden layers, with the number of neurons decreasing progressively in each layer. For example, the first hidden layer has 256 neurons, the second hidden layer has 128 neurons, and the third hidden layer has 64 neurons. Each hidden layer can use the ReLU activation function for non-linear transformation, and during network training, a Batch Normalization layer is used to normalize the input of each layer to accelerate convergence. A Dropout layer is also introduced to randomly discard some neurons to prevent overfitting.

[0111] Feature extraction refers to the computational process in which the feature extraction layer performs a layer-by-layer nonlinear transformation on the concatenated feature vector. This process maps the input data through the activation functions of multiple layers of neurons, with each layer taking the output of the previous layer as its input, gradually transforming the original concatenated feature vector into a higher-level, more abstract feature representation.

[0112] The feature extraction vector refers to the final feature representation output after the concatenated feature vectors are transformed layer by layer in the feature extraction layer of a neural network.

[0113] In S306, the model output layer refers to the last neural network layer in the model being trained, located after the feature extraction layer, and used to generate the final prediction result. This layer is specifically implemented as a fully connected layer with one neuron, receiving the feature extraction vector from the feature extraction layer as input and mapping it to a scalar value through a linear transformation. This scalar value then undergoes a non-linear transformation, including a sigmoid activation function, and is compressed into a probability range between 0 and 1, ultimately outputting the predicted push probability.

[0114] The model input layer takes the first feature vector corresponding to the user profile features, the second feature vector corresponding to the user scenario features, the third feature vector corresponding to the information content consumption statistics features, the fourth feature vector corresponding to the satisfied consumption information content statistics features, and the fifth feature vector corresponding to the first click information content statistics features as inputs into the model to be trained. This ensures that the model can fully utilize multi-dimensional information and provide a rich data foundation for subsequent predictions.

[0115] The first, second, third, fourth, and fifth feature vectors are concatenated by the feature concatenation layer to obtain a concatenated feature vector. This preserves the original information of each feature and forms a comprehensive vector that contains the whole picture of the user, enabling subsequent processing to learn from all features simultaneously.

[0116] By extracting features from the concatenated feature vectors through a feature extraction layer, a feature extraction vector is obtained, thereby automatically mining the deep correlation between features and extracting the discriminative features most relevant to the active user conversion target, thus improving the model's predictive ability.

[0117] The predicted push probability is obtained by extracting the feature vector through the model output layer. This intuitively represents the confidence that pushing the target information content once in the current information stream can encourage the user to become an active user, which is convenient for comparing with the standard push label to calculate the loss.

[0118] Figure 4 This is a flowchart illustrating an information flow push method according to an embodiment of this disclosure. This embodiment is applicable to situations where intelligent first-screen penetration decisions are made to increase the number of active users when new information content is introduced into an information flow recommendation platform. The method can be executed by an information flow push device, which can be implemented in hardware and / or software and can be configured in an electronic device. (Reference) Figure 4 The method specifically includes the following: S401. Obtain the target user characteristics corresponding to the current information flow refresh.

[0119] S402. Input the target user features into the push user screening model to obtain the predicted push probability corresponding to the target information content in the target user's current information stream refresh.

[0120] S403. If the predicted push probability is greater than or equal to the probability threshold, then obtain the target information stream containing the target information content and push the target information stream to the first screen of the information stream application corresponding to the current information stream refresh time of the target user.

[0121] In S401, the target user refers to the platform user who needs to make push decisions during the model application phase. Corresponding to the sample users used for data collection during the training phase, the target user is the user who actually receives the information stream service.

[0122] The current feed refresh rate refers to the single exposure opportunity corresponding to a target user initiating this content request in the feed application. This refresh rate is usually triggered by user actions such as launching the application, performing a refresh operation, or pulling down to load, and is the basic processing unit for the model to make real-time push decisions.

[0123] Target user features refer to user characteristic data collected or calculated by the system in real time when the target user's current information feed refresh occurs. These features maintain consistency in type and dimension with the sample user features from the training phase, specifically including at least one of the following: target user profile features, user scenario features, information content consumption statistics, satisfied consumption information content statistics, and first-click information content statistics. Target user features serve as real-time input to the push user selection model, used to generate the predicted push probability for the target information content in the current information feed refresh.

[0124] In one implementation of S401, when the current information stream refresh of the target user occurs, the real-time scene features of that refresh are first collected, including information such as the current time, network environment, and geographical location. Simultaneously, based on the target user's identifier, user profile features, information content consumption statistics, satisfied consumption information content statistics, and first-click information content statistics are queried from an offline feature database. The real-time collected scene features are integrated with the offline queried features to form a complete target user profile for the current refresh.

[0125] In step S402, the push user screening model is trained using any of the push user screening model training methods provided in the embodiments of this disclosure.

[0126] In one implementation of S402, after obtaining the complete target user features corresponding to the current information flow refresh, the target user features are input into a pre-trained push user screening model. The model first receives feature vectors corresponding to user profile features, user scenario features, information content consumption statistics, satisfied consumption information content statistics, and first-click information content statistics through the model input layer. These feature vectors are then concatenated through a feature concatenation layer to form a concatenated feature vector containing comprehensive user information. Subsequently, the concatenated feature vector undergoes a layer-by-layer nonlinear transformation by a multilayer perceptron in the feature extraction layer, gradually abstracting it into a feature extraction vector related to the active user conversion target. Finally, this feature extraction vector is passed to the model output layer, mapped by an activation function, and outputs a predicted push probability value between 0 and 1. This probability value represents the estimated likelihood that pushing target information content to the target user in the current information flow refresh will lead to the target user becoming an active user.

[0127] In S403, the probability threshold refers to a preset critical value used to decide whether to push target information content to the target user. This threshold is a value between 0 and 1, and the push behavior is determined by comparing it with the predicted push probability output by the model. The probability threshold can be dynamically adjusted according to the platform's operation strategy. For example, increasing the threshold can reduce the number of pushes to ensure user experience, while decreasing the threshold can increase the number of pushes to accelerate the penetration of target information content.

[0128] The target information stream refers to the information stream containing target information pushed to the target user during the model application phase when the decision result is a push notification. This information stream can be generated in the same way as during the training phase, such as inserting target information content into the regular information stream, or generating it by uniformly sorting the target information content with other regular content. The target information stream is the sequence of content that the target user ultimately sees on the first screen of the information stream application.

[0129] The first screen of the news feed application corresponding to the current news feed refresh by the target user refers to the first screen area that the target user can see directly without scrolling after the news feed application is launched or refreshed in the current news feed refresh.

[0130] In one implementation of S403, when the predicted push probability output by the push user filtering model is greater than or equal to a preset dynamic threshold, it is determined to penetrate the target information content onto the target user's feed application homepage during the current feed refresh. Specifically, firstly, the regular feed originally to be pushed for the current feed refresh is obtained. This feed is generated according to the original recommendation logic and contains regular information content arranged according to a regular sorting strategy. Subsequently, according to preset penetration rules, a target information content is inserted into this regular feed, for example, inserted into a specified position on the feed application homepage (such as the first or third position), generating a target feed containing the target information content. Finally, this target feed is pushed to the target user's feed application homepage corresponding to the current feed refresh for display.

[0131] This embodiment of the disclosure obtains the target user characteristics corresponding to the current information flow refresh time of the target user; inputs the target user characteristics into the push user filtering model to obtain the predicted push probability corresponding to the target information content in the current information flow refresh time of the target user; if the predicted push probability is greater than or equal to the probability threshold, then the target information flow containing the target information content is obtained, and the target information flow is pushed to the first screen of the information flow application corresponding to the current information flow refresh time of the target user. The beneficial effects are: Firstly, by comparing the predicted push probability with a preset threshold, personalized penetration decisions can be made based on the specific state of each target user in each information flow refresh, achieving refined push that is "tailored to the time and individual" and avoiding the "one-size-fits-all" extensive penetration of traditional solutions.

[0132] Secondly, it effectively avoids the negative impact on the user experience caused by blindly pushing target information content when the target user is in a bad state or the scenario is inappropriate.

[0133] Thirdly, by deploying the push user screening model in the online inference process, the predictive capabilities of the push user screening model can be directly applied to real-time decision-making, enabling the complex patterns learned by the push user screening model to be transformed into actual business value, thus maximizing the role of the trained model.

[0134] Optionally, after inputting the target user features into the push user filtering model to obtain the predicted push probability corresponding to the target information content in the target user's current information stream refresh, the model further includes: If the predicted push probability is less than the probability threshold, then retrieve the original information stream to be pushed to the home screen of the information stream application when the target user refreshes the current information stream; push the original information stream to the home screen of the information stream application corresponding to the target user's current information stream refresh.

[0135] The "original information stream" refers to the regular information stream pushed to the target user during the model application phase when the decision result is not to push the target information content. This information stream is generated according to the platform's original recommendation logic and consists only of regular information content arranged according to the conventional sorting strategy. The original information stream is the platform's natural recommendation result in a non-interventional state, conceptually consistent with the initial information stream obtained in the previous steps, and serves as the default push content when the predicted push probability does not reach the probability threshold in this step.

[0136] In one implementation, when the predicted push probability output by the user filtering model is less than a preset probability threshold, it is determined that the target information content will not be pushed to the target user during the current information feed refresh. At this time, the regular recommendation engine retrieves the original information feed to be pushed for the current information feed refresh. Subsequently, the original information feed is pushed to the target user's home screen of the information feed application corresponding to the current information feed refresh, so that the target user sees the platform's natural recommended content without intervention after launching the application or refreshing.

[0137] If the predicted push probability is less than a probability threshold, then the original information feed to be pushed to the home screen of the information feed application at the current time the target user refreshes the information feed; pushing the original information feed to the home screen of the information feed application corresponding to the current time the target user refreshes the information feed has the following benefits: When the predicted push probability is less than the probability threshold, it indicates that the likelihood of pushing the target information to the target user after the current information feed refresh is low, and it may even negatively impact the user experience due to mismatched information content. By pushing the original information feed, the platform maintains its original recommendation experience and avoids forcibly pushing new content when the user is in a bad state or the scenario is inappropriate, thus ensuring the stability of the user experience.

[0138] Figure 5 This is a schematic diagram of the overall process of information flow push according to an embodiment of this disclosure, such as... Figure 5 As shown, the overall information feed push process includes five stages, from the first stage to the fifth stage.

[0139] The first phase includes: Step 1: Randomly selecting sample users; Step 2: Infiltrating the target information content on the first screen; and Step 3: Collecting interactive behavior data of the sample users.

[0140] The second stage includes: Step 1: Constructing sample user features; Step 2: Generating standard push tags.

[0141] The third stage includes: Step 1: Input sample user feature vectors into the model to be trained; Step 2: Output predicted push probability into the model to be trained.

[0142] The fourth stage includes: Step 1: Target users initiate information flow refreshes; Step 2: Real-time acquisition of target user characteristics; Step 3: Inputting the push user screening model to obtain the predicted push probability; Step 4: Comparing the predicted push probability with the probability threshold; Step 5: Deciding whether to penetrate the target information content.

[0143] The specific implementation methods of the above steps can be found in the description of the embodiments section of this disclosure, and will not be repeated here.

[0144] Figure 6 This is a schematic diagram of an information flow push decision provided according to an embodiment of this disclosure, such as... Figure 6 As shown, it is implemented in five parts: user interface, feature implementation service, model prediction service, decision engine, and recommendation system. The specific steps include: 1. The user initiates a request for the first screen of the information stream.

[0145] 2. Call the real-time feature service to generate complete target user features.

[0146] 3. Send the target user characteristics to the model prediction service.

[0147] 4. Call the model prediction service to output the predicted push probability based on the target user characteristics, and return the predicted push probability to the decision engine.

[0148] 5. Call the decision engine to compare the predicted push probability with the probability threshold. If the predicted push probability is greater than or equal to the probability threshold, then decide to push the target content information; if the predicted push probability is less than the probability threshold, then decide not to push the target content information; send the decision result to the recommendation system.

[0149] 6. Call the recommendation system to generate a target information stream based on the decision results, and return the target information stream to the user.

[0150] The specific implementation methods of the above steps can be found in the description of the embodiments section of this disclosure, and will not be repeated here.

[0151] Figure 7 This is a schematic diagram of a training device for a push user screening model provided according to an embodiment of the present disclosure. It can be applied to situations where intelligent first-screen penetration decisions are made to increase the number of active users when new information content is introduced into an information flow recommendation platform. The device of this embodiment can be implemented in software and / or hardware and can be integrated into any electronic device with computing capabilities.

[0152] like Figure 7 As shown, the training device 70 for the push user screening model disclosed in this embodiment may include a sample information stream acquisition module 71, a standard push tag generation module 72, and a model training module 73, wherein: The sample information stream acquisition module 71 is used to acquire a sample information stream containing target information content and push the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user swipes the sample information stream; The standard push tag generation module 72 is used to determine whether the target information content has made the sample user an active user based on the interaction behavior data of the sample user in the sample information stream, and generate a standard push tag corresponding to the target information content in the sample user's sample information stream based on the determination result. The model training module 73 is used to input the sample user features corresponding to the sample user when the sample information stream is refreshed into the model to be trained, to obtain the predicted push probability of the target information content corresponding to the sample user's refresh rate in the sample information stream, and to train the model to be trained according to the predicted push probability and the standard push label to obtain the push user screening model corresponding to the target information content.

[0153] Optionally, the sample information stream acquisition module 71 is specifically used for: When the sample information stream of the sample user is refreshed, the initial information stream to be pushed to the first screen of the information stream is applied to the information stream application; The target information content is inserted into the initial information stream to obtain the sample information stream.

[0154] Optional, the standard push label generation module 72 is specifically used for: When the interaction behavior data includes first behavior data and second behavior data, it is determined that the target information content facilitates the sample user to become an active user. The first behavioral data refers to the information click or information refresh behavior generated by the sample user on the first screen of the information flow application. The second behavioral data refers to the satisfactory consumption behavior of the sample users towards any information content in the information flow application within a preset time period after the occurrence of the first behavioral data. The satisfactory consumption behaviors include liking, commenting, sharing, and prolonged viewing. The prolonged viewing behavior indicates that the sample users' viewing time for the information content exceeds a preset multiple of the average viewing time of the information content.

[0155] Optionally, the standard push label generation module 72 is also used for: If it is determined that the target information content leads the sample user to become an active user, then a standard push label with a label content of one is generated; If it is determined that the target information content does not lead to the sample user becoming an active user, then a standard push label with a label content of zero is generated.

[0156] Optionally, the sample user characteristics include at least one of user profile characteristics, user scenario characteristics, information content consumption statistics, satisfied consumption information content statistics, and first-click information content statistics. The statistical features of the satisfactory consumption information content are the set of information content in which the sample users generate satisfactory consumption behaviors in the information flow application. The satisfactory consumption behaviors include liking behavior, commenting behavior, sharing behavior, and extended viewing behavior. The extended viewing behavior means that the viewing time of the sample user for any information content exceeds a preset multiple of the average viewing time of that information content. The statistical features of the first click information content are the set of information content that the sample user clicked for the first time in the historical information stream refresh.

[0157] Optionally, the information content consumption statistical characteristics include at least one of the following: The percentage of time the sample users spent consuming each type of information content in the information flow application. The effective click-through rate of the sample users for each type of information content in the information flow application; wherein, the effective click-through rate is determined based on the number of effective clicks for each type of information content and the total number of pushes, and the number of effective clicks is determined based on the consumption time of the information content after the click; The average consumption time of the sample users for each type of information content in the information flow application; The sample users' completion rate for each type of information content in the information flow application.

[0158] Optionally, the model to be trained includes a model input layer, a feature concatenation layer, a feature extraction layer, and a model output layer; Model training module 73 is specifically used for: The first feature vector corresponding to the user profile features, the second feature vector corresponding to the user scenario features, the third feature vector corresponding to the information content consumption statistics features, the fourth feature vector corresponding to the satisfied consumption information content statistics features, and the fifth feature vector corresponding to the first click information content statistics features are input into the model to be trained through the model input layer. The first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector are concatenated through the feature concatenation layer to obtain a concatenated feature vector; The concatenated feature vector is extracted by the feature extraction layer to obtain the feature extraction vector; The predicted push probability is obtained through the model output layer based on the feature extraction vector.

[0159] The training apparatus 70 for the push user screening model disclosed in this embodiment can execute the training method for the push user screening model disclosed in this embodiment, and has the corresponding functional modules and beneficial effects of the method execution. Content not described in detail in this embodiment can be referred to the description in the method embodiments of this disclosure.

[0160] Figure 8 This is a schematic diagram of an information flow push device according to an embodiment of the present disclosure. It can be applied to situations where intelligent first-screen penetration decisions are made to increase the number of active users when new information content is introduced into an information flow recommendation platform. The device of this embodiment can be implemented in software and / or hardware and can be integrated into any electronic device with computing capabilities.

[0161] like Figure 8 As shown, the information stream push device 80 disclosed in this embodiment may include a target user feature acquisition module 81, a predicted push probability determination module 82, and a target information stream push module 83, wherein: The target user feature acquisition module 81 is used to acquire the target user features corresponding to the current information stream refresh. The prediction push probability determination module 82 is used to input the target user features into the push user screening model to obtain the predicted push probability of the target information content corresponding to the current information stream refresh of the target user; wherein, the push user screening model is trained using the training method of the push user screening model as described in any of the embodiments of this disclosure. The target information stream push module 83 is used to obtain a target information stream containing the target information content if the predicted push probability is greater than or equal to a probability threshold, and push the target information stream to the target user's first screen of the information stream application corresponding to the current information stream refresh.

[0162] Optionally, the device further includes the original push logic processing module, specifically used for: If the predicted push probability is less than the probability threshold, then when the target user refreshes the current information stream, the original information stream to be pushed to the first screen of the information stream application is obtained; The original information stream is pushed to the target user on the home screen of the information stream application corresponding to the current information stream refresh.

[0163] The information stream push device 80 disclosed in this embodiment can execute the information stream push method disclosed in this embodiment, and has the corresponding functional modules and beneficial effects of executing the method. Content not described in detail in this embodiment can be referred to the description in the method embodiments of this disclosure.

[0164] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0165] Figure 9 A schematic block diagram of an example electronic device 900 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.

[0166] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 902 or a computer program loaded from storage unit 908 into random access memory (RAM) 903. RAM 903 may also store various programs and data required for the operation of device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0167] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0168] The computing unit 901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 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 901 performs the various methods and processes described above, such as the push user screening model training method and / or information stream push method. For example, in some embodiments, the push user screening model training method and / or information stream push method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the push user screening model training method and / or information stream push method described above can be performed. Alternatively, in other embodiments, the computing unit 901 may be configured by any other suitable means (e.g., by means of firmware) to perform a training method for a push user screening model and / or a news feed push method.

[0169] Various implementations 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), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations 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.

[0170] 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.

[0171] 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.

[0172] 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).

[0173] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users 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., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0174] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service ecosystem that addresses the management difficulties and weak business scalability inherent in traditional physical hosting and VPS services. Servers can also be servers for distributed systems or servers integrated with blockchain technology.

[0175] Artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies mainly include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0176] Cloud computing refers to a technology system that enables access to a shared pool of physical or virtual resources via a network. These resources can include servers, operating systems, networks, software, applications, and storage devices, and can be deployed and managed on demand and in a self-service manner. Cloud computing technology can provide efficient and powerful data processing capabilities for applications such as artificial intelligence and blockchain, as well as for model training.

[0177] It should be understood that the various forms of processes shown above can be used to reorder, 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.

[0178] 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. A method for training a user filtering model, comprising: Obtain a sample information stream containing target information content, and push the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user swipes the sample information stream; Based on the interaction behavior data of the sample users in the sample information stream, determine whether the target information content has led the sample users to become active users, and generate standard push tags corresponding to the target information content in the sample users' sample information stream based on the determination result. The sample user features corresponding to the sample user's refresh rate in the sample information stream are input into the model to be trained to obtain the predicted push probability of the target information content corresponding to the sample user's refresh rate in the sample information stream. Based on the predicted push probability and the standard push tag, the model to be trained is trained to obtain the push user screening model corresponding to the target information content.

2. The method according to claim 1, wherein, The acquisition of the sample information stream containing the target information content includes: When the sample information stream of the sample user is refreshed, the initial information stream to be pushed to the first screen of the information stream is applied to the information stream application; The target information content is inserted into the initial information stream to obtain the sample information stream.

3. The method according to claim 1, wherein, The step of determining whether the target information content contributes to making the sample user an active user based on the sample user's interaction behavior data during the sample information stream refresh includes: When the interaction behavior data includes first behavior data and second behavior data, it is determined that the target information content facilitates the sample user to become an active user. The first behavioral data refers to the information click or information refresh behavior generated by the sample user on the first screen of the information flow application. The second behavioral data refers to the satisfactory consumption behavior of the sample users towards any information content in the information flow application within a preset time period after the occurrence of the first behavioral data. The satisfactory consumption behaviors include liking, commenting, sharing, and prolonged viewing. The prolonged viewing behavior indicates that the sample users' viewing time for the information content exceeds a preset multiple of the average viewing time of the information content.

4. The method according to claim 1, wherein, The step of generating a standard push tag corresponding to the target information content in the sample information stream of the sample user based on the determined result includes: If it is determined that the target information content leads the sample user to become an active user, then a standard push label with a label content of one is generated; If it is determined that the target information content does not lead to the sample user becoming an active user, then a standard push label with a label content of zero is generated.

5. The method according to any one of claims 1-4, wherein, The sample user characteristics include at least one of user profile characteristics, user scenario characteristics, information content consumption statistics characteristics, satisfied consumption information content statistics characteristics, and first-click information content statistics characteristics; The statistical features of the satisfactory consumption information content are the set of information content in which the sample users generate satisfactory consumption behaviors in the information flow application. The satisfactory consumption behaviors include liking behavior, commenting behavior, sharing behavior, and extended viewing behavior. The extended viewing behavior means that the viewing time of the sample user for any information content exceeds a preset multiple of the average viewing time of that information content. The statistical features of the first click information content are the set of information content that the sample user clicked for the first time in the historical information stream refresh.

6. The method according to claim 5, wherein, The statistical characteristics of information content consumption include at least one of the following: The percentage of time the sample users spent consuming each type of information content in the information flow application. The effective click-through rate of the sample users for each type of information content in the information flow application; wherein, the effective click-through rate is determined based on the number of effective clicks for each type of information content and the total number of pushes, and the number of effective clicks is determined based on the consumption time of the information content after the click; The average consumption time of the sample users for each type of information content in the information flow application; The sample users' completion rate for each type of information content in the information flow application.

7. The method according to claim 5, wherein, The model to be trained includes a model input layer, a feature concatenation layer, a feature extraction layer, and a model output layer; The step of inputting the sample user features corresponding to the sample user's refresh rate in the sample information stream into the model to be trained, and obtaining the predicted push probability of the target information content corresponding to the sample user's refresh rate in the sample information stream, includes: The first feature vector corresponding to the user profile features, the second feature vector corresponding to the user scenario features, the third feature vector corresponding to the information content consumption statistics features, the fourth feature vector corresponding to the satisfied consumption information content statistics features, and the fifth feature vector corresponding to the first click information content statistics features are input into the model to be trained through the model input layer. The first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector are concatenated through the feature concatenation layer to obtain a concatenated feature vector; The concatenated feature vector is extracted by the feature extraction layer to obtain the feature extraction vector; The predicted push probability is obtained through the model output layer based on the feature extraction vector.

8. A method for pushing information streams, comprising: Obtain the target user characteristics corresponding to the current information flow refresh time; The target user features are input into the push user filtering model to obtain the predicted push probability of the target information content corresponding to the current information stream refresh of the target user; wherein, the push user filtering model is trained using the method described in any one of claims 1-7; If the predicted push probability is greater than or equal to the probability threshold, then the target information stream containing the target information content is obtained, and the target information stream is pushed to the target user's first screen of the information stream application corresponding to the current information stream refresh.

9. The method according to claim 8, further comprising, after inputting the target user features into the push user filtering model to obtain the predicted push probability of the target information content corresponding to the current information stream refresh of the target user: If the predicted push probability is less than the probability threshold, then when the target user refreshes the current information stream, the original information stream to be pushed to the first screen of the information stream application is obtained; The original information stream is pushed to the target user on the home screen of the information stream application corresponding to the current information stream refresh.

10. A training device for a push user screening model, comprising: The sample information stream acquisition module is used to acquire a sample information stream containing target information content and push the sample information stream to the first screen of the information stream application corresponding to the number of times the sample user swipes the sample information stream; The standard push tag generation module is used to determine whether the target information content has led the sample user to become an active user based on the interaction behavior data of the sample user in the sample information stream, and generate a standard push tag corresponding to the target information content in the sample user's sample information stream based on the determination result. The model training module is used to input the sample user features corresponding to the sample user when the sample information stream is refreshed into the model to be trained, to obtain the predicted push probability of the target information content corresponding to the sample user's refresh rate in the sample information stream, and to train the model to be trained based on the predicted push probability and the standard push tag to obtain the push user screening model corresponding to the target information content.

11. An information stream push device, comprising: The target user feature acquisition module is used to acquire the target user features corresponding to the current information stream refresh. The predictive push probability determination module is used to input the target user features into the push user filtering model to obtain the predicted push probability of the target information content corresponding to the current information stream refresh of the target user; wherein, the push user filtering model is trained using the method described in any one of claims 1-7; The target information stream push module is used to obtain a target information stream containing the target information content if the predicted push probability is greater than or equal to a probability threshold, and push the target information stream to the target user's first screen of the information stream application corresponding to the current information stream refresh.

12. 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 1-9.

13. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-9.

14. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the method according to any one of claims 1-9.

Images