Content real-time pushing optimization method based on short-term user behavior prediction

Abstract

The invention relates to a content real-time pushing optimization method based on short-term user behavior prediction. The method comprises the following steps: S1, log data index standard and normative; s2, constructing a log data acquisition system; s3, behavior prediction model training: providing personalized support for recommending short-term user behaviors through a behavior prediction model; s4, short-term behavior feature extraction: quickly capturing short-term user behaviors so as to enhance the timeliness of short-term user behavior recommendation; s5, a content optimization mechanism: screening and sorting recommended contents by using historical and real-time behavior data; s6, real-time content pushing: executing pushing and dynamic adjustment on the candidate content in the candidate pool; and S7, updating a feedback mechanism and a model: according to the real-time feedback of the short-term user behavior, performing incremental training on the student model, and gradually optimizing the behavior prediction model by adopting the student model, thereby improving the prediction capability of the behavior prediction model on the short-term user behavior.

Description

Technical Field

[0001] The present invention relates to the field of software information system design, and in particular to a real-time content push optimization method based on short-term user behavior prediction. Background Art

[0002] With the development of the Internet and the popularization of information flow push technology, real-time push systems have become the key to providing personalized content, pushing content that users may be interested in, thereby improving platform user stickiness and conversion rates.

[0003] However, traditional push methods mainly rely on long-term user behavior or simple real-time signals. This method ignores the dynamic changes of users' short-term interests and is difficult to accurately capture changes in users' short-term interests, resulting in reduced push content relevance and user satisfaction.

[0004] In addition, short-term behaviors are time-sensitive and dynamic. Existing methods lack the ability to model short-term behaviors in real time. They usually adopt batch processing and cannot quickly respond to users' immediate needs, resulting in a lack of targeted push content.

[0005] Therefore, how to predict users' interests based on their short-term behavior and optimize real-time content push has become an important direction of technological development.

[0006] However, the existing mainstream technology is a recommendation method based on long-term user behavior. Long-term behavior data can reflect the general interests of users, but it is difficult to capture changes in user interests in a short period of time. The update cycle of long-term behavior data is long, resulting in a relatively slow response speed of the recommendation system, which is difficult to meet real-time requirements.

[0007] Currently, there is no relevant solution on the market. Summary of the invention

[0008] In view of the above technical problems in the related art, the present invention proposes a real-time content push optimization method based on short-term user behavior prediction, which can overcome the above-mentioned shortcomings of the prior art.

[0009] To achieve the above technical objectives, the technical solution of the present invention is implemented as follows: A method for optimizing content real-time push based on short-term user behavior prediction, comprising the following steps: S1 log data indicator standards and specifications: S1.1 Index design: Develop an extensible log data index system, which includes several indicators covering multi-dimensional data analysis needs. The log data index system dynamically expands the included indicators according to specific business scenarios; S1.2 Data collection specifications: formulate data collection specifications, which include technical specifications and log format standardization; S1.2.1 Technical Specifications: Execute event tracking collection and non-tracking collection respectively; event tracking collection refers to the unified collection of user operation behavior data at the front-end tracking through the self-developed unified log framework, so as to ensure that the log upload complies with the established specifications; tracking collection refers to the automatic capture of user behavior data by the back-end through the proxy layer, so as to ensure the flexible and efficient collection process; S1.2.2 Log format standardization: Use a unified data format to record user behavior data. User behavior data follows the principle of consistency in field naming and structure design to facilitate subsequent processing and analysis; S2 builds a log data collection system: Through the log collection system, multi-terminal behavior data convergence and unified management are achieved; S3 behavior prediction model training: Provide personalized support for recommending short-term user behaviors through behavior prediction models, including the following steps: S3.1 Model input design: Use the characteristics of short-term user behavior as the main input, including historical behavior characteristics and real-time behavior characteristics in short-term user behavior; S3.2 Data standardization: Perform normalization of numerical features and embedding of categorical features on short-term user behaviors; normalization of numerical features is used to eliminate the impact of dimensions; embedding of categorical features is used to improve feature representation capabilities; S3.3 Model training: using a supervised learning method to train the behavior prediction model based on the cleaned historical behavior data; S3.4 Model miniaturization: By using the knowledge distillation technology to transfer knowledge from the teacher model (large model) to the student model (small model), a corresponding teacher model (large model) and a student model (small model) are constructed for the behavior prediction model, and the behavior prediction model is miniaturized to ensure that the miniaturized behavior prediction model significantly improves the operating efficiency while maintaining high accuracy; S4 Short-term behavior feature extraction: Quickly capture short-term user behavior to enhance the timeliness of short-term user behavior recommendations, including the following steps: S4.1 Data extraction: Obtain the user's behavior data for the last 10 minutes in real time through the query interface - that is, short-term user behavior; S4.2 Feature standardization: normalize and embed the short-term user behavior data extracted in S4.1 to ensure data consistency; S4.3 Weight enhancement: assigning a higher weight to the short-term features of the short-term user behavior, thereby highlighting the impact of the real-time nature of the behavior on the prediction results; S5 content selection mechanism: Use historical and real-time behavior data to filter and sort recommended content, including the following steps: S5.1 Candidate content generation: Combining user historical data and content features, a collaborative filtering algorithm is used to generate a candidate pool including several candidate contents; S5.2 Content sorting: Based on the student model (small model) and short-term behavior characteristics of the behavior prediction model, score and sort each candidate content in the candidate pool; S5.3 Content screening: removing candidate contents with lower rankings in the candidate pool, thereby ensuring the relevance and accuracy of the recommendation; S6 real-time content push: push and dynamically adjust the candidate content in the candidate pool; S7 Feedback mechanism and model update: Based on the real-time feedback of the short-term user behavior, the student model (small model) is incrementally trained, and the behavior prediction model is gradually optimized using the student model (small model), thereby improving the prediction ability of the behavior prediction model for short-term user behavior.

[0010] Preferably, in S1.1, The log data indicator system includes basic interaction indicators, behavior path indicators, conversion indicators, etc. The basic interaction indicators include click count, dwell time, sliding depth, etc. The behavior path indicators include jump path, bounce rate, etc.; the conversion indicators include recommendation data click rate, etc.

[0011] Preferably, in S1.2, the unified data format is preferably in JSON format.

[0012] Preferably, in S2, The log collection system includes a data source integration module and a data processing module; The data source integration module supports collecting user behavior data from the front-end embedding framework and the back-end proxy layer to ensure full data coverage; The data processing module includes an ETL module, a log storage module, and a query service module; ETL module: supports data extraction, conversion and loading, and cleans and standardizes raw data; Log storage module: adopts distributed storage architecture to ensure efficient storage and access; Query service module: Through a unified query interface, it provides fast query support for historical behavior data, providing a data foundation for subsequent analysis and model training.

[0013] Preferably, the S4 comprises the following steps: S4.1 Data extraction: define a short-term period T, and obtain the user's recent behavior data of the short-term period T in real time through the query interface - that is, short-term user behavior; S4.2 Feature standardization: normalize and embed the short-term user behavior data extracted in S4.1 to ensure data consistency; S4.3 Weight enhancement: assigning higher weights to the short-term features of the short-term user behavior, thereby highlighting the impact of the real-time nature of the behavior on the prediction results.

[0014] Preferably, in S4.1, the value range of the short-term period T depends on actual needs, and is preferably 5 minutes to 180 minutes.

[0015] Preferably, the S6 comprises the following steps: S6.1 Push method: Push the candidate content in real time through home page recommendation or pop-up window; S6.2 Push optimization: According to the characteristics of the short-term user behavior, dynamically adjust the push frequency and method to improve the user experience.

[0016] Preferably, the S7 comprises the following steps: S7.1 Feedback data collection: record the user interaction data after S6 push, and divide the real-time feedback in the user interaction data into positive feedback and negative feedback; the positive feedback includes click, like, and forward; the negative feedback includes close, disinterest, and report; S7.2 incremental model update: performing incremental training on the student model (small model) using the positive feedback and negative feedback; S7.3: The behavior prediction model is gradually optimized through the student model (small model) to improve the short-term user behavior prediction capability of the behavior prediction model and enhance the adaptability of the behavior prediction model.

[0017] Beneficial effects of the present disclosure: The present disclosure collects user behavior data in real time within a short period of time (for example, a few minutes or hours), introduces deep learning technology, builds a predictive student model (small model) for short-term user behavior, predicts user content of interest through the student model (small model), and adjusts the content push strategy in real time through a comprehensive scoring mechanism, thereby improving the accuracy and real-time nature of the push and meeting users' immediate needs.

[0018] This paper solves the shortcomings of existing content push technology in terms of real-time, accuracy and dynamism, and proposes a real-time content push optimization method based on short-term user behavior prediction. By quickly collecting, analyzing and predicting user behavior data in a short period of time, the push strategy is dynamically adjusted to achieve more efficient and accurate content recommendations.

[0019] Based on short-term user behavior, the present invention captures the dynamic changes of user interests in real time, pushes more relevant content to users, and improves user experience. Real-time optimized content push can extend user stay time and increase user activity, thereby bringing more advertising exposure and click opportunities. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 This is a logical flow chart of the method for optimizing real-time content push based on short-term user behavior prediction described in the present disclosure. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0023] like Figure 1 As shown, in order to facilitate understanding of the above technical solutions of the present invention, the above technical solutions of the present invention are described in detail below through specific usage methods.

[0024] A method for optimizing content real-time push based on short-term user behavior prediction, comprising the following steps: S1 log data indicator standards and specifications: S1.1 Index design: Develop an extensible log data index system, which includes several indicators covering multi-dimensional data analysis needs. The log data index system dynamically expands the included indicators according to specific business scenarios; S1.2 Data collection specifications: formulate data collection specifications, which include technical specifications and log format standardization; S1.2.1 Technical Specifications: Execute event tracking collection and non-tracking collection respectively; event tracking collection refers to the unified collection of user operation behavior data at the front-end tracking through the self-developed unified log framework, so as to ensure that the log upload complies with the established specifications; tracking collection refers to the automatic capture of user behavior data by the back-end through the proxy layer, so as to ensure the flexible and efficient collection process; S1.2.2 Log format standardization: Use a unified data format to record user behavior data. User behavior data follows the principle of consistency in field naming and structure design to facilitate subsequent processing and analysis; S2 builds a log data collection system: Through the log collection system, multi-terminal behavior data convergence and unified management are achieved; S3 behavior prediction model training: Provide personalized support for recommending short-term user behaviors through behavior prediction models, including the following steps: S3.1 Model input design: Use the characteristics of short-term user behavior as the main input, including historical behavior characteristics and real-time behavior characteristics in short-term user behavior; S3.2 Data standardization: Perform normalization of numerical features and embedding of categorical features on short-term user behaviors; normalization of numerical features is used to eliminate the impact of dimensions; embedding of categorical features is used to improve feature representation capabilities; S3.3 Model training: using a supervised learning method to train the behavior prediction model based on the cleaned historical behavior data; S3.4 Model miniaturization: By using the knowledge distillation technology to transfer knowledge from the teacher model (large model) to the student model (small model), a corresponding teacher model (large model) and a student model (small model) are constructed for the behavior prediction model, and the behavior prediction model is miniaturized to ensure that the miniaturized behavior prediction model significantly improves the operating efficiency while maintaining high accuracy; S4 Short-term behavior feature extraction: Quickly capture short-term user behavior to enhance the timeliness of short-term user behavior recommendations, including the following steps: S4.1 Data extraction: Obtain the user's behavior data for the last 10 minutes in real time through the query interface - that is, short-term user behavior; S4.2 Feature standardization: normalize and embed the short-term user behavior data extracted in S4.1 to ensure data consistency; S4.3 Weight enhancement: assigning a higher weight to the short-term features of the short-term user behavior, thereby highlighting the impact of the real-time nature of the behavior on the prediction results; S5 content selection mechanism: Use historical and real-time behavior data to filter and sort recommended content, including the following steps: S5.1 Candidate content generation: Combining user historical data and content features, a collaborative filtering algorithm is used to generate a candidate pool including several candidate contents; S5.2 Content sorting: Based on the student model (small model) and short-term behavior characteristics of the behavior prediction model, score and sort each candidate content in the candidate pool; S5.3 Content screening: removing candidate contents with lower rankings in the candidate pool, thereby ensuring the relevance and accuracy of the recommendation; S6 real-time content push: push and dynamically adjust the candidate content in the candidate pool; S7 Feedback mechanism and model update: Based on the real-time feedback of the short-term user behavior, the student model (small model) is incrementally trained, and the behavior prediction model is gradually optimized using the student model (small model), thereby improving the prediction ability of the behavior prediction model for short-term user behavior.

[0025] In one embodiment, in S1.1, The log data indicator system includes basic interaction indicators, behavior path indicators, conversion indicators, etc. The basic interaction indicators include click count, dwell time, sliding depth, etc. The behavior path indicators include jump path, bounce rate, etc.; the conversion indicators include recommendation data click rate, etc.

[0026] In a certain embodiment, in S1.2, the unified data format is preferably in JSON format.

[0027] In one embodiment, in S2, The log collection system includes a data source integration module and a data processing module; The data source integration module supports collecting user behavior data from the front-end embedding framework and the back-end proxy layer to ensure full data coverage; The data processing module includes an ETL module, a log storage module, and a query service module; ETL module: supports data extraction, conversion and loading, and cleans and standardizes raw data; Log storage module: adopts distributed storage architecture to ensure efficient storage and access; Query service module: Through a unified query interface, it provides fast query support for historical behavior data, providing a data foundation for subsequent analysis and model training.

[0028] In one embodiment, in S3, the behavior prediction model is preferably.

[0029] In one embodiment, the S4 comprises the following steps: S4.1 Data extraction: define a short-term period T, and obtain the user's recent behavior data of the short-term period T in real time through the query interface - that is, short-term user behavior; S4.2 Feature standardization: normalize and embed the short-term user behavior data extracted in S4.1 to ensure data consistency; S4.3 Weight enhancement: assigning higher weights to the short-term features of the short-term user behavior, thereby highlighting the impact of the real-time nature of the behavior on the prediction results.

[0030] In one embodiment, in S4.1, the value range of the short-term period T depends on actual needs, and is preferably 5 minutes to 180 minutes.

[0031] In one embodiment, the S6 comprises the following steps: S6.1 Push method: Push the candidate content in real time through home page recommendation or pop-up window; S6.2 Push optimization: According to the characteristics of the short-term user behavior, dynamically adjust the push frequency and method to improve the user experience.

[0032] In one embodiment, the S7 comprises the following steps: S7.1 Feedback data collection: record the user interaction data after S6 push, and divide the real-time feedback in the user interaction data into positive feedback and negative feedback; the positive feedback includes click, like, and forward; the negative feedback includes close, disinterest, and report; S7.2 incremental model update: performing incremental training on the student model (small model) using the positive feedback and negative feedback; S7.3: The behavior prediction model is gradually optimized through the student model (small model) to improve the short-term user behavior prediction capability of the behavior prediction model and enhance the adaptability of the behavior prediction model.

[0033] To sum up, through the above-mentioned unique technical solution, the present invention collects the user's behavior data in real time within a short period of time (for example, a few minutes or hours), introduces deep learning technology, builds a predictive student model (small model) for short-term user behavior, predicts the user's interest content through the student model (small model), and adjusts the content push strategy in real time through a comprehensive scoring mechanism, thereby improving the accuracy and real-time nature of the push and meeting the user's immediate needs.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for optimizing content push in real time based on short-term user behavior prediction, characterized in that: The following steps are involved: S1 log data indicator standards and specifications: S1.1 Index design: Develop an extensible log data index system, which includes several indicators covering multi-dimensional data analysis needs. The log data index system dynamically expands the included indicators according to specific business scenarios; S1.2 Data collection specifications: formulate data collection specifications, which include technical specifications and log format standardization; S1.2.1 Technical Specifications: Execute event tracking collection and non-tracking collection respectively; event tracking collection refers to the unified collection of user operation behavior data at the front-end tracking through the self-developed unified log framework, so as to ensure that the log upload complies with the established specifications; tracking collection refers to the automatic capture of user behavior data by the back-end through the proxy layer, so as to ensure the flexible and efficient collection process; S1.2.2 Log format standardization: Use a unified data format to record user behavior data. User behavior data follows the principle of consistency in field naming and structure design to facilitate subsequent processing and analysis; S2 builds a log data collection system: Through the log collection system, multi-terminal behavior data convergence and unified management are achieved; S3 behavior prediction model training: Provide personalized support for recommending short-term user behaviors through behavior prediction models, including the following steps: S3.1 Model input design: Use the characteristics of short-term user behavior as the main input, including historical behavior characteristics and real-time behavior characteristics in short-term user behavior; S3.2 Data standardization: Perform normalization of numerical features and embedding of categorical features on short-term user behaviors; normalization of numerical features is used to eliminate the impact of dimensions; embedding of categorical features is used to improve feature representation capabilities; S3.3 Model training: using a supervised learning method to train the behavior prediction model based on the cleaned historical behavior data; S3.4 Model miniaturization: By using the knowledge distillation technology to transfer knowledge from the teacher model (large model) to the student model (small model), a corresponding teacher model (large model) and a student model (small model) are constructed for the behavior prediction model, and the behavior prediction model is miniaturized to ensure that the miniaturized behavior prediction model significantly improves the operating efficiency while maintaining high accuracy; S4 Short-term behavior feature extraction: Quickly capture short-term user behavior to enhance the timeliness of short-term user behavior recommendations, including the following steps: S4.1 Data extraction: Obtain the user's behavior data for the last 10 minutes in real time through the query interface - that is, short-term user behavior; S4.2 Feature standardization: normalize and embed the short-term user behavior data extracted in S4.1 to ensure data consistency; S4.3 Weight enhancement: assigning a higher weight to the short-term features of the short-term user behavior, thereby highlighting the impact of the real-time nature of the behavior on the prediction results; S5 content selection mechanism: Use historical and real-time behavior data to filter and sort recommended content, including the following steps: S5.1 Candidate content generation: Combining user historical data and content features, a collaborative filtering algorithm is used to generate a candidate pool including several candidate contents; S5.2 Content sorting: Based on the student model (small model) and short-term behavior characteristics of the behavior prediction model, score and sort each candidate content in the candidate pool; S5.3 Content screening: removing candidate contents with lower rankings in the candidate pool, thereby ensuring the relevance and accuracy of the recommendation; S6 real-time content push: push and dynamically adjust the candidate content in the candidate pool; S7 Feedback mechanism and model update: Based on the real-time feedback of the short-term user behavior, the student model is incrementally trained, and the behavior prediction model is gradually optimized using the student model, thereby improving the prediction ability of the behavior prediction model for short-term user behavior.

2. The method for optimizing content push in real time according to claim 1, characterized in that: In S1.1, The log data indicator system includes basic interaction indicators, behavior path indicators, and conversion indicators; The basic interaction indicators include click counts, dwell time, and sliding depth; The behavior path indicators include jump paths and bounce rates; The conversion indicator includes the click-through rate of the recommendation data.

3. The content real-time push optimization method according to claim 1 is characterized in that: In S1.2, the unified data format is JSON format.

4. The method for optimizing content push in real time according to claim 1, characterized in that: In S2, The log collection system includes a data source integration module and a data processing module; The data source integration module supports collecting user behavior data from the front-end embedding framework and the back-end proxy layer to ensure full data coverage; The data processing module includes an ETL module, a log storage module, and a query service module; ETL module: supports data extraction, conversion and loading, and cleans and standardizes raw data; Log storage module: adopts distributed storage architecture to ensure efficient storage and access; Query service module: Through a unified query interface, it provides fast query support for historical behavior data, providing a data foundation for subsequent analysis and model training.

5. The method for optimizing content push in real time according to claim 1, characterized in that: The S4 comprises the following steps: S4.1 Data extraction: define a short-term period T, and obtain the user's recent behavior data of the short-term period T in real time through the query interface - that is, short-term user behavior; S4.2 Feature standardization: normalize and embed the short-term user behavior data extracted in S4.1 to ensure data consistency; S4.3 Weight enhancement: assigning higher weights to the short-term features of the short-term user behavior, thereby highlighting the impact of the real-time nature of the behavior on the prediction results.

6. The content real-time push optimization method according to claim 7 is characterized in that: In S4.1, the value range of the short-term period T is 5 minutes to 180 minutes.

7. The method for optimizing content push in real time according to claim 1, characterized in that: The S6 comprises the following steps: S6.1 Push method: Push the candidate content in real time through home page recommendation or pop-up window; S6.2 Push optimization: According to the characteristics of the short-term user behavior, dynamically adjust the push frequency and method to improve the user experience.

8. The method for optimizing content push in real time according to claim 1, characterized in that: The S7 comprises the following steps: S7.1 Feedback data collection: record the user interaction data after S6 push, and divide the real-time feedback in the user interaction data into positive feedback and negative feedback; the positive feedback includes click, like, and forward; the negative feedback includes close, disinterest, and report; S7.2 incremental model update: performing incremental training on the student model (small model) using the positive feedback and negative feedback; S7.3: The behavior prediction model is gradually optimized through the student model (small model) to improve the short-term user behavior prediction capability of the behavior prediction model and enhance the adaptability of the behavior prediction model.

Images