Personalized sorting and pushing method based on user behavior feature capture learning

By collecting user behavior data and using Thompson sampling method to build a sorting model, integrating operator connection relationships and user behavior characteristics, the accuracy problem of intelligent push system when processing large-scale and complex data is solved, and the dynamic of personalized information push is realized and efficient.

CN119046526APending Publication Date: 2024-11-29INST OF WAR STUDIES ACAD OF MILITARY SCI OF THE CHINESE PEOPLES LIBERATION ARMY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410999671.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

When existing intelligent push systems process large-scale and complex data, it is difficult to effectively integrate operator connection relationships and user behavior characteristics, resulting in inaccurate information push and difficult to dynamically capture user preferences.

Method used

By collecting user behavior data, obtaining timestamp behavior buried points, counting user behavior characteristics, using Thompson sampling method to build an operator connection matrix, combining logistic regression model, a personalized sorting model based on user behavior is constructed, and operator push is performed.

Benefits of technology

It realizes the effective integration of operator connection relationships with user behavior characteristics, improves the accuracy and dynamic nature of information push, solves the cold start problem, and continuously corrects the probability prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046526A_ABST
    Figure CN119046526A_ABST
Patent Text Reader

Abstract

The invention relates to a personalized sorting and pushing method based on user behavior feature capture learning, and belongs to the technical field of information retrieval. The method comprises the following steps of: 1, triggering at a front end by a user to carry out acquisition, and acquiring behavior burying points in a timestamp form by utilizing user behavior data; 2, acquiring user behavior characteristics consisting of operator offline characteristics, user offline characteristics and cross offline characteristics by counting behavior burying points; 3, operator connection adjacent matrix elements are obtained through probability estimation of Thompson sampling; 4, constructing a sample training sorting model based on user behaviors; 5, performing prediction according to a sequencing model result and a Thompson sampling probability estimation result in the adjacent matrix; wherein the Thompson sampling probability estimation result is that rows where the Thompson sampling probability estimation matrix of the recommended operator is located are ranked. According to the invention, the Thompson sampling and logistic regression method is used to solve the technical problem of operator pushing after the operator connection relation and the user behavior characteristics are fused.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a personalized sorting and pushing method based on user behavior feature capture and learning, and belongs to the technical field of information retrieval. Background Art

[0002] With the development of Internet technology, the phenomenon of information overload has gradually become prominent. Intelligent push uses machine learning and deep learning methods as means, uses user behavior data and user document attributes as features, and provides users with accurate matching in scenarios such as keyword search, related recommendations, and hot lists. The push system is an effective tool to solve information overload. It can help users discover valuable information and has been applied to platforms such as e-commerce and social media. The push system has become a popular modern information system for user-centric intelligent applications. Research on intelligent push systems can be traced back to the 1990s. At first, content-based and collaborative filtering-based methods were the main methods. Later, influenced by the Netflix competition, matrix decomposition technology became the mainstream in the first decade of the 21st century. However, due to the inherent linear nature of the factor decomposition model, its performance in processing large-scale and complex data is lacking, such as complex user interactions and items that may contain complex semantic information. Around 2010, deep learning technology began to have a revolutionary impact in computer vision, natural language processing, speech recognition and other fields, and then many deep learning methods that are still popular today were born.

[0003] The push system plays the role of accurately matching user needs and pushing useful information to users in a timely manner. The information in the system is constantly accumulated and updated over time, and the user's focus will also change continuously. Therefore, how to dynamically capture user behavior preferences and information, push appropriate information to users, and obtain user feedback (such as clicks, browsing, favorites, etc.) has become the key to whether the push system is efficient and easy to use.

[0004] Therefore, how to integrate operator connection relationships with user behavior characteristics and then push operators has become a problem that needs to be solved urgently. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of pushing operators after fusing operator connection relationships with user behavior features, and to propose a personalized sorting and pushing method based on user behavior feature capture and learning.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A personalized sorting and pushing method based on user behavior feature capture and learning of the present invention comprises the following steps:

[0008] Step 1: Collect data through user triggers on the front end, and use user behavior data to obtain behavior tracking points in the form of timestamps;

[0009] Step 2: Obtain user behavior features consisting of operator offline features, user offline features, and cross-offline features by counting behavior tracking points;

[0010] Step 3: Obtain the operator connection adjacency matrix elements using the probability estimation of Thompson sampling;

[0011] Step 3.1: Initialize the adjacency matrix and initialize the operators according to the operator type;

[0012] Step 3.2: Connect the operators in the data pipeline, and use the Thompson sampling method to update the alpha adjacency matrix elements and the beta adjacency matrix elements by traversing the operator connection information;

[0013] Step 3.3: Divide the alpha adjacency matrix by the beta adjacency matrix to obtain the Thompson sampling probability estimation matrix;

[0014] Step 4: Build a sample training ranking model based on user behavior;

[0015] Step 4.1: Construct training samples; select the user dragging behavior information and user connection behavior information within the sampling period as positive samples; other candidate operators that are not dragged and connected are used as negative samples, and the sample threshold is set to ensure the sampling volume of positive and negative samples. Use equations (1) and (2) to make the samples obey the logistic distribution and form a sample space;

[0016]

[0017] Where F(x) is the sample distribution function, f(x) is the probability density function, μ is the location parameter, f is the density function, X is the sample, γ is the shape parameter, and γ>0;

[0018] Step 4.2: Use sklearn to train the sorting model for the above samples and the user behavior features described in step 2;

[0019] Step 5: Predict based on the ranking model results obtained through training and the Thompson sampling probability estimation results in the adjacency matrix; the Thompson sampling probability estimation results in the adjacency matrix are ranked for the rows of the Thompson sampling probability estimation matrix of the recommended operator;

[0020] Beneficial effects:

[0021] 1. The present invention applies the Thompson sampling method to the probability estimation modeling of interconnected directed graphs, uses the adjacency matrix to represent the parameters of Thompson's sampling, and provides a new directed graph probability estimation method.

[0022] 2. The present invention combines the Thompson sampling and logistic regression methods, taking into account both spatial information and user preferences, and can sort the results more comprehensively and reasonably.

[0023] 3. In actual use of the present invention, the Thompson sampling method can solve the deviation caused by directly calculating the probability, solve the cold start problem, and can continuously supplement and correct the probability prediction results as the pipeline is added. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the process of the present invention;

[0025] Figure 2 It is the adjacency matrix parameter update graph of Thompson sampling of the present invention. DETAILED DESCRIPTION

[0026] In order to better illustrate the purpose and advantages of the present invention, the invention is further described below in conjunction with the accompanying drawings and examples. It should be noted that the implementation of the present invention is not limited to the following embodiments, and any form of modification or change made to the present invention will fall within the protection scope of the present invention.

[0027] Example

[0028] like Figure 1 As shown, the personalized sorting and pushing method based on user behavior feature capture and learning of the present invention has the following specific implementation steps:

[0029] Step 1: Collect data through user triggers on the front end, and use user behavior data to obtain behavior tracking points in the form of timestamps;

[0030] In the embodiment, the tracking points record the behaviors between the user and the operated object, including exposure, clicking, dragging, connecting and other behaviors, the ID information and location information of the user and the operated object, and the timestamp of the event.

[0031] Step 2: Obtain user behavior features consisting of operator offline features, user offline features, and cross-offline features by counting behavior tracking points;

[0032] In the embodiment, the system can calculate offline user features, offline operator features and offline cross features. The offline user features are the statistics of the buried point information in the user dimension, such as the number of operators clicked by a certain user within 7 days; the offline operator features refer to the statistics of the buried point information in the operator dimension, such as the number of times a JSON input operator is connected within 7 days; the cross feature refers to the statistics of the user on a certain dimension attribute of the operator, such as the proportion of operators clicked by users with built-in operators in 7 days to all operators clicked within 7 days.

[0033] Step 3: Obtain the operator connection adjacency matrix elements using the probability estimation of Thompson sampling;

[0034] Step 3.1: Initialize the adjacency matrix and initialize the operators according to the operator type;

[0035] Step 3.2: If Figure 2 As shown, the operators in the data pipeline are connected, and the traversal operator connection information is used to update the alpha adjacency matrix elements and the beta adjacency matrix elements using the Thompson sampling method;

[0036] Step 3.3: Divide the alpha adjacency matrix by the beta adjacency matrix to obtain the Thompson sampling probability estimation matrix;

[0037] In the embodiment, the embodiment is constructed according to all operators in the system. If the number of operators in the system is n, two n*n adjacency matrices are constructed, namely the alpha matrix and the beta matrix. The alpha matrix records the alpha parameter in the Thompson sampling rate, and the beta matrix records the beta parameter in the Thompson sampling. During initialization, if the operator is a starting operator, the alpha of other operators connected to the starting operator is 0, and beta is any positive number; if the operator is a terminating operator, the alpha of the terminating operator connected to other operators is 0, and beta is any positive number; if the operator is a processing operator, the alpha of all connections except the first two cases is assigned to 5, and beta is 5*n. The reason for this assignment is to give alpha and beta a larger parameter, and the ratio of their parameters, that is, the value of alpha / beta, is the connection probability estimate of the operator. Afterwards, when traversing the operator connection in each pipeline in the system, if a connection is connected from operator a to operator b, operator a connects to operator b's alpha+1 and beta+1, and operator a connects to other operators except that the alpha of other operators remains unchanged and beta+1.

[0038] Step 4: Build a sample training ranking model based on user behavior;

[0039] Step 4.1: Construct training samples; select the user dragging behavior information and user connection behavior information within the sampling period as positive samples; other candidate operators that are not dragged and connected are used as negative samples, and the sample threshold is set to ensure the sampling volume of positive and negative samples. Use equations (1) and (2) to make the samples obey the logistic distribution and form a sample space;

[0040]

[0041] Where F(x) is the sample distribution function, f(x) is the probability density function, μ is the location parameter, f is the density function, X is the sample, γ is the shape parameter, and γ>0;

[0042] Step 4.2: Use sklearn to train the sorting model for the above samples and the user behavior features described in step 2;

[0043] In the embodiment: determine the sampling time period, select dragging, connection, etc. as positive samples; other candidate operators that are not dragged and connected are used as negative samples, and set the sample threshold to ensure the sampling volume of positive and negative samples. By using the features obtained in step 3 as input and the sample values ​​as output to train the logistic regression model, a personalized ranking model based on user behavior is obtained.

[0044] Step 5: Predict based on the ranking model results obtained through training and the Thompson sampling probability estimation results in the adjacency matrix; the Thompson sampling probability estimation results in the adjacency matrix are ranked for the rows of the Thompson sampling probability estimation matrix of the recommended operator;

[0045] In the embodiment, the sorting result obtained by Thompson sampling and the corresponding probability estimation sorting result in the adjacency matrix are fused through the RRF algorithm (Reciprocal Rank Fusion is used in hybrid search) as shown in formula (3) to finally obtain the final sorting result.

[0046]

[0047] Among them, k is a hyperparameter, D is the complete set of items, R is the set of sorting functions, and r(d) is the ranking of the items.

[0048] The results show that the personalized sorting and pushing method based on user behavior feature capture and learning proposed in the present invention can integrate the operator connection relationship with the user behavior features to push operators.

Claims

1. A personalized sorting and pushing method based on user behavior feature capture and learning, characterized by: The following steps are included: Step 1: Collect data through user triggers on the front end, and use user behavior data to obtain behavior tracking points in the form of timestamps; Step 2: Obtain user behavior features consisting of operator offline features, user offline features, and cross-offline features by counting behavior tracking points; Step 3: Obtain the operator connection adjacency matrix elements using the probability estimation of Thompson sampling; Step 4: Build a sample training ranking model based on user behavior; Step 4.1: Construct training samples; The user dragging behavior information and user connection behavior information within the sampling time period are selected as positive samples; other candidate operators that are not dragged and connected are selected as negative samples, and the sample threshold is set to ensure the sampling volume of positive and negative samples; the samples are subjected to logistic distribution using equations (1) and (2) to form a sample space; Where F(x) is the sample distribution function, f(x) is the probability density function, μ is the location parameter, f is the density function, X is the sample, γ is the shape parameter, and γ>0; Step 4.2: Use sklearn to train the sorting model for the above samples and the user behavior features described in step 2; Step 5: Make predictions based on the ranking model results obtained through training and the Thompson sampling probability estimation results in the adjacency matrix; the Thompson sampling probability estimation results in the adjacency matrix are ranked for the rows of the Thompson sampling probability estimation matrix of the recommended operator.

2. A personalized sorting and pushing method based on user behavior feature capture and learning as claimed in claim 1, characterized in that: Step 3 is implemented as follows: Step 3.1: Initialize the adjacency matrix and initialize the operators according to the operator type; Step 3.2: Connect the operators in the data pipeline, and use the Thompson sampling method to update the alpha adjacency matrix elements and the beta adjacency matrix elements by traversing the operator connection information; Step 3.3: Divide the alpha adjacency matrix by the beta adjacency matrix to obtain the Thompson sampling probability estimation matrix.