[go: up one dir, main page]

WO2025118131A1 - Procédé de recommandation intelligente basé sur des caractéristiques de comportement d'utilisateur - Google Patents

Procédé de recommandation intelligente basé sur des caractéristiques de comportement d'utilisateur Download PDF

Info

Publication number
WO2025118131A1
WO2025118131A1 PCT/CN2023/136354 CN2023136354W WO2025118131A1 WO 2025118131 A1 WO2025118131 A1 WO 2025118131A1 CN 2023136354 W CN2023136354 W CN 2023136354W WO 2025118131 A1 WO2025118131 A1 WO 2025118131A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
user
cin
ctr
layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2023/136354
Other languages
English (en)
Chinese (zh)
Inventor
蔡洪斌
卢光辉
赵晨
胡耀东
艾鑫
晏小虎
段雅俊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to PCT/CN2023/136354 priority Critical patent/WO2025118131A1/fr
Publication of WO2025118131A1 publication Critical patent/WO2025118131A1/fr
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation

Definitions

  • the present invention belongs to the technical field of big data analysis, and in particular relates to an intelligent recommendation method based on user behavior characteristics.
  • the present invention provides an intelligent recommendation method based on user behavior characteristics, which includes the following main steps:
  • Step 1 Collect user data.
  • the user data includes the user's basic information, implicit information, and behavior information. This step includes:
  • Step 1.1 collect basic information of users: when users register, obtain basic descriptive information characteristics of users, including gender, age, education level and other information.
  • Step 1.2 collect the user's implicit information: the user completes the necessary assessment process, and the user's implicit information is extracted according to the assessment process, including various assessment ability values, personality tendencies and other information.
  • Step 1.3 collect user behavior information: When the user enters the recommendation page, classify and mark the user's behavior information on different recommended content feedback, including behavior types such as ignore, view, collect, purchase, etc., and record the time of the operation.
  • behavior types such as ignore, view, collect, purchase, etc.
  • Step 2 Reshape the user data. This step includes:
  • Step 2.1 data cleaning: Use the average value or special marking method to clean invalid data in the user data.
  • Step 2.2 combination construction: According to the needs of subsequent screening of CTR models, different combinations of the data collected in step 1 are constructed.
  • Step 2.3 data preprocessing: perform corresponding encoding conversion operations and embedding processing operations on the categorical feature data and continuous feature data in the data collected in step 1, so that they can be used as input of the CTR model.
  • Step 3 Construct a CTR model based on LR. Based on the logistic regression algorithm, construct a CTR model. This step includes:
  • Step 3.1 input information features: select the user's basic information features, the user's implicit information features, the user's behavior information features, whether to click on the label, and the product information as model input.
  • Step 3.2 combine informative features: combine these features using a first-order linear relationship function.
  • Step 3.3 perform training calculation: according to the actual result of whether the click is made, perform training calculation to obtain the cross entropy of the linear calculation result.
  • Step 4 Build a CTR model based on CIN. Based on the cross network principle, build a CTR model. This step includes:
  • Step 4.1 input information features. Select the user's basic information features, the user's implicit information features, the user's behavior information features (whether to click on the label) and the product information as model input.
  • Step 4.2 construct the first layer vector. Construct the output vector of the first layer of the compressed interaction network CIN. This step includes:
  • Step 4.2.1 calculate the Hadamard product of the embedded input vector with itself to obtain multiple two-dimensional matrices.
  • Step 4.2.2 multiply multiple two-dimensional matrices by the corresponding coefficient matrix to obtain a feature map vector.
  • Step 4.2.3 Repeat step 4.2.2 as needed to obtain multiple feature map vectors to form the first layer of the CIN network.
  • Step 4.3 construct multi-layer vectors.
  • the CIN network vector of the previous layer is calculated with the input vector as in step 4.2 to continue constructing the next layer of CIN network.
  • Step 4.4 compress the results.
  • a random value pooling method according to a certain coefficient, compress the results of each layer of the CIN network to obtain the output of the CIN network part.
  • Step 4.5 Joint training: Construct the DNN part and the linear part of the CTR model and perform joint training with the CIN part.
  • Step 4.6 calculate the cross entropy. Use the sigmoid function to jointly calculate the outputs of the three parts, and perform cross entropy calculation as the output of the entire model.
  • Step 4.7 regularize the objective function. Regularize the objective function, and add Batch to the deep neural network part.
  • the Normalization layer is used to improve the generalization ability, and Dropout is used for neurons to prevent overfitting in training and L2 regularization is adopted; while the Embedding layer and CIN network part only use L2 regularization.
  • Step 5 Apply the model to filter the recommendations.
  • the two-layer recommendation model is used to filter the recommendation results. This step includes:
  • Step 5.1 Preliminary screening: Based on the category selected by the user, the corresponding recommended content is put into the LR-based CTR model for preliminary screening.
  • Step 5.2 Secondary screening: The results after the initial screening are sent to the CTR model based on CIN for final screening. Finally, the results are sorted according to the output probability, and the top 20 results are obtained as the final output.
  • the beneficial effect of the present invention is: proposing an intelligent recommendation method based on user behavior characteristics, which can reasonably utilize user basic information characteristics, user implicit information characteristics, and user behavior characteristics, greatly reduce the training cost when training the recommendation model, and improve the accuracy of recommendation.
  • FIG1 shows a flow chart of an intelligent recommendation method based on user behavior characteristics of the present invention
  • FIG2 shows the use process of an intelligent recommendation method based on user behavior characteristics of the present invention
  • FIG3 shows the specific structure of the core algorithm model of the intelligent recommendation method of the present invention
  • FIG4 shows the collection and conversion process of the input data of the model of the present invention
  • Figures 5 and 6 show the calculation process of the feature map part of the CIN network part of the present invention.
  • FIG1 shows the basic process of the implementation of the present invention
  • the network structure diagram shown in FIG2 shows the structure of the entire innovative model of the present invention:
  • Step 1 Collect user data. Collect the user's input feature data in the CTR model. This step includes:
  • Step 1.1 Collect basic information of users.
  • users register obtain basic descriptive information characteristics of users, such as gender, age, and education level.
  • Step 1.2 Collect the user's implicit information.
  • the user completes the necessary assessment process and uses the assessment form to extract the user's implicit information from the information obtained from the assessment process, including various assessment ability values, personality tendencies and other information.
  • Step 1.3 Collect user behavior information.
  • a user enters the recommendation page classify and mark the user's behavior on different recommended content information, including: ignore, view, collect, purchase and other behavior types, and record the time of the operation, so as to update the user's behavior information in real time.
  • Step 2 Reshape the user data. Reshape the data obtained in step 1 to make it conform to the requirements of the following steps.
  • This step includes:
  • Step 2.1 Clean the data. Use the average value or special marking method to clean the invalid data collected in step 1.
  • Step 2.2 Combination construction: According to the needs of subsequent screening of CTR models, different combinations of the data collected in step 1 are constructed.
  • Step 2.3 Data preprocessing. Perform corresponding encoding conversion and embedding operations on the categorical feature data and continuous feature data collected in step 1, so that they can be used as input for the CTR model. This step includes:
  • Step 2.3.1 Perform a simple normalization operation on continuous data and use label encoding conversion operation on categorical data.
  • Step 2.3.2 Based on the LR model, the categorical data that has undergone the label encoding conversion operation in step 2.3.1 is normalized, and the continuous data that has undergone a simple normalization operation in step 2.3.1 is jointly trained with the model.
  • Step 2.3.3 Based on the CIN model, use one-hot encoding to convert the categorical data that has undergone label encoding conversion in step 2.3.1 into sparse vector data, and then perform embedding operation on it and the continuous data that has undergone simple normalization operation in step 2.3.1 to obtain dense vectors respectively, and use the two dense vectors for model training.
  • Step 3 Construct a CTR model based on LR. Based on the logistic regression algorithm, construct a CTR model. This step includes:
  • Step 3.1. Input information features Select the user's basic information features, the user's implicit information features, the user's behavior information features (whether to click on the label) and the product information as the model input.
  • Step 3.2 Combine informative features. Combine these features using a first-order linear relationship function.
  • Step 3.3 Perform training calculations. Perform training calculations to find cross entropy on the linear calculation results based on the actual click results.
  • Step 4 Build a CTR model based on CIN. Based on the cross network principle, build a CTR model. This step includes:
  • Step 4.1 Input information features. Select the user's basic information features, the user's implicit information features, and the user's behavior The information features of whether the label is clicked and the product information are used as model inputs.
  • Step 4.2. Construct the first layer vector. Construct the output vector of the first layer of the compressed interaction network CIN. This step includes:
  • Step 4.2.1 Calculate the Hadamard product of the embedded input vector with itself to obtain multiple two-dimensional matrices.
  • Step 4.2.2 Multiply multiple matrices by the corresponding coefficient matrix to obtain a feature map vector.
  • Step 4.2.3 Repeat step 4.2.2 as required to obtain multiple feature map vectors to form the first layer of CIN network.
  • the calculation process diagram is shown in Figures 5 and 6, and the calculation formula is as follows:
  • Step 4.3 Construct multi-layer vectors.
  • the CIN network vector of the previous layer is calculated with the input vector as in step 4.2, and the next layer of CIN network is constructed.
  • the calculation formula is as follows:
  • Step 4.4. Result compression By using a random value pooling method according to a certain coefficient, the result of each layer of the CIN network is compressed to obtain the output of the CIN network part.
  • Step 4.5 Joint training: Construct the DNN part and the linear part of the CTR model, and perform joint training with the CIN part.
  • Step 4.6 Cross entropy calculation. Use the sigmoid function to jointly calculate the outputs of the three parts, and perform cross entropy calculation as the output of the entire model.
  • the calculation formula is as follows:
  • Step 4.7 Regularize the objective function. Regularize the objective function.
  • Add a Batch Normalization layer to improve the generalization ability, use Dropout for neurons to prevent overfitting and use L2 regularization; while only use L2 regularization for the Embedding layer and CIN network part.
  • Step 5 Apply the model to filter the recommendations.
  • the two-layer recommendation model is used to filter the recommendation results. This step includes:
  • Step 5.1 Preliminary screening: Based on the category selected by the user, the corresponding recommended content is put into the LR-based CTR model for preliminary screening.
  • Step 5.2. Secondary screening The results after the initial screening are sent to the CIN-based CTR model for final screening, and finally sorted according to the output probability, and the top 20 results are obtained as the final output.

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention concerne un procédé de recommandation intelligente basé sur des caractéristiques de comportement d'utilisateur. Le procédé comprend la collecte de données, l'organisation de données, la construction d'un modèle CTR basé sur LR, la construction d'un modèle CTR basé sur CIN, et la sélection et la recommandation de modèle. Sur la base de l'idée de conception d'un modèle de recommandation de CTR, le procédé prend en compte des caractéristiques de comportement d'utilisateur telles que la visualisation, l'achat et la mise en favoris, ce qui permet de réaliser des recommandations raisonnables sur la base de comportements de groupes d'utilisateurs ; de plus, des données d'utilisateur massives sont nettoyées et remises en forme pour filtrer des données valides, ce qui permet d'améliorer la vitesse d'apprentissage ; de plus, afin d'améliorer l'efficacité de prédiction et la précision de prédiction, un algorithme de termes d'interaction de caractéristiques explicites d'ordre élevé est en outre introduit. Le procédé améliore efficacement le processus d'algorithmes de recommandation classiques, améliorant ainsi la précision de recommandation.
PCT/CN2023/136354 2023-12-05 2023-12-05 Procédé de recommandation intelligente basé sur des caractéristiques de comportement d'utilisateur Pending WO2025118131A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2023/136354 WO2025118131A1 (fr) 2023-12-05 2023-12-05 Procédé de recommandation intelligente basé sur des caractéristiques de comportement d'utilisateur

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2023/136354 WO2025118131A1 (fr) 2023-12-05 2023-12-05 Procédé de recommandation intelligente basé sur des caractéristiques de comportement d'utilisateur

Publications (1)

Publication Number Publication Date
WO2025118131A1 true WO2025118131A1 (fr) 2025-06-12

Family

ID=95981399

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/136354 Pending WO2025118131A1 (fr) 2023-12-05 2023-12-05 Procédé de recommandation intelligente basé sur des caractéristiques de comportement d'utilisateur

Country Status (1)

Country Link
WO (1) WO2025118131A1 (fr)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108182597A (zh) * 2017-12-27 2018-06-19 银橙(上海)信息技术有限公司 一种基于决策树和逻辑回归的点击率预估方法
CN111563770A (zh) * 2020-04-27 2020-08-21 杭州金智塔科技有限公司 一种基于特征差异化学习的点击率预估方法
WO2021159640A1 (fr) * 2020-02-13 2021-08-19 平安科技（深圳）有限公司 Procédé de recommandation de médicament basé sur l'intelligence artificielle et dispositif associé
CN114861045A (zh) * 2022-04-18 2022-08-05 北京快乐茄信息技术有限公司 数据处理方法及装置、电子设备及存储介质
CN114925270A (zh) * 2022-05-09 2022-08-19 华南师范大学 一种会话推荐方法和模型
CN115271789A (zh) * 2022-06-29 2022-11-01 沈阳建筑大学 一种融合压缩交叉网络的混合推荐方法
WO2023051678A1 (fr) * 2021-09-29 2023-04-06 华为技术有限公司 Procédé de recommandation et dispositif associé

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108182597A (zh) * 2017-12-27 2018-06-19 银橙(上海)信息技术有限公司 一种基于决策树和逻辑回归的点击率预估方法
WO2021159640A1 (fr) * 2020-02-13 2021-08-19 平安科技（深圳）有限公司 Procédé de recommandation de médicament basé sur l'intelligence artificielle et dispositif associé
CN111563770A (zh) * 2020-04-27 2020-08-21 杭州金智塔科技有限公司 一种基于特征差异化学习的点击率预估方法
WO2023051678A1 (fr) * 2021-09-29 2023-04-06 华为技术有限公司 Procédé de recommandation et dispositif associé
CN114861045A (zh) * 2022-04-18 2022-08-05 北京快乐茄信息技术有限公司 数据处理方法及装置、电子设备及存储介质
CN114925270A (zh) * 2022-05-09 2022-08-19 华南师范大学 一种会话推荐方法和模型
CN115271789A (zh) * 2022-06-29 2022-11-01 沈阳建筑大学 一种融合压缩交叉网络的混合推荐方法

Similar Documents

Publication Publication Date Title
CN112579778A (zh) 基于多层次的特征注意力的方面级情感分类方法
CN111275521A (zh) 一种基于用户评论与满意层面嵌入的商品推荐方法
CN111538761A (zh) 基于注意力机制的点击率预测方法
CN112000771B (zh) 一种面向司法公开服务的句子对智能语义匹配方法和装置
CN112860998B (zh) 一种基于多任务学习机制的点击率预估方法
CN117196763A (zh) 基于时序感知自注意力和对比学习的商品序列推荐方法
CN114238758B (zh) 一种基于多源跨界数据融合的用户画像预测方法
CN109064294B (zh) 一种融合时间因素、文本特征和相关性的药品推荐方法
CN108492118A (zh) 汽车售后服务质量评价回访文本数据的两阶段抽取方法
CN113254782A (zh) 问答社区专家推荐方法及系统
CN114491248B (zh) 一种基于用户多意图演进的序列推荐方法
CN115712777A (zh) 一种基于逻辑回归的文献推荐系统的排序方法
CN112667824A (zh) 基于多语义学习的知识图谱补全方法
CN117333037A (zh) 一种面向出版大数据的产业大脑构建方法及装置
CN115062727A (zh) 一种基于多阶超图卷积网络的图节点分类方法及系统
CN116468460A (zh) 基于人工智能的消费金融客户画像识别系统及其方法
CN118014652A (zh) 一种基于人工智能和大数据分析技术的广告创意设计方法及其系统
CN106202515A (zh) 一种基于排序学习的移动应用推荐方法及其推荐系统
Jin GraphCNNpred: A stock market indices prediction using a Graph based deep learning system
CN118395989A (zh) 一种基于超图注意力机制的人格检测方法
CN111950717B (zh) 一种基于神经网络的舆情量化方法
CN113362034A (zh) 一种职位推荐方法
CN117574915A (zh) 基于多方数据源的公共数据平台及其数据分析方法
CN119809724B (zh) 基于人工智能的跨地域广告投放分析决策方法及其系统
CN114880538A (zh) 基于自监督的属性图社团检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23960446

Country of ref document: EP

Kind code of ref document: A1