User interest degree prediction method and interest degree prediction model training method

Abstract

The embodiment of the invention discloses an interest degree prediction model training method and a user interest degree prediction method.The interest degree prediction model training method comprises the steps that multi-dimensional data of a user are obtained, and the multi-dimensional data comprise eye movement data of the user for a target object and user behavior data; the user behavior data comprises function information of the target object used by the user; the users are clustered according to the function information of the target objects used by the users, a plurality of user groups are obtained, and functions used by the users in different user groups are different; and performing iterative training on the initial interest degree prediction model by using the user behavior data and the eye movement data corresponding to each user group to obtain an interest degree prediction model, the interest degree prediction model being used for predicting the interest degree of the target user to the target object.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method for predicting a user's interest level and a method for training an interest level prediction model. Background Art

[0002] With the popularization of the Internet and the growing demand for personalized services, understanding user interests has become the key to online platforms and e-commerce. Determining user interests can help operators better understand user interest preferences, facilitate target object design and function development to better meet user needs, and help operators develop personalized recommendation strategies for users to improve user experience.

[0003] In some scenarios, when determining the interests of users, questionnaires are often used to directly collect users' personal attributes, such as age, occupation, interests, and gender, and then infer users' interests based on their personal attributes. However, the information obtained from questionnaires is usually static and cannot reflect changes in users' interests in real time. Users' interests may change due to various factors such as social events, personal experiences, and seasonal changes, but questionnaires can often only be conducted at specific time points and cannot reflect users' real-time changes in interests. In addition, questionnaires are usually limited to a pre-set range of questions and may not cover more of the user's interests. Therefore, inferring users' interests based on questionnaire data has low reliability and accuracy. Summary of the invention

[0004] The present application provides a method for predicting the degree of user interest, a method and device for training an interest degree prediction model, and solves the problem of low reliability and accuracy when inferring the user's interest.

[0005] In a first aspect, the present application provides a method for training an interest level prediction model, comprising: obtaining multi-dimensional data of users, the multi-dimensional data including eye movement data and user behavior data of users for a target object, the user behavior data including functional information of the target object used by the users; clustering users according to the functional information of the target object used by the users to obtain multiple user groups, wherein users in different user groups use different functions; iteratively training an initial interest level prediction model using the user behavior data and eye movement data corresponding to each user group to obtain an interest level prediction model, wherein the interest level prediction model is used to predict the target user's interest level in the target object.

[0006] In a second aspect, the present application provides a method for predicting a user's interest level, comprising: obtaining the target eye movement data, target behavior data and the target user group to which the target user belongs for a target object; inputting the target eye movement data, target user group and target behavior data into an interest level prediction model, wherein the interest level prediction model is trained using the interest level prediction model training method mentioned in the first aspect; and predicting the target user's interest level in the target object using the interest level prediction model.

[0007] In a third aspect, the present application provides a training device for an interest level prediction model, comprising: an acquisition module for acquiring multi-dimensional data of users, the multi-dimensional data including eye movement data and user behavior data of users for a target object, the user behavior data including functional information of the target object used by the users; a clustering module for clustering users according to the functional information of the target object used by the users, and obtaining multiple user groups, wherein users in different user groups use different functions; a training module for iteratively training an initial interest level prediction model using the user behavior data and eye movement data corresponding to each user group, and obtaining an interest level prediction model, wherein the interest level prediction model is used to predict the target user's interest level in the target object.

[0008] In a fourth aspect, the present application provides a device for predicting a user's interest level, comprising: an acquisition module for acquiring a target user's target eye movement data, target behavior data, and a target user group to which the target user belongs for a target object; a prediction module for inputting the target eye movement data, target user group, and target behavior data into an interest level prediction model, wherein the interest level prediction model is trained using the interest level prediction model training method mentioned in the first aspect; and predicting the target user's interest level in the target object using the interest level prediction model.

[0009] In a fifth aspect, the present application provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute instructions to implement a method as in the first aspect or the second aspect.

[0010] In a sixth aspect, the present application provides a computer-readable storage medium. When instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute the method of the first aspect or the second aspect.

[0011] In the seventh aspect, an embodiment of the present application provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the method steps mentioned in the first aspect or the second aspect.

[0012] Through the technical solution disclosed in the embodiment of the present application, by obtaining the multi-dimensional data of the user, the multi-dimensional data includes the user's eye movement data and user behavior data for the target object, and the user behavior data includes the user's use of the function information of the target object; then the user is clustered according to the function information of the target object used by the user, and multiple user groups are obtained, and the functions used by users in different user groups are different; then the initial interest level prediction model is iteratively trained using the user behavior data and eye movement data corresponding to each user group to obtain the interest level prediction model, and the interest level prediction model is used to predict the target user's interest level in the target object. In this way, by refining the user group, and using the eye movement data and user behavior data of various user groups to train the prediction model, the prediction model is prevented from overfitting, and the generalization ability of the model is improved. When the prediction model is applied to the prediction of the user's interest level, the prediction accuracy can be effectively improved. And the user's eye movement data and user behavior data can reflect the user's interest changes in real time. When the prediction model predicts the interest level based on the user's multi-dimensional data, the predicted user interest level has high reliability and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings described herein are used to provide a further understanding of this specification and constitute a part of this specification. The illustrative embodiments and descriptions of this specification are used to explain this specification and do not constitute an improper limitation on this specification. In the drawings:

[0014] Figure 1 A flowchart of a method for training an interest level prediction model provided in an embodiment of the present application;

[0015] Figure 2 A flowchart of a method for predicting user interest level provided in an embodiment of the present application;

[0016] Figure 3 A schematic diagram of the structure of a training device for an interest level prediction model provided in an embodiment of the present application;

[0017] Figure 4 A schematic diagram of the structure of a device for predicting user interest level provided in an embodiment of the present application;

[0018] Figure 5 A schematic diagram of the structure of an electronic device provided for one embodiment of the present specification. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0020] The terms "first", "second", etc. in this specification and claims are used to distinguish similar objects, and are not used to describe a specific order or precedence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein. In addition, "and / or" in this specification and claims means at least one of the connected objects, and the character " / " generally means that the objects associated with each other are in an "or" relationship.

[0021] For example, Figure 1 As shown, the embodiment of the present application provides a method for training an interest level prediction model. The execution subject of the method may be a server. The method for training the interest level prediction model may specifically include the following steps:

[0022] Step S101, obtaining multi-dimensional data of the user.

[0023] Among them, the multi-dimensional data includes the user's eye movement data and user behavior data on the target object, and the user behavior data includes the functional information of the user using the target object.

[0024] Specifically, wearable eye tracker devices can be used to obtain eye tracker-related data, such as the user's line of sight, gaze, Automated Optical Inspection (AOI) related indicators, eye saccade indicators, gaze duration, and angular velocity. After collecting the eye movement data, in order to remove the noise in the eye movement data, that is, to smooth the data curve, a sliding median filter method is used. Each eye movement data is the average value of the defined number before and after the current data. The specific formula is as follows:

[0025]

[0026] In the above formula, y[n] represents the average value of the eye movement data, N represents the manually set window sample size, generally N=3, n is the number of eye movement data, and x represents the eye movement data.

[0027] Finally, a trajectory map is generated based on the speed threshold algorithm, specifically including: generating a gaze sequence based on the gaze position and time information of the eyes and visually representing it based on the speed threshold algorithm. The generation of the gaze sequence depends on where the user's eyes stay, how long they stay at each position, and how much attention the user pays. Through the gaze sequence, the user's gaze order can be known.

[0028] Furthermore, the multi-dimensional data also includes the user's subjective evaluation data on the target object, which is used to identify and adjust the eye movement data obtained through the wearable eye tracker device. Since users with different familiarity levels have different sight lines when using the same function, the eye movement data is identified and adjusted by collecting interview data on the user's eye movement data to improve the accuracy and reliability of the eye movement data.

[0029] Specifically, users fill out a questionnaire before participating in the assessment. Based on the questionnaire results, users are divided into user groups with different familiarity levels. For each user group, their eye movement data measured using an eye tracker is collected. After preliminary screening and analysis of the eye movement data, further screening and adjustment are performed based on the subjective evaluation data to evaluate the reliability and accuracy of the eye movement data.

[0030] User behavior data includes but is not limited to the user's use of the target object's functions and the operation of the target object's functions. Functional use behavior includes but is not limited to important operation record information such as uploading, downloading first-level backups, downloading videos, and using documents. Operation behavior includes but is not limited to browsing albums, uploading, and backing up. By constructing multi-dimensional data, including eye movement data, user behavior data, and user interview data, and associating eye movement data through the user's unique identification code, the multi-dimensional data is shown in the following table:

[0031]

[0032] In a possible implementation, in order to improve the accuracy of the data, after acquiring the multi-dimensional data of the user, the method further includes: performing data cleaning on the multi-dimensional data.

[0033] Specifically, the multi-dimensional data is cleaned and associated through conventional cleaning methods. Data cleaning includes but is not limited to: processing missing values, outliers, performing mean interpolation and Lagrange interpolation for null values ​​of indicators, and deleting characteristic indicators with a large number of missing values.

[0034] By cleaning the multi-dimensional data, the obtained multi-dimensional data is converted into data that can be trained by the prediction model. Since directly inputting multi-dimensional data will result in information redundancy and unclear features, redundancy is eliminated through cleaning methods to emphasize important features. For eye movement data, the functions corresponding to the target objects that the user has looked at are recorded in the order of the user's line of sight. Construct feature engineering for eye movement data and splice the user's eye movement data into vector input. In this way, by cleaning the eye movement data in a targeted manner, the use effect of the eye movement data by the later prediction model is improved, making the prediction of user interests more accurate.

[0035] Finally, the cleaned eye movement data, user behavior data, and subjective evaluation data are linked through the user's unique identification field to obtain a complete historical training data set. User behavior data also includes the maximum usage space, 30-day active times, and monthly active days.

[0036] Step S103: clustering users according to the function information of the target objects used by the users to obtain multiple user groups.

[0037] Among them, users in different user groups use different functions.

[0038] Specifically, the cleaned user behavior data can be input into the clustering model, and users can be clustered through the clustering model. For example, the functional demand types of user behavior data can be divided into five user groups, such as content storage user groups, content viewing user groups, content processing user groups, content acquisition and sharing user groups, and platform welfare user groups. The types of user groups are added to the cleaned historical training data set as labels, and are mainly associated with eye movement data. Combined with the user group type and the survey of users' familiarity with the functions of the target object, the parameters of the subsequent prediction model are accurately adjusted.

[0039] Among them, the user group of content storage will focus on the behavioral information corresponding to functional modules such as user upload and backup; the user group of content viewing will focus on the information corresponding to the functions of modules such as albums and videos; the user group of content processing will focus on the functions of photo tools and document tools; the user group of content acquisition and sharing will focus on indicators such as discovery and circles; the user group of platform benefits will focus on activity and rights-related indicator information.

[0040] In the process of feature selection and subsequent prediction model setting, the feature weight configuration of different user groups will be different. For example, for the user group of content storage, their data concentration is more affected by indicators such as upload and backup. In this way, based on the feature form required by the tree model, the eye movement data is cleaned in a targeted manner and the weight is adjusted by combining the type of user group through data features, which improves the effect of the tree model on the eye movement data and makes the model more accurate in predicting the interests of different types of user groups.

[0041] Step S105, iteratively training the initial interest level prediction model using the user behavior data and eye movement data corresponding to each user group to obtain an interest level prediction model.

[0042] Among them, the interest level prediction model is used to predict the interest level of the target user in the target object.

[0043] Specifically, the initial interest level prediction model can be a weak classifier, and the weak classifier can be a decision tree. The user behavior data and eye movement data are input into the initial interest level prediction model for iterative training to obtain the interest level prediction model. In the embodiment of the present application, the user's interest level in the target object is used as the prediction target, and the target object can be a certain product. The initial interest level prediction model is accurately optimized in combination with multi-dimensional data, and the weight parameters of the initial interest level prediction model are dynamically adjusted to achieve fine control, and the loss function of the initial interest level prediction model is optimized based on the eye movement data.

[0044] In a possible implementation, iterative training of the initial interest level prediction model using the user behavior data and eye movement data corresponding to each user group includes: constructing a weak classifier and using the weak classifier as the initial interest level prediction model; iteratively training the weak classifier using the user behavior data and eye movement data corresponding to each user group until a predetermined number of iterations is reached to obtain the interest level prediction model; wherein each iterative training generates a new weak classifier, and the weak classifiers generated by multiple iterative trainings are superimposed to obtain the interest level prediction model.

[0045] Specifically, the weak classifier can be a decision tree. A new decision tree will be constructed each time the weak classifier is iterated. The decision tree is constructed through the residual between the true value and the predicted value. In the cleaned historical data training set, various feature combinations are formed by matrices of different values ​​of various indicators such as the type of user group, the order of sight of eye movement data, the maximum usage space, the number of active times in 30 days, the number of active days in a month, etc. The branches of the decision tree will be different. The residual will be considered when constructing the second decision tree, that is, the new weak classifier is constructed by the residual between the predicted value and the true value of the previous weak classifier. When the number of iterations reaches the upper limit or the residual no longer decreases, it stops, and a strong classifier model with multiple decision trees is obtained. This strong classifier model is the interest level prediction model.

[0046] In one possible implementation, constructing a weak classifier includes: constructing a loss function, which indicates the difference between a predicted value and a true value; determining a leaf weight item, a number of leaf nodes, a leaf weight penalty regularization item, and a leaf tree penalty regularization item of a tree model; determining a regularization item based on the leaf weight item, the number of leaf nodes, the leaf weight penalty regularization item, and the leaf tree penalty regularization item of the tree model; constructing an objective function based on the loss function and the regularization item; and constructing a weak classifier based on the objective function.

[0047] Specifically, when constructing weak classifiers and building an integrated model in a multiple stacking manner, as a gradient boosting algorithm, by step-by-step forward stacking, each iteration generates a weak learner and forms the following objective function L of the t-th decision tree: (t) formula:

[0048]

[0049] In the above formula, is the regularization term, To reduce the risk of overfitting, is the loss function, y i is the predicted value, is the true value, f k is the tree model of the kth decision tree, T is the number of leaf nodes in the tree model, ω is the leaf weight term, γ is the leaf tree penalty regularization term, and λ is the leaf weight penalty regularization term to prevent model overfitting.

[0050] In one possible implementation, before constructing the objective function according to the loss function and the regularization term, the method also includes: determining anchored positive samples, positive samples and negative samples according to the eye movement data of each user in each user group; constructing a contrast loss term according to the first Euclidean distance from the anchored positive sample to the positive sample, the second Euclidean distance from the anchored positive sample to the negative sample and the interval parameter; constructing the objective function according to the loss function and the regularization term includes: constructing the objective function according to the loss function, the regularization term and the contrast loss term.

[0051] Specifically, when constructing weak classifiers and building an integrated model in a multiple stacking manner, as a gradient boosting algorithm, by step-by-step forward stacking, each iteration generates a weak learner and forms the following objective function L of the t-th decision tree: (t) formula:

[0052]

[0053] In the above formula, is the regularization term, To reduce the risk of overfitting, is the loss function, yi is the predicted value, is the true value, f k is the tree model of the kth decision tree, T is the number of leaf nodes in the tree model, ω is the leaf weight term, γ is the leaf tree penalty regularization term, and λ is the leaf weight penalty regularization term to prevent the model from overfitting. max(D p -D n )+margin,0) is the contrast loss term, D p is the first Euclidean distance from the anchored positive sample to the positive sample, D n is the second Euclidean distance from the anchored positive sample to the negative sample, and margin is the interval parameter.

[0054] The interval parameter may be determined in the following manner: corresponding interval parameters are set for each user group according to the type of each user group.

[0055] Specifically, during the model training process, due to the constraints of the loss function, the direction of the model gradient optimization will evolve in the direction of the probability distribution distance, and the accuracy of the model can be improved by dividing samples with similar features. In the embodiment of the present application, the loss function is optimized in combination with the characteristics of the eye movement data. Since the sample differences of the eye movement data are small, the metric learning method is used to improve the loss function of the prediction model. Construct a triple as training data, including [anchored positive sample, positive sample, negative sample]. During the training process, obtain the outputs corresponding to the anchored positive sample, positive sample, and negative sample, respectively, and then calculate the Euclidean distance D between the anchored positive sample and the positive sample, respectively. p and the Euclidean distance D between the anchored positive sample and the negative sample n , according to the situation of the five user groups, adjust and set different interval parameters margin in the prediction model. For example, through the prediction model effect, it is found that the model prediction effect is best when the five prediction models are [5, 4, 8, 5, 6], that is, by comparing the loss items and taking different interval parameters according to different types of user groups to improve the loss function, and finally by optimizing the sample distribution, the prediction probability distribution distance between positive samples and negative samples is maximized, and the samples with small training differences are optimized. At the same time, for the corresponding five different user groups, the loss function of the tree model is optimized, and the accuracy of the tree model is improved by dividing samples with similar features, which has a better prediction effect.

[0056] In one possible implementation, after constructing the objective function according to the loss function and the regularization term, the method further includes: defining model parameters according to the type of each user group, the model parameters including learning rate, minimum loss threshold and model depth, the minimum loss threshold being the maximum value of the learning rates of each user group; constructing a weak classifier based on the objective function includes: constructing a weak classifier based on the objective function and the model parameters.

[0057] Specifically, in terms of model parameter definition, we combine the features of eye movement data for precise optimization, and traverse and test each model parameter to obtain the parameters of the optimal model based on eye movement data. We adjust the weight of the model, introduce weight influencing features (eye movement data, classification of five user groups, and user familiarity with functions) into the features of eye movement data, and dynamically adjust the weight parameters to achieve fine control and avoid excessive reliance on a single feature.

[0058] In the embodiment of the present application, "gbtree" is used as a weak learner. Considering that setting a smaller learning rate can make more reference to weak learners to compensate for different residuals and effectively reduce overfitting, the learning rates "eta" of the five user groups of content storage, content viewing, content processing, content acquisition and sharing, and platform welfare are defined as [0.12, 0.10, 0.08, 0.18, 0.10] respectively for the classification models of the five user groups in combination with the data characteristics of different user groups for eye movement data. "Gamma" is the minimum threshold for the loss caused by node splitting. When the loss of node splitting is less than the minimum threshold for the loss, it will no longer split. The larger the "gamma", the more conservative the prediction model. Therefore, the "gamma" parameter of the prediction model of the five user groups is uniformly set to 0.1. "max_depth" is the depth of the prediction model. When the depth of the prediction model is not enough, the training data cannot be fully fitted. When the depth of the prediction model is large, it will lead to overfitting and reduce the generalization ability of the model. Therefore, through multiple experiments, the highest verification accuracy is selected for the five user groups, and "max_depth" [10, 8, 6, 12, 8] is used as a training parameter.

[0059] Through the technical solution disclosed in the embodiment of the present application, by obtaining the multi-dimensional data of the user, the multi-dimensional data includes the user's eye movement data and user behavior data for the target object, and the user behavior data includes the user's use of the function information of the target object; then the user is clustered according to the function information of the target object used by the user, and multiple user groups are obtained, and the functions used by users in different user groups are different; then the initial interest level prediction model is iteratively trained using the user behavior data and eye movement data corresponding to each user group to obtain the interest level prediction model, and the interest level prediction model is used to predict the target user's interest level in the target object. In this way, by refining the user group, and using the eye movement data and user behavior data of various user groups to train the prediction model, the prediction model is prevented from overfitting, and the generalization ability of the model is improved. When the prediction model is applied to the prediction of the user's interest level, the prediction accuracy can be effectively improved. And the user's eye movement data and user behavior data can reflect the user's interest changes in real time. When the prediction model predicts the interest level based on the user's multi-dimensional data, the predicted user interest level has high reliability and accuracy.

[0060] For example, Figure 2 As shown, the embodiment of the present application provides a method for predicting the user's interest level. The execution subject of the method may be a server. The method for predicting the user's interest level may specifically include the following steps:

[0061] Step S201 , obtaining target eye movement data, target behavior data of a target user for a target object, and a target user group to which the target user belongs.

[0062] Specifically, as described in the above embodiments, the target user group can be determined based on the user's target behavior data, which includes but is not limited to the user's use of the function of the target object used by the user and the behavior of operating the function of the target object. The user group to which the user belongs is determined based on the user behavior data, and the user group includes but is not limited to the user group for content storage, the user group for content viewing, the user group for content processing, the user group for content acquisition and sharing, and the user group for platform benefits.

[0063] Step S203, inputting the target eye movement data, the target user group and the target behavior data into the interest level prediction model.

[0064] The interest level prediction model is a prediction model trained by the interest level prediction model training method in the above embodiment.

[0065] Specifically, the prediction model obtained by training in the above embodiment inputs the target eye movement data, the target user group and the target behavior data into the interest level prediction model to predict the user's interest level.

[0066] Step S205: predicting the target user's interest level in the target object through an interest level prediction model.

[0067] Specifically, the target object refers to a certain product, and the user's interest in the target object can be the user's interest in the product. The interest level can be represented by a numerical value from 0 to 1. When the interest level is greater than a threshold, it means that the user is more interested in the product and can be recommended to the user. The threshold can be determined according to actual conditions, and the embodiments of the present application are not limited thereto.

[0068] Through the technical solution disclosed in the embodiments of the present application, the user's eye movement data, user behavior data and the user group to which the user belongs can reflect the user's interest changes in real time. When the prediction model predicts the interest level based on the user's eye movement data, user behavior data and the user group to which the user belongs, the predicted user interest level has high reliability and accuracy.

[0069] Corresponding to the above-mentioned training method of the interest level prediction model, the embodiment of the present application also provides a training device for the interest level prediction model. Figure 3 It is a structural schematic diagram of a training device 300 for an interest level prediction model provided in an embodiment of the present application, comprising: an acquisition module 301, used to acquire multi-dimensional data of users, the multi-dimensional data including eye movement data and user behavior data of users for a target object, the user behavior data including functional information of the target object used by the users; a clustering module 302, used to cluster users according to the functional information of the target object used by the users, and obtain multiple user groups, where users in different user groups use different functions; a training module 303, used to iteratively train an initial interest level prediction model using the user behavior data and eye movement data corresponding to each user group, and obtain an interest level prediction model, wherein the interest level prediction model is used to predict the interest level of the target user in the target object.

[0070] Through the technical solution disclosed in the embodiment of the present application, by obtaining the multi-dimensional data of the user, the multi-dimensional data includes the user's eye movement data and user behavior data for the target object, and the user behavior data includes the user's use of the function information of the target object; then the user is clustered according to the function information of the target object used by the user, and multiple user groups are obtained, and the functions used by users in different user groups are different; then the initial interest level prediction model is iteratively trained using the user behavior data and eye movement data corresponding to each user group to obtain the interest level prediction model, and the interest level prediction model is used to predict the target user's interest level in the target object. In this way, by refining the user group, and using the eye movement data and user behavior data of various user groups to train the prediction model, the prediction model is prevented from overfitting, and the generalization ability of the model is improved. When the prediction model is applied to the prediction of the user's interest level, the prediction accuracy can be effectively improved. And the user's eye movement data and user behavior data can reflect the user's interest changes in real time. When the prediction model predicts the interest level based on the user's multi-dimensional data, the predicted user interest level has high reliability and accuracy.

[0071] In one possible implementation, the training module 303 is also used to construct a weak classifier, and use the weak classifier as an initial interest level prediction model; the weak classifier is iteratively trained using the user behavior data and eye movement data corresponding to each user group until a predetermined number of iterations is reached to obtain an interest level prediction model; wherein each iterative training generates a new weak classifier, and the weak classifiers generated by multiple iterative trainings are superimposed to obtain an interest level prediction model.

[0072] In a possible implementation, a new weak classifier is constructed by the residual between the predicted value and the true value of the previous weak classifier.

[0073] In one possible implementation, the training module 303 is also used to construct a loss function, which indicates the difference between the predicted value and the true value; determine the leaf weight item, the number of leaf nodes, the leaf weight penalty regularization item and the leaf tree penalty regularization item of the tree model; determine the regularization item based on the leaf weight item, the number of leaf nodes, the leaf weight penalty regularization item and the leaf tree penalty regularization item of the tree model; construct an objective function based on the loss function and the regularization item; and construct a weak classifier based on the objective function.

[0074] In a possible implementation, the training module 303 is also used to determine anchored positive samples, positive samples and negative samples based on the eye movement data of each user in each user group; construct a contrast loss term based on the first Euclidean distance from the anchored positive sample to the positive sample, the second Euclidean distance from the anchored positive sample to the negative sample and the interval parameter; constructing an objective function based on the loss function and the regularization term includes: constructing an objective function based on the loss function, the regularization term and the contrast loss term.

[0075] In a possible implementation, the training module 303 is further configured to set corresponding interval parameters for each user group according to the type of each user group.

[0076] In a possible implementation, the training module 303 is also used to define model parameters according to the type of each user group, the model parameters including learning rate, minimum loss threshold and model depth, the minimum loss threshold is the maximum value of the learning rate of each user group; a weak classifier is constructed based on the objective function and the model parameters.

[0077] In a possible implementation, the multi-dimensional data also includes: subjective evaluation data of the user on the target object; the training module 303 is further used to optimize the eye movement data using the subjective evaluation data.

[0078] In a possible implementation, it also includes: a cleaning module, which is used to clean the multi-dimensional data.

[0079] Corresponding to the above-mentioned method for predicting the user's interest level, the embodiment of the present application also provides a device for predicting the user's interest level. Figure 4 It is a structural schematic diagram of a user interest level prediction device 400 provided in an embodiment of the present application, including: an acquisition module 401, used to obtain the target eye movement data, target behavior data of the target user for the target object and the target user group to which the target user belongs; a prediction module 402, used to input the target eye movement data, the target user group and the target behavior data into an interest level prediction model, the interest level prediction model is trained by the interest level prediction model training method mentioned in the above embodiment; the interest level prediction model is used to predict the target user's interest level in the target object.

[0080] Through the technical solution disclosed in the embodiments of the present application, the user's eye movement data, user behavior data and the user group to which the user belongs can reflect the user's interest changes in real time. When the prediction model predicts the interest level based on the user's eye movement data, user behavior data and the user group to which the user belongs, the predicted user interest level has high reliability and accuracy.

[0081] Corresponding to the method provided in the above embodiment, based on the same technical concept, the embodiment of the present application further provides an electronic device, which is used to execute the above method. Figure 5 A schematic diagram of the structure of an electronic device for implementing each embodiment of the present application is shown in FIG. Figure 5 As shown. The electronic device may have relatively large differences due to different configurations or performances, and may include one or more processors 501 and memory 502, and the memory 502 may store one or more storage applications or data. Among them, the memory 502 may be a temporary storage or a permanent storage. The application stored in the memory 502 may include one or more modules (not shown in the figure), and each module may include a series of computer executable instructions in the electronic device.

[0082] Furthermore, the processor 501 may be configured to communicate with the memory 502 to execute a series of computer executable instructions in the memory 502 on the electronic device. The electronic device may also include one or more power supplies 503, one or more wired or wireless network interfaces 504, one or more input and output interfaces 505, and one or more keyboards 506.

[0083] Specifically in this embodiment, the electronic device includes a processor, a communication interface, a memory and a communication bus; wherein the processor, the communication interface and the memory communicate with each other through the bus; the memory is used to store computer programs; the processor is used to execute the programs stored in the memory to achieve the above Figure 1 or Figure 2 The various steps in the method embodiment are similar to those in the method embodiment, and have the beneficial effects of the above method embodiments. To avoid repetition, the embodiments of the present application will not be described again here.

[0084] This embodiment also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above Figure 1 or Figure 2 Each step in the method embodiment has the beneficial effects of the above method embodiments. To avoid repetition, the embodiments of the present application will not be described again here.

[0085] The present application provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the above Figure 1 or Figure 2 Each step in the method embodiment has the beneficial effects of the above method embodiments. To avoid repetition, the embodiments of the present application will not be described again here.

[0086] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, devices, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0087] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0088] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0090] In a typical configuration, an electronic device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0091] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0092] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0093] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

Claims

1. A method for training an interest level prediction model, characterized in that: include: Acquire multi-dimensional data of the user, wherein the multi-dimensional data includes eye movement data of the user with respect to the target object and user behavior data, wherein the user behavior data includes functional information of the user using the target object; Clustering users according to the functional information of the target objects used by them to obtain multiple user groups, where users in different user groups use different functions; The initial interest level prediction model is iteratively trained using the user behavior data and the eye movement data corresponding to each of the user groups to obtain an interest level prediction model, wherein the interest level prediction model is used to predict the interest level of the target user in the target object.

2. The method for training the interest level prediction model according to claim 1, characterized in that: The iterative training of the initial interest level prediction model using the user behavior data and the eye movement data corresponding to each of the user groups includes: Constructing a weak classifier, and using the weak classifier as the initial interest level prediction model; Iteratively training the weak classifier using the user behavior data and the eye movement data corresponding to each of the user groups until a predetermined number of iterations is reached to obtain the interest level prediction model; Each iterative training generates a new weak classifier, and the weak classifiers generated by multiple iterative trainings are superimposed to obtain the interest level prediction model.

3. The method for training the interest level prediction model according to claim 2, characterized in that: The new weak classifier is constructed by the residual between the predicted value and the true value of the previous weak classifier.

4. The method for training the interest level prediction model according to claim 2, characterized in that: The construction of a weak classifier comprises: constructing a loss function indicating the difference between the predicted value and the true value; Determine the leaf weight term, the number of leaf nodes, the leaf weight penalty regularization term, and the leaf tree penalty regularization term of the tree model; Determining a regularization term according to a leaf weight term, a number of leaf nodes, a leaf weight penalty regularization term, and a leaf tree penalty regularization term of the tree model; Constructing an objective function according to the loss function and the regularization term; The weak classifier is constructed based on the objective function.

5. The method for training the interest level prediction model according to claim 4, characterized in that: Before constructing the objective function according to the loss function and the regularization term, the method further includes: Determine anchor positive samples, positive samples and negative samples according to the eye movement data of each user in each of the user groups; Constructing a contrast loss term according to a first Euclidean distance from the anchored positive sample to the positive sample, a second Euclidean distance from the anchored positive sample to the negative sample, and an interval parameter; The constructing the objective function according to the loss function and the regularization term comprises: The objective function is constructed according to the loss function, the regularization term and the contrastive loss term.

6. The method for training the interest level prediction model according to claim 5, characterized in that: Before constructing the contrast loss term according to the first Euclidean distance from the anchored positive sample to the positive sample, the second Euclidean distance from the anchored positive sample to the negative sample, and the interval parameter, the method further includes: Corresponding interval parameters are set for each user group according to the type of each user group.

7. The method for training the interest level prediction model according to claim 4, characterized in that: After constructing the objective function according to the loss function and the regularization term, the method further includes: Defining model parameters according to the type of each user group, the model parameters including learning rate, minimum loss threshold and model depth, the minimum loss threshold being the maximum value among the learning rates of each user group; The constructing the weak classifier based on the objective function comprises: The weak classifier is constructed based on the objective function and the model parameters.

8. The method for training the interest level prediction model according to claim 1, characterized in that: The multi-dimensional data also includes: subjective evaluation data of the user on the target object; Before iteratively training the initial interest level prediction model using the user behavior data and the eye movement data corresponding to each of the user groups, the method further includes: The eye movement data is optimized using the subjective evaluation data.

9. The method for training the interest level prediction model according to claim 1, characterized in that: After obtaining the multi-dimensional data of the user, the method further includes: The multi-dimensional data is cleaned.

10. A method for predicting user interest level, characterized in that: include: Obtaining target eye movement data, target behavior data of a target user for a target object, and a target user group to which the target user belongs; Inputting the target eye movement data, the target user group and the target behavior data into an interest level prediction model, wherein the interest level prediction model is trained by the interest level prediction model training method according to any one of claims 1 to 9; The interest level of the target user in the target object is predicted by the interest level prediction model.

Images