CN117112870A

CN117112870A - User interaction image classification method and AI session interaction system based on artificial intelligence

Info

Publication number: CN117112870A
Application number: CN202311169420.8A
Authority: CN
Inventors: 帅丽兰; 华龙杰
Original assignee: Ningxia Zhiboyuan Education Technology Co ltd
Current assignee: Zhou Xianghong
Priority date: 2023-09-12
Filing date: 2023-09-12
Publication date: 2023-11-24
Anticipated expiration: 2043-09-12
Also published as: CN117112870B

Abstract

The embodiment of the application provides a user interaction image classification method and an AI conversation interaction system based on artificial intelligence, which can firstly perform user preference data positioning on template user conversation interaction behavior flow data based on a user preference data positioning model, and load undetermined user interest point data obtained in the user preference data positioning process into a user preference range observation model, so that the user preference range observation model can generate estimated interest point distribution data based on undetermined user interest point data. The user preference data positioning model can focus on the user preference range information of the learning target session interaction behavior data, so that the observation capability of using the user preference data positioning model with the user preference range can be deployed, the user preference data positioning accuracy can be improved by combining the user preference range of the session interaction behavior data when the user preference data positioning is carried out, and the accuracy of the subsequent user interaction image classification can be improved.

Description

User interaction image classification method and AI session interaction system based on artificial intelligence

Technical Field

The embodiment of the application relates to the technical field of artificial intelligence, in particular to a user interaction image classification method and an AI session interaction system based on artificial intelligence.

Background

With the development of internet information technology and network communication technology, the number of users of various network online service applications is rapidly increased, and the users are effectively analyzed by user interaction images so as to push network service contents matched with the user interaction images for the users, thereby improving the service experience of the network online service applications. In the process of classifying the user interaction images, in order to improve the classification efficiency, in the related art, user preference data needs to be positioned from user session interaction behavior flow data and then classified. However, the existing user preference data positioning model does not have a user preference range sensing capability, and cannot perform positioning of the user preference data by sensing the behavior data user preference range, which may result in lower accuracy of the positioning result of the user preference data. Therefore, how to train the user preference data positioning model with the user preference range perception capability so as to ensure the accuracy of the user preference data positioning, and further facilitate the subsequent classification of the user interactive portraits of the target user is a technical problem to be solved in the technical field.

Disclosure of Invention

In order to at least overcome the above-mentioned shortcomings in the prior art, an object of an embodiment of the present application is to provide a user interaction image classification method and an AI session interaction system based on artificial intelligence.

According to one aspect of the embodiment of the application, there is provided an artificial intelligence-based user interaction image classification method, comprising:

the method comprises the steps of obtaining template user session interaction behavior flow data and template user preference data, wherein the template user session interaction behavior flow data comprises first session interaction behavior data in a directed graph format and second session interaction behavior data in a directed graph format, and the user preference range of the first session interaction behavior data is different from the user preference range of the second session interaction behavior data; the template user preference data is generated by calibrating preference data of target session interaction behavior data in the template user session interaction behavior flow data, and the target session interaction behavior data comprises at least one of the first session interaction behavior data and the second session interaction behavior data;

user preference data positioning is carried out on the template user session interaction behavior flow data based on a user preference data positioning model, and the data of the undetermined user interest points analyzed in the user preference data positioning process are loaded into a user preference range observation model;

Generating estimated interest point distribution data by combining the undetermined user interest point data based on the user preference range observation model, wherein the estimated interest point distribution data characterizes an estimated user preference range of the target session interaction behavior data;

generating reference interest point distribution data based on the user preference range observation model in combination with the observation interest points of the target session interaction behavior data in the template user preference data; the reference interest point distribution data characterizes a reference user preference range of the target session interaction behavior data;

updating weight information of the user preference data positioning model based on distinguishing distribution data between the estimated interest point distribution data and the reference interest point distribution data;

the method comprises the steps that user preference data positioning is conducted on target user session interaction behavior flow data of any target user based on a user preference data positioning model updated by weight information, user preference data of the target user are generated, user interaction portrait classification is conducted on the user preference data, and user interaction portrait information of the target user is generated; the target user session interaction behavior flow data comprises at least two session interaction behavior data which are in a directed graph format and have different user preference ranges.

In an implementation manner of the first aspect, the pending user interest point data includes: the user preference data positioning model comprises a plurality of K coding-level conversation interaction behavior vectors generated by K encoders, wherein one encoder generates a coding-level conversation interaction behavior vector, and the data volume of the conversation interaction behavior vectors of any two coding levels is different;

the user preference range observation model comprises a point of interest estimation network and K feature quantity optimization networks, wherein the network loading positions of the feature quantity optimization networks are respectively connected with different encoders in the K encoders, and the network output positions of the feature quantity optimization networks are respectively connected with the network loading positions of the point of interest estimation network;

the generating estimated interest point distribution data based on the user preference range observation model and the undetermined user interest point data comprises the following steps:

performing data quantity optimization on the session interaction behavior vectors of each coding level based on each characteristic quantity optimization network to generate optimized K session interaction behavior vectors; a feature quantity optimizing network optimizes the session interaction behavior vectors of a coding level, wherein the optimized data quantity of the K session interaction behavior vectors is the same;

Converging the optimized K session interaction behavior vectors to generate a converging behavior vector;

and observing the interest point distribution data based on the converging behavior vector on the basis of the interest point estimation network, and generating estimated interest point distribution data.

In an implementation of the first aspect, the template user preference data includes one or more of the following: first user preference data, second user preference data, and third user preference data; the first user preference data is generated by calibrating preference data of the first session interaction behavior data in the template user session interaction behavior traffic data, the second user preference data is generated by calibrating preference data of the second session interaction behavior data in the template user session interaction behavior traffic data, and the third user preference data is generated by calibrating preference data of the first session interaction behavior data and the second session interaction behavior data in the template user session interaction behavior traffic data;

the estimated point of interest distribution data includes one or more of the following: first estimated point of interest distribution data for the first session interaction behavior data, second estimated point of interest distribution data for the second session interaction behavior data, and third estimated point of interest distribution data for the first session interaction behavior data and the second session interaction behavior data;

The reference point of interest distribution data includes one or more of the following: first reference point of interest distribution data for the first session interaction behavior data, second reference point of interest distribution data for the second session interaction behavior data, and third reference point of interest distribution data for the first session interaction behavior data and the second session interaction behavior data.

In an implementation of the first aspect, if the target session interaction behavior data is the first session interaction behavior data or the second session interaction behavior data, the template user preference data comprises the first user preference data or the second user preference data, and the reference point of interest distribution data comprises the first reference point of interest distribution data or the second reference point of interest distribution data;

the generating reference interest point distribution data based on the user preference range observation model and the observation interest points of the target session interaction behavior data in the template user preference data comprises the following steps:

determining observation interest points of the target session interaction behavior data in the template user preference data based on the user preference range observation model, and determining a plurality of target user preference knowledge points belonging to the target session interaction behavior data in the template user preference data;

Based on trigger matching degree between the behavior trigger nodes of the preference knowledge points of each target user and the observation interest points of the interaction behavior data of the target session, respectively optimizing preference characteristic values of the preference knowledge points of each target user, and generating the reference interest point distribution data; and the reference preference characteristic value of any target user preference knowledge point in the reference interest point distribution data is positively correlated with the trigger matching degree of any target user preference knowledge point.

In an implementation of the first aspect, if the target session interaction behavior data comprises the first session interaction behavior data and the second session interaction behavior data, the template user preference data comprises the third user preference data, the reference point of interest distribution data comprises the third reference point of interest distribution data;

determining an observation interest point of the first session interaction behavior data and an observation interest point of the second session interaction behavior data in the third user preference data based on the user preference range observation model;

Determining a plurality of first user preference knowledge points belonging to the first session interaction behavior data and a plurality of second user preference knowledge points belonging to the second session interaction behavior data in the third user preference data;

based on trigger matching degree between the behavior trigger nodes of the first user preference knowledge points and the observation interest points of the first session interaction behavior data, respectively optimizing preference characteristic values of the first user preference knowledge points;

based on trigger matching degree between the behavior trigger nodes of the second user preference knowledge points and the observation interest points of the second session interaction behavior data, respectively optimizing preference characteristic values of the second user preference knowledge points, and generating the reference interest point distribution data;

the method comprises the steps that a reference preference characteristic value of any first user preference knowledge point in the reference interest point distribution data is positively correlated with the trigger matching degree of any first user preference knowledge point, and a reference preference characteristic value of any second user preference knowledge point in the reference interest point distribution data is positively correlated with the trigger matching degree of any second user preference knowledge point.

In an implementation manner of the first aspect, the updating the weight information of the user preference data positioning model based on the distinguishing distribution data between the estimated interest point distribution data and the reference interest point distribution data includes:

performing weight parameter optimization error calculation based on the estimated interest point distribution data and the reference interest point distribution data based on a cross entropy cost calculation function, and generating a first user preference observation error value, wherein the first user preference observation error value characterizes distinguishing distribution data between the estimated interest point distribution data and the reference interest point distribution data;

calculating a second user preference observation error value based on the first user preference observation error value;

and optimizing the weight information of the user preference data positioning model by taking the minimization of the second user preference observation error value as a training target.

In an implementation manner of the first aspect, the calculating a weight parameter optimization error based on the estimated interest point distribution data and the reference interest point distribution data by the cross entropy cost calculation function, generating a first user preference observation error value includes:

Determining a plurality of target user preference knowledge points belonging to the target session interaction behavior data from the template user preference data;

acquiring estimated preference characteristic values of each target user preference knowledge point in the estimated interest point distribution data and reference preference characteristic values of each target user preference knowledge point in the reference interest point distribution data;

and respectively calculating characteristic training error values between the reference preference characteristic values and the corresponding estimated preference characteristic values of the target user preference knowledge points, and carrying out weight fusion calculation on the calculated characteristic training error values to generate a first user preference observation error value.

In an implementation manner of the first aspect, the calculating a second user preference observation error value based on the first user preference observation error value includes:

based on the user preference range observation model and combining M feature expansion parameters, carrying out feature expansion on the template user preference data to generate M feature expanded user preference data;

calculating to obtain M candidate interest point distribution data with expanded characteristics corresponding to the user preference data with expanded characteristics;

estimating and generating M estimated interest point distribution data with expanded characteristics based on the user preference range observation model and the undetermined user interest point data; the estimated interest point distribution data after the feature expansion corresponds to a feature expansion knowledge base;

Performing weight parameter optimization error calculation based on the estimated interest point distribution data after feature expansion and the candidate interest point distribution data after feature expansion corresponding to each feature expansion parameter by the cross entropy cost calculation function, and generating M feature expansion user preference observation error values;

and carrying out weight fusion calculation on the first user preference observation error value and the M characteristic extension user preference observation error values to generate a second user preference observation error value.

For example, in an implementation manner of the first aspect, the optimizing the weight information of the user preference data positioning model with the minimizing of the second user preference observation error value as the training target includes:

obtaining a significant training error value, the significant training error value comprising at least a first training error value; the first training error value characterizes a distinguishing distribution data between a significant correlation of the target session interaction behavior data in the template user preference data and a significant correlation of the target session interaction behavior data in the estimated user preference data; the estimated user preference data is obtained by the user preference data positioning model for positioning the user preference data of the template user session interaction behavior flow data;

Calculating a target training error value for the user preference data positioning model based on the second user preference observation error value and the significant training error value;

minimizing the target training error value as a training target, and optimizing weight information of the user preference data positioning model;

for example, the obtaining significant training error values includes:

generating a first behavior traffic data cluster to be learned based on the template user session interaction behavior traffic data and the template user preference data, and generating a second behavior traffic data cluster to be learned based on the template user session interaction behavior traffic data and the estimated user preference data,

based on the behavior attention characteristics in the first to-be-learned behavior flow data cluster and the behavior attention characteristics in the second to-be-learned behavior flow data cluster, the attention mechanism unit performs significant relevance ranking on the template user preference data and the estimated user preference data, and generates significant relevance ranking information;

calculating the first training error value based on a significance assessment index and the significance relevance ranking information; the saliency assessment index characterizes the saliency of the target session interaction behavior data in the template user preference data to be greater than the saliency of the target session interaction behavior data in the estimated user preference data;

The attention-based mechanism unit performs significant relevance ranking on the template user preference data and the estimated user preference data based on the behavior attention feature in the first behavior traffic data cluster to be learned and the behavior attention feature in the second behavior traffic data cluster to be learned, and generates significant relevance ranking information, including:

performing time sequence mode feature scrambling on behavior attention features in the template user preference data to generate scrambled user preference data; and generating a third behavior traffic data cluster to be learned based on the template user session interaction behavior traffic data and the cluttered user preference data,

based on the behavior attention characteristics in the first to-be-learned behavior traffic data cluster, the behavior attention characteristics in the second to-be-learned behavior traffic data cluster and the behavior attention characteristics in the third to-be-learned behavior traffic data cluster, the attention mechanism unit performs significant relevance ranking on the template user preference data, the estimated user preference data and the disorder user preference data, and generates significant relevance ranking information;

the saliency assessment index characterizes the saliency of the target session interaction behavior data in the template user preference data, and is larger than the saliency of the target session interaction behavior data in the estimated user preference data; and the significant correlation of the target session interaction behavior data in the estimated user preference data is greater than the significant correlation of the target session interaction behavior data in the cluttered user preference data;

The salient training error values further comprise second training error values characterizing distinguishing distribution data between salient preference features in the template user preference data and salient preference features of the estimated user preference data; the obtaining the significant training error value further includes:

performing feature extraction on the template user preference data based on an attention mechanism unit, and generating salient preference features in the template user preference data;

performing feature extraction on the estimated user preference data based on the attention mechanism unit to generate salient preference features in the estimated user preference data;

the second training error value is calculated based on the salient preference features in the template user preference data and the salient preference features in the estimated user preference data.

In accordance with one aspect of an embodiment of the present application, there is provided an AI session interaction system comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement an artificial intelligence-based user interaction image classification method in any of the foregoing possible implementations.

In accordance with one aspect of embodiments of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The computer instructions are read from a computer-readable storage medium by a processor of a computer device, which executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations of the three aspects described above.

In the technical schemes provided by some embodiments of the present application, user preference data positioning may be performed on template user session interaction behavior traffic data based on a user preference data positioning model, and undetermined user interest point data obtained in the user preference data positioning process is loaded into a user preference range observation model, so that the user preference range observation model may generate estimated interest point distribution data based on undetermined user interest point data. Because the template user preference data is generated by calibrating preference data of target session interaction behavior data in the template user session interaction behavior flow data, the user preference range observation model can also generate reference interest point distribution data which can be used as a training learning target based on observation interest points of the target session interaction behavior data in the template user preference data. Since the estimated point of interest distribution data may characterize an estimated user preference range for the target session interaction behavior data, while the reference point of interest distribution data may characterize a reference user preference range (i.e., an actual user preference range) for the target session interaction behavior data; therefore, the difference distribution data between the estimated interest point distribution data and the reference interest point distribution data can express the difference distribution data between the estimated user preference range and the actual user preference range of the target session interaction behavior data, so that when the weight information of the user preference data positioning model is updated according to the difference distribution data, the user preference data positioning model can focus and learn the user preference range information of the target session interaction behavior data, the user preference data positioning model can be deployed to have the observation capability of the user preference range, and the user preference data positioning accuracy can be improved by combining the user preference range of the session interaction behavior data when the user preference data positioning is carried out, so that the accuracy of the classification of the subsequent user interaction images can be improved.

Drawings

For a clearer description of the technical solutions of the embodiments of the present application, reference will be made to the accompanying drawings, which are needed to be activated in the embodiments, and it should be understood that the following drawings only illustrate some embodiments of the present application, and therefore should not be considered as limiting the scope, and other related drawings can be extracted by those skilled in the art without the inventive effort.

FIG. 1 is a flow chart of a user interaction image classification method based on artificial intelligence according to an embodiment of the application;

fig. 2 is a schematic block diagram of an AI session interaction system for implementing the above-mentioned artificial intelligence-based user interaction image classification method according to an embodiment of the present application.

Detailed Description

The following description is presented to enable one of ordinary skill in the art to make and use the application and is provided in the context of a particular application and its requirements. It will be apparent to those having ordinary skill in the art that various changes can be made to the disclosed embodiments and that the general principles defined herein may be applied to other embodiments and applications without departing from the principles and scope of the application. Therefore, the present application is not limited to the described embodiments, but is to be accorded the widest scope consistent with the claims.

FIG. 1 is a flow chart of an artificial intelligence based user interaction image classification method according to an embodiment of the present application, and the detailed description of the artificial intelligence based user interaction image classification method is provided below.

Step S101, template user session interaction behavior flow data and template user preference data are obtained.

The template user session interaction behavior flow data is used for updating weight information of a user preference data positioning model; the template user session interaction behavior flow data comprises first session interaction behavior data in a directed graph format and second session interaction behavior data in a directed graph format, and the user preference range of the first session interaction behavior data is different from the user preference range of the second session interaction behavior data. For example, the template user session interaction behavior traffic data may be user live session interaction behavior traffic data, or shopping session interaction behavior traffic data; then, the first session interaction behavior data and the second session interaction behavior data may be dialogue behavior data and sharing behavior data, respectively, having different ranges of user preferences.

The template user preference data is generated by preference data calibration of target session interaction behavior data in the template user session interaction behavior traffic data, where the target session interaction behavior data may include one or more of the following: the first session interaction behavior data and the second session interaction behavior data; accordingly, the template user preference data may include one or more of the following: first user preference data, second user preference data, and third user preference data. Illustratively, when the target session interaction behavior data includes only the first session interaction behavior data, the template user preference data includes the first user preference data; when the target session interaction behavior data only includes second session interaction behavior data, the template user preference data includes second user preference data; when the target session interaction behavior data comprises first session interaction behavior data and second session interaction behavior data, the template user preference data may comprise third user preference data, in which case the template user preference data may also comprise first user preference data and second user preference data, for example. The first user preference data is generated by calibrating preference data of first session interaction behavior data in the template user session interaction behavior flow data, namely the first user preference data only comprises the first session interaction behavior data; the second user preference data is generated by calibrating preference data of second session interaction behavior data in the template user session interaction behavior flow data, namely the second user preference data only comprises the second session interaction behavior data; the third user preference data is generated by calibrating preference data of the first session interaction behavior data and the second session interaction behavior data in the template user session interaction behavior flow data, namely the third user preference data comprises the first session interaction behavior data and the second session interaction behavior data.

By way of example, assume two types of template user session interaction behavior data are collected: one is a user who likes to purchase home products, and the other is a user who prefers to purchase sports equipment. Such data may include keywords for the user to search for merchandise, records for browsing and purchasing merchandise, etc., as directed graphs.

For example, the conversational interaction behavior data of household product preference users may show that they often browse from "bed sheets" to "pillows" to "carpets"; and sports equipment preference users may often browse from "yoga mats" to "dumbbells" to "jump ropes".

Template user preference data is then generated using these data. For example, if a user frequently browses and purchases household products, he may be marked as having a preference for household products.

The user preference range generally refers to the degree of interest of a user in a certain product, service or content within a certain time range, and can also be regarded as interest distribution of the user. Such preferences may cover various aspects of merchandise category, brand, price space, style, color, size, etc.

The following are several specific examples:

1. On an e-commerce website, the user preference range may include preferences for a certain class of goods (e.g., household items, electronic devices, clothing, etc.), or preferences for a certain brand, even for goods in a certain price range.

2. On a music streaming platform, the user's range of preferences may include preferences for a certain music type (e.g., popular, classical, rock, etc.), or preferences for a certain artist.

3. On news or social media platforms, the user preferences may include preferences for a certain topic (e.g., sports, entertainment, etc.), or preferences for a certain content format (e.g., video, articles, pictures, etc.).

4. On a travel reservation platform, the user's range of preferences may include preferences for a certain travel destination or preferences for a certain type of activity (e.g., hiking, surfing, bird watching, etc.).

The above are specific examples of the range of user preferences that may help understand and predict the needs of the user and provide more accurate recommendation and personalization services.

Step S102, user preference data positioning is carried out on the template user session interaction behavior flow data based on the user preference data positioning model, and the undetermined user interest point data analyzed in the user preference data positioning process is loaded into the user preference range observation model.

In an alternative embodiment, the implementation step of step S102 may be: and carrying out feature extraction on the template user session interaction behavior flow data based on the user preference data positioning model, and generating session interaction behavior extraction features of each user preference knowledge point in the template user session interaction behavior flow data. Secondly, the association degree information of the session interaction behavior data of each user preference knowledge point can be determined based on the session interaction behavior extraction characteristics of each user preference knowledge point; correlation information characterization of session interaction behavior data of any user preference knowledge point: the confidence that any user preference knowledge point belongs to the first session interaction behavior data and the confidence that any user preference knowledge point belongs to the second session interaction behavior data. Then, based on the association degree information of the session interaction behavior data of each user preference knowledge point, a plurality of first user preference knowledge points belonging to the first session interaction behavior data and a plurality of second user preference knowledge points belonging to the second session interaction behavior data can be determined from the template user session interaction behavior flow data. Finally, the first user preference knowledge points and/or the second user preference knowledge points can be specifically marked in the template user session interaction behavior flow data, so that the user preference data can be positioned from the template user session interaction behavior flow data by the first session interaction behavior data and/or the second session interaction behavior data, and estimated user preference data can be generated. For example, assume that the template user session interaction behavior flow data includes two categories of users, one category being users who like to purchase household products and the other category being users who prefer to purchase sports equipment. Such data may include keywords for the user to search for merchandise, records for browsing and purchasing merchandise, and the like. Firstly, the user preference data positioning model is utilized to conduct characteristic extraction on the template user session interaction behavior flow data, for example, the browsing and purchasing frequency of users on certain commodities in different time periods is extracted. Relevance information is then determined for session interaction behavior data for each user preference knowledge point (e.g., specific to a certain type of household product or piece of athletic equipment) based on the features, that is, a degree of interest of the user in the knowledge point is determined. For example, if a user frequently browses and purchases yoga mats at night, the user may be interested in yoga. Next, according to the association degree information of the session interaction behavior data of each user preference knowledge point, a plurality of first user preference knowledge points (e.g., bedsheets, pillows, carpets) belonging to the first session interaction behavior data (e.g., home product related data) and a plurality of second user preference knowledge points (e.g., yoga mats, dumbbells, jump ropes) belonging to the second session interaction behavior data (e.g., sports equipment related data) may be determined from the template user session interaction behavior flow data. And finally, carrying out specific labeling on the first user preference knowledge points and/or the second user preference knowledge points in the template user session interaction behavior flow data, separating the first user preference knowledge points and/or the second user preference knowledge points from the whole session interaction behavior flow data, and generating estimated user preference data. For example, by labeling, it may be possible to ascertain which users are more inclined to purchase household products and which users are more inclined to purchase sports equipment, and to further precisely push related goods to the target users.

The user preference data location model may be a long and short term memory network model, a decision tree network model, a random forest tree network model, an unbalanced data classification model of the base Yu Mengte Carlo neural network.

In the description of the above embodiments, estimating user preference data may each include one or more of the following: the first estimated user session interaction behavior traffic data, the second estimated user session interaction behavior traffic data, and the third estimated user session interaction behavior traffic data. Wherein, the first estimated user session interaction behavior traffic data refers to: positioning the generated user preference data only from the template user session interaction behavior flow data based on the user preference data positioning model; the second estimated user session interaction behavior traffic data refers to: positioning the generated user preference data only from the template user session interaction behavior flow data by using the second session interaction behavior data based on the user preference data positioning model; the third estimated user session interaction behavior traffic data refers to: the first session interaction behavior data and the second session interaction behavior data are located from the template user session interaction behavior traffic data based on the user preference data location model. In the description of the above embodiments, the user preference data positioning model may extract the to-be-determined user interest point data in the user preference data positioning process. The user preference data positioning model can load undetermined user interest point data analyzed in the user preference data positioning process into the user preference range observation model, so that the user preference range observation model can estimate the user preference range based on the undetermined user interest point data on the first session interaction behavior data and/or the second session interaction behavior data in the template user session interaction behavior traffic data. The above user preference data positioning model may be a user preference data positioning model updated by the weight information multiple times, or may be a user preference data positioning model not updated by the weight information.

For example, still taking the foregoing example as an example, the first estimated user session interaction behavior traffic data may include user preference data related only to purchasing household products, e.g., a user often searches for "modern conclusive style" household products. The second estimated user session interaction flow data may then include user preference data associated with purchasing only athletic equipment, e.g., a user frequently browsing and purchasing treadmills. And the third estimated user session interaction flow data may include user preference data associated with purchasing both household products and sports equipment, such as a user having both browsed "modern conciseness style" household products and searched for "aerobic sports equipment.

In this process, the pending user point of interest data can also be parsed, for example, a user may be interested in "extremely abbreviated design" furniture or "intelligent exercise equipment". These pending user point of interest data are loaded into a user preference range observation model so that the model can estimate the user's likely preference range based on these data.

Based on the above, the observation of the user preference range is performed according to the to-be-determined user interest point data based on the user preference range observation model, so that the weight information of the user preference data positioning model is updated based on the observation result, and the updated user preference data positioning model is used for classifying the user interaction images after the user preference data positioning, and specifically, the following steps S103-S106 can be referred to.

Step S103, generating estimated interest point distribution data based on the pending user interest point data based on the user preference range observation model.

The pending user point of interest data may include: any encoder in the user preference data positioning model performs feature extraction on the template user session interaction behavior flow data to obtain a session interaction behavior vector with user interest features; alternatively, the pending user point of interest data may comprise session interaction behavior vectors for K encoding levels generated by K encoders in the user preference data positioning model, K being a positive integer greater than 1. In addition, as the number of session interaction behavior vectors included in the pending user interest point data is different, different distribution data exists in the architecture of the user preference range observation model, so that the manner in which the user preference range observation model generates estimated interest point distribution data based on the pending user interest point data is also different.

In an alternative embodiment, when pending user point of interest data comprises: when any encoder in the user preference data positioning model encodes the conversation interaction behavior flow data of the template user, a conversation interaction behavior vector is generated; a point of interest estimation network may be included in the user preference range observation model. Thus, step S103 may specifically include: the interest point estimation network in the observation model based on the user preference range can directly observe the interest point distribution data based on the interest point data of the undetermined user, and estimated interest point distribution data is generated.

In an alternative embodiment, the user point of interest data comprises: when the user prefers the session interaction behavior vectors of K coding grades generated by K encoders in the data positioning model; from the foregoing, the user preference range observation model may include a point of interest estimation network and K feature quantity optimization networks, where network loading positions of each feature quantity optimization network are respectively connected to different encoders in the K encoders, and network output positions of each feature quantity optimization network are all connected to network loading positions of the point of interest estimation network. Thus, step S103 may specifically include:

firstly, carrying out data quantity optimization on session interaction behavior vectors of all coding grades based on all feature quantity optimization networks to generate optimized K session interaction behavior vectors; a feature quantity optimizing network optimizes the conversation interaction behavior vector of a coding level, and the data quantity of the optimized K conversation interaction behavior vectors is the same. For example, a data amount of preset target user session interaction behavior flow data may be obtained, where the data amount of the target user session interaction behavior flow data is greater than the data amount of the session interaction behavior vector of each coding level; and then, according to the data quantity of the target user session interaction behavior flow data, respectively carrying out feature quantity optimization on the session interaction behavior vectors of all the coding grades based on each feature quantity optimization network so as to adjust the data quantity of the session interaction behavior vectors of all the coding grades to the data quantity of the target user session interaction behavior flow data, thereby obtaining optimized K session interaction behavior vectors. Or, any session interaction behavior vector can be selected from the K session interaction behavior vectors to be used as a reference session interaction behavior vector; and according to the data quantity of the reference session interaction behavior vector, respectively optimizing the characteristic quantity of other session interaction behavior vectors based on the characteristic quantity optimizing network corresponding to the other session interaction behavior vectors except the reference session interaction behavior vector so as to optimize the data quantity of the other session interaction behavior vectors to the data quantity of the reference session interaction behavior vector, thereby generating K optimized session interaction behavior vectors.

And secondly, the optimized K session interaction behavior vectors can be aggregated to generate an aggregated behavior vector. For example, the optimized session interaction behavior vectors can be sequentially aggregated according to the number of modes of the optimized K session interaction behavior vectors, so as to generate an aggregated behavior vector. It can be seen that the number of modes of the aggregate behavior vector is equal to the sum of the number of modes of the optimized K session interaction behavior vectors.

Then, the point of interest estimation network may observe the point of interest distribution data based on the aggregate behavior vector, generating estimated point of interest distribution data.

The data for the interaction behavior of the target session may include one or more of the following: the first session interaction behavior data and the second session interaction behavior data; thus, estimating the point of interest distribution data may include one or more of the following: the first estimated point of interest distribution data with respect to the first session interaction behavior data, the second estimated point of interest distribution data with respect to the second session interaction behavior data, and the third estimated point of interest distribution data with respect to the first session interaction behavior data and the second session interaction behavior data. In other words, the first estimated point of interest distribution data may characterize an estimated user preference range of the first session interaction behavior data, the second estimated point of interest distribution data may characterize an estimated user preference range of the second session interaction behavior data, and the third estimated point of interest distribution data may be used to reflect both the estimated user preference range of the first session interaction behavior data and the estimated user preference range of the second session interaction behavior data. Illustratively, when the target session interaction behavior data includes only the first session interaction behavior data, the estimated point of interest distribution data includes first estimated point of interest distribution data; when the target session interaction behavior data only comprises second session interaction behavior data, the estimated interest point distribution data comprises second estimated interest point distribution data; when the target session interaction behavior data comprises the first session interaction behavior data and the second session interaction behavior data, the estimated point of interest distribution data may comprise third estimated point of interest distribution data, in which case the estimated point of interest distribution data may also comprise the first estimated point of interest distribution data and the second estimated point of interest distribution data, for example.

Step S104, generating reference interest point distribution data based on the observation interest points of the target session interaction behavior data in the template user preference data based on the user preference range observation model.

Wherein the reference point of interest distribution data may characterize a reference user preference range (i.e., an actual user preference range) of the target session interaction behavior data. The data for the interaction behavior of the target session may include one or more of the following: the first session interaction behavior data and the second session interaction behavior data; thus, the reference point of interest distribution data may include one or more of the following: the first reference point of interest distribution data with respect to the first session interaction behavior data, the second reference point of interest distribution data with respect to the second session interaction behavior data, and the third reference point of interest distribution data with respect to the first session interaction behavior data and the second session interaction behavior data. In other words, the first reference point of interest distribution data may characterize a reference user preference range of the first session interaction behavior data, the second reference point of interest distribution data may characterize a reference user preference range of the second session interaction behavior data, and the third reference point of interest distribution data may be used to reflect both the reference user preference range of the first session interaction behavior data and the reference user preference range of the second session interaction behavior data.

In an alternative embodiment, if the target session interaction behavior data is the first session interaction behavior data or the second session interaction behavior data, the template user preference data includes first user preference data or second user preference data; then the reference point of interest distribution data includes the first reference point of interest distribution data or the second reference point of interest distribution data, respectively. The step S104 may specifically include: first, observation points of interest of target session interaction behavior data in template user preference data may be determined based on a user preference range observation model, and a plurality of target user preference knowledge points belonging to the target session interaction behavior data may be determined in the template user preference data. When the target session interaction behavior data is the first session interaction behavior data, the observation interest point of the target session interaction behavior data refers to the observation interest point of the first session interaction behavior data, and the target user preference knowledge point refers to the first user preference knowledge point belonging to the first session interaction behavior data; when the target session interaction behavior data is second session interaction behavior data, the observation interest point of the target session interaction behavior data refers to the observation interest point of the second session interaction behavior data, and the target user preference knowledge point refers to a second user preference knowledge point belonging to the second session interaction behavior data. Then, the preference characteristic values of the preference knowledge points of each target user can be optimized based on the trigger matching degree between the behavior trigger node of the preference knowledge point of each target user and the observation interest point of the interaction behavior data of the target session, and the reference interest point distribution data (namely, the first reference interest point distribution data or the second reference interest point distribution data) can be generated. The method comprises the steps that a reference preference characteristic value of any target user preference knowledge point in reference interest point distribution data is positively correlated with a trigger matching degree of any target user preference knowledge point; in other words, among the plurality of target user preference knowledge points, the reference preference feature value of the target user preference knowledge point of the term in the interest set located on the observation interest point of the target session interaction behavior data is the largest, and the reference preference feature value of the target user preference knowledge point of the term in the non-interest set having the largest observation interest point deviating from the target session interaction behavior data is the smallest.

Illustratively, based on the trigger matching degree between the behavior trigger node of each target user preference knowledge point and the observation interest point of the target session interaction behavior data, respectively optimizing the preference feature value of each target user preference knowledge point, generating the reference interest point distribution data specifically may include: first, a target user preference knowledge point of an item in an interest set located on an observation interest point of target session interaction behavior data and a target user preference knowledge point of an item in a non-interest set other than the target user preference knowledge point of the item in the interest set may be determined from a plurality of target user preference knowledge points. The preference feature values of the target user preference knowledge points of the terms in the interest set may then be optimized to the set feature values. The specific step of optimizing the preference feature value of the target user preference knowledge point of the term in the interest set to the set feature value may be: adjusting the preference characteristic value of the target user preference knowledge point of the vocabulary entry in the interest set to a set characteristic value; alternatively, the preference feature value of the target user preference knowledge point of the term in the interest set may be directly determined as the set feature value, and in this case, the specific step of optimizing the preference feature value of the target user preference knowledge point of the term in the interest set to the set feature value may be: the preference characteristic value of the target user preference knowledge point of the vocabulary entry in the interest set is kept unchanged. In addition, according to an optimization strategy that the trigger matching degree of the target user preference knowledge points of the non-interest concentrated vocabulary entries and the preference feature values are in positive correlation, the preference feature values of the target user preference knowledge points of the non-interest concentrated vocabulary entries are optimized along the direction from the set feature values to the failure values (such as 0) so as to obtain the reference interest point distribution data. It can be seen that the reference preference feature values of the target user preference knowledge points of the terms in the non-interest sets in the reference interest point distribution data are smaller than the set feature values.

For example, the target session interaction behavior data is set as first session interaction behavior data, and the first session interaction behavior data is live behavior data; and the preference characteristic values of the first user preference knowledge points belonging to the first session interaction behavior data in the template user preference data (namely the first user preference data) are set to be 1, and the preference characteristic values of the other user preference knowledge points except the first user preference knowledge points are set to be 0. Then, while generating the first reference point of interest distribution data, the preference feature value of the first user preference knowledge point (e.g., user preference knowledge point u) of the vocabulary entry in the interest set located on the observation point of the live behavior data may be kept unchanged. And the preference characteristic values of the first user preference knowledge points of the non-interest concentrated vocabulary entries can be optimized stepwise and sequentially according to the trigger matching degree between the first user preference knowledge points (such as the user preference knowledge point b and the user preference knowledge point c) of the non-interest concentrated vocabulary entries and the observation interest points of the first session interaction behavior data along the direction gradually decreasing from 1 to 0 in accordance with the optimization strategy. Since the trigger matching degree of the user preference knowledge point b is smaller than that of the user preference knowledge point c, the preference characteristic value of the user preference knowledge point b can be optimized to 0.8, the preference characteristic value of the user preference knowledge point c can be optimized to 0.7, and so on.

Alternatively, if the target session interaction behavior data includes first session interaction behavior data and second session interaction behavior data, the template user preference data may include at least third user preference data; wherein the reference point of interest distribution data may comprise at least third reference point of interest distribution data. Thus, step S104 may specifically include: first, an observation point of interest of first session interaction behavior data and an observation point of interest of second session interaction behavior data in the third user preference data may be determined based on a user preference range observation model; and determining a plurality of first user preference knowledge points belonging to the first session interaction behavior data and a plurality of second user preference knowledge points belonging to the second session interaction behavior data in the third user preference data. Then, the preference characteristic values of the first user preference knowledge points can be optimized respectively based on the trigger matching degree between the behavior trigger nodes of the first user preference knowledge points and the observation interest points of the first session interaction behavior data; and optimizing preference characteristic values of the second user preference knowledge points based on trigger matching degrees between the behavior trigger nodes of the second user preference knowledge points and the observation interest points of the second session interaction behavior data, respectively, and generating reference interest point distribution data (namely distribution data of entries in a third reference line interest set). The method comprises the steps that a reference preference characteristic value of any first user preference knowledge point in reference interest point distribution data is positively correlated with the trigger matching degree of any first user preference knowledge point, and a reference preference characteristic value of any second user preference knowledge point in reference interest point distribution data is positively correlated with the trigger matching degree of any second user preference knowledge point. In other words, among the plurality of first user preference knowledge points (or the plurality of second user preference knowledge points) within the third reference point of interest distribution data, the reference preference feature value of the first user preference knowledge point (or the second user preference knowledge point of the term in the interest set) located on the observation point of the first session interaction behavior data (or the second session interaction behavior data) is the largest, and the reference preference feature value of the first user preference knowledge point (or the second user preference knowledge point of the term in the non-interest set) located farthest from the observation point of the first session interaction behavior data (or the second session interaction behavior data) is the smallest.

For example, based on the trigger matching degree between the behavior trigger node of each first user preference knowledge point and the observation interest point of the first session interaction behavior data, optimizing the preference feature value of each first user preference knowledge point may specifically include: first, first user preference knowledge points of the terms in the interest set that are located on the observation interest point of the first session interaction behavior data, and first user preference knowledge points of the terms in the non-interest set other than the first user preference knowledge points of the terms in the interest set, may be determined from a plurality of first user preference knowledge points. Then, optimizing the preference characteristic value of the first user preference knowledge point of the vocabulary entry in the interest set to the set characteristic value; and optimizing the preference feature values of the first user preference knowledge points of the non-interest concentrated vocabulary entries along the direction gradually decreasing from the set feature values to the failure values according to the optimization strategy that the trigger matching degree of the first user preference knowledge points of the non-interest concentrated vocabulary entries and the preference feature values are in positive correlation. Similarly, the specific embodiment of optimizing the preference feature value of each first user preference knowledge point based on the trigger matching degree between the behavior trigger node of each first user preference knowledge point and the observation interest point of the first session interaction behavior data may be referred to this specific embodiment, and will not be described herein.

When the target session interaction behavior data includes first session interaction behavior data and second session interaction behavior data, the template user preference data may include first user preference data and second user preference data in addition to third user preference data; the above reference point of interest distribution data may also include first reference point of interest distribution data and second reference point of interest distribution data.

Step S105, updating weight information of the user preference data positioning model based on the distinguishing distribution data between the estimated interest point distribution data and the reference interest point distribution data.

Step S106, user preference data positioning is carried out on target user session interaction behavior flow data of any target user based on the user preference data positioning model updated by the weight information, user preference data of the target user are generated, user interaction portrait classification is carried out on the user preference data, and user interaction portrait information of the target user is generated; the target user session interaction behavior flow data comprises at least two session interaction behavior data which are in a directed graph format and have different user preference ranges.

For example, in generating user interaction image information of the target user, the target user may be added to an interaction group corresponding to the user interaction image information for subsequent session interaction activities.

Since the estimated point of interest distribution data may characterize an estimated user preference range for the target session interaction behavior data, while the reference point of interest distribution data may characterize a reference user preference range (i.e., an actual user preference range) for the target session interaction behavior data; thus, the distinguishing distribution data between the estimated user preference range and the actual user preference range of the target session interaction behavior data can be reflected by the distinguishing distribution data between the estimated point of interest distribution data and the reference point of interest distribution data. The embodiment can update the weight information of the user preference data positioning model based on the distinguishing distribution data between the estimated interest point distribution data and the reference interest point distribution data. For example, weight parameter optimization error calculation may be performed based on the estimated point of interest distribution data and the reference point of interest distribution data based on a cross entropy cost calculation function, and a first user preference observation error value may be generated, where the first user preference observation error value characterizes distinguishing distribution data between the estimated point of interest distribution data and the reference point of interest distribution data. Second, a second user preference observation error value may be calculated based on the first user preference observation error value; the weight information of the user preference data positioning model may then be optimized in accordance with the direction in which the second user preference observation error value is reduced.

In the implementation process, the above steps S101 to S105 may be repeated multiple times, to obtain the user preference data positioning model after the weight information update. The user preference data positioning model with updated weight information can be used for positioning user preference data of target user session interaction behavior flow data, wherein the target user session interaction behavior flow data comprises session interaction behavior data with at least two directed graph formats and different user preference ranges; that is, the target user session interaction behavior traffic data refers to user session interaction behavior traffic data including at least two session interaction behavior data of a directed graph format and different user preference ranges.

By adopting the steps, the user preference data positioning model can firstly position the user preference data of the template user session interaction behavior flow data based on the user preference data positioning model, and the undetermined user interest point data obtained in the user preference data positioning process is loaded into the user preference range observation model, so that the user preference range observation model can generate estimated interest point distribution data based on the undetermined user interest point data. Because the template user preference data is generated by calibrating preference data of target session interaction behavior data in the template user session interaction behavior flow data, the user preference range observation model can also generate reference interest point distribution data which can be used as a training learning target based on observation interest points of the target session interaction behavior data in the template user preference data. Since the estimated point of interest distribution data may characterize an estimated user preference range for the target session interaction behavior data, while the reference point of interest distribution data may characterize a reference user preference range (i.e., an actual user preference range) for the target session interaction behavior data; therefore, the difference distribution data between the estimated interest point distribution data and the reference interest point distribution data can express the difference distribution data between the estimated user preference range and the actual user preference range of the target session interaction behavior data, so that when the weight information of the user preference data positioning model is updated according to the difference distribution data, the user preference data positioning model can focus and learn the user preference range information of the target session interaction behavior data, the user preference data positioning model can be deployed to have the observation capability of the user preference range, and the user preference data positioning accuracy can be improved by combining the user preference range of the session interaction behavior data when the user preference data positioning is carried out, so that the accuracy of the classification of the subsequent user interaction images can be improved.

In another embodiment, the above method may further include the following step S201-step S206:

step S201, template user session interaction behavior flow data and template user preference data are obtained, the template user session interaction behavior flow data comprise first session interaction behavior data in a directed graph format and second session interaction behavior data in a directed graph format, and the user preference range of the first session interaction behavior data is different from the user preference range of the second session interaction behavior data; the template user preference data is generated by preference data calibration of target session interaction behavior data in the template user session interaction behavior flow data, and the target session interaction behavior data comprises one or more of the following data: the first session interaction behavior data and the second session interaction behavior data.

Step S202, user preference data positioning is carried out on the template user session interaction behavior flow data based on the user preference data positioning model, and the undetermined user interest point data analyzed in the user preference data positioning process is loaded into the user preference range observation model.

Step S203, generating estimated interest point distribution data based on the undetermined user interest point data based on the user preference range observation model, and generating reference interest point distribution data based on the observed interest points of the target session interaction behavior data in the template user preference data based on the user preference range observation model; the estimated interest point distribution data represents the estimated user preference range of the target session interaction behavior data, and the reference interest point distribution data represents the reference user preference range of the target session interaction behavior data.

Step S204, performing weight parameter optimization error calculation based on the estimated interest point distribution data and the reference interest point distribution data based on the cross entropy cost calculation function, and generating a first user preference observation error value.

In an alternative embodiment, the weight parameter optimization error calculation may be directly performed based on the cross entropy cost calculation function, and the weight parameter optimization error calculation may be performed based on the estimated preference feature value of each user preference knowledge point in the template user session interaction behavior flow data in the estimated interest point distribution data and the reference preference feature value in the reference interest point distribution data, so as to generate the first user preference observation error value. Illustratively, step S204 may specifically include: the estimated preference characteristic values of all user preference knowledge points in the template user session interaction behavior flow data in the estimated interest point distribution data and the reference preference characteristic values of all user preference knowledge points in the reference interest point distribution data can be respectively obtained. Second, feature training error values between the reference preference feature values and the corresponding estimated preference feature values for each user preference knowledge point may be calculated separately. And then, carrying out weight fusion calculation on the feature training error values corresponding to the knowledge points of the user preference to generate a first user preference observation error value.

Step S204 may specifically include: first, a plurality of target user preference knowledge points belonging to target session interaction behavior data may be determined from the template user preference data. Then, the estimated preference characteristic values of the target user preference knowledge points in the estimated interest point distribution data and the reference preference characteristic values of the target user preference knowledge points in the reference interest point distribution data are obtained. Feature training error values between the reference preference feature values and the corresponding estimated preference feature values of the respective target user preference knowledge points are calculated respectively. And then carrying out weight fusion calculation on the calculated characteristic training error value to generate a first user preference observation error value.

In step S20, if the target session interaction behavior data includes the first session interaction behavior data and the second session interaction behavior data, the number of the reference point of interest distribution data and the estimated point of interest distribution data may be at least two; for example, the reference point of interest distribution data includes first, second, and third reference point of interest distribution data, and the estimated point of interest distribution data also includes first, second, and third estimated point of interest distribution data. Whereby any one of the above possible implementations may be selected to calculate a first user preference observation error value for each combined point of interest distribution data; the first user preference observation error values for each of the combined point of interest distribution data are then summed to generate a final first user preference observation error value. For example, a first user preference observation error value for a first combined point of interest distribution data (i.e., first reference point of interest distribution data and first estimated point of interest distribution data), a first user preference observation error value for a second combined point of interest distribution data (i.e., second reference point of interest distribution data and second estimated point of interest distribution data), and a first user preference observation error value for a third combined point of interest distribution data (i.e., third reference point of interest distribution data and third estimated point of interest distribution data) may be determined in sequence; and then, carrying out weight fusion calculation on the first user preference observation error value of the first combined interest point distribution data, the first user preference observation error value of the second combined interest point distribution data and the first user preference observation error value of the third combined interest point distribution data to generate a final first user preference observation error value.

Step S205 calculates a second user preference observation error value based on the first user preference observation error value.

In an alternative embodiment, the first user preference observation error value may be directly taken as the second user preference observation error value.

In an alternative embodiment, the model may first perform feature expansion on the template user preference data based on the M feature expansion parameters based on the user preference range observation model, and generate M feature-expanded user preference data. And then, calculating to obtain M candidate interest point distribution data with the expanded characteristics corresponding to the user preference data with the expanded characteristics. And, each feature-expanded candidate point of interest distribution data may include one or more of the following: the method comprises the steps of feature-expanded first reference interest point distribution data, feature-expanded second reference interest point distribution data and feature-expanded third reference interest point distribution data.

Then, based on the user interest point data, the estimated interest point distribution data after the M feature expansion can be estimated and generated based on the user preference range observation model, and one feature expansion knowledge base corresponds to one feature expansion estimated interest point distribution data. It can be seen that each estimated interest point distribution data may include three modes of estimated interest point distribution data including first estimated interest point distribution data, second estimated interest point distribution data, and third estimated interest point distribution data. Then, each feature extended estimated point of interest distribution data may include one or more of the following: the method comprises the steps of feature-expanded first estimated interest point distribution data, feature-expanded second estimated interest point distribution data and feature-expanded third estimated interest point distribution data.

And then, carrying out weight parameter optimization error calculation based on the estimated interest point distribution data after feature expansion and the candidate interest point distribution data after feature expansion corresponding to each feature expansion parameter by a cross entropy cost calculation function, and generating M feature expansion user preference observation error values. And finally, carrying out weight fusion calculation on the first user preference observation error value and the M characteristic extension user preference observation error values to generate a second user preference observation error value.

Based on the above steps, performing feature expansion on the template user preference data based on at least one feature expansion knowledge base, and calculating the interest point distribution data of the feature-expanded user preference data (namely candidate interest point distribution data after feature expansion and estimated interest point distribution data after feature expansion); the feature information of the user preference range can be effectively enriched, so that the feature information of the second user preference observation error value is effectively enriched, the accuracy of the second user preference observation error value is improved, and further the subsequent model training effect is improved.

Step S206, optimizing the weight information of the user preference data positioning model according to the direction of reducing the second user preference observation error value.

In an alternative embodiment, the weight information of the user preference data positioning model may be optimized directly in dependence of the direction in which the second user preference observation error value is reduced.

The smaller the remarkable correlation of the target session interaction behavior data in the estimated user preference data is, the more interruption features among estimated user preference knowledge points which belong to the target session interaction behavior data and are estimated in the estimated user preference data are indicated, so that the lower the accuracy of positioning the user preference data is, the worse the positioning effect of the user preference data of the session interaction behavior data is indicated. That is, the magnitude of the significant correlation of the target session interaction behavior data in estimating the user preference data is correlated with the user preference data positioning accuracy of the session interaction behavior data of the user preference data positioning model. Therefore, the application can further acquire the significant training error value generated by the user preference data positioning model in the user preference data positioning process of the session interaction behavior data, and train and optimize the user preference data positioning model by combining the significant training error value and the second user preference observation error value, thereby improving the estimation performance of the user preference range of the user preference data positioning model. Illustratively, a specific embodiment of step S206 may include the following steps A-C:

And A, obtaining a significant training error value.

Wherein the significant training error value may include at least a first training error value; the first training error value may characterize a distinguishing distribution data between a significant correlation of the target session interaction behavior data in the template user preference data and a significant correlation of the target session interaction behavior data in the estimated user preference data; the estimated user preference data is obtained by user preference data positioning of the template user session interaction behavior traffic data by a user preference data positioning model. The first to-be-learned behavior flow data cluster can be generated based on the template user session interaction behavior flow data and the template user preference data, and the second user session interaction behavior flow data pair can be generated based on the template user session interaction behavior flow data and the estimated user preference data. Secondly, the module user preference data and the estimated user preference data can be subjected to remarkable relevance ranking based on the behavior attention characteristics in the first behavior flow data cluster to be learned and the behavior attention characteristics in the second behavior flow data cluster to be learned by the attention mechanism unit, and remarkable relevance ranking information is generated. Then, a first training error value may be calculated based on the saliency assessment index and the saliency ranking information:

The embodiment of the application can directly sort the remarkable relevance of the template user preference data and the estimated user preference data based on the behavior attention characteristics in the first behavior flow data cluster to be learned and the behavior attention characteristics in the second behavior flow data cluster to be learned by the attention mechanism unit, and generate the remarkable relevance sorting information. In the above scheme, the saliency assessment index may characterize a significant correlation of the target session interaction behavior data in the template user preference data, which is greater than a significant correlation of the target session interaction behavior data in the estimated user preference data; it may in particular comprise training direction data estimating user preference data and training direction data of template user preference data.

In order to enrich the user preference data with different obvious correlations, the method can also generate disorder user preference data with lower obvious correlations based on the behavior attention characteristics of the real template user preference data, and load the template user session interaction behavior flow data and the disorder user preference data into a characteristic discrimination network in a correlation manner to sort the obvious correlations. The generating, based on the attention mechanism unit, the significant relevance ranking for the template user preference data and the estimated user preference data based on the behavior attention feature in the first behavior traffic data cluster to be learned and the behavior attention feature in the second behavior traffic data cluster to be learned may specifically include: firstly, carrying out time sequence mode feature scrambling on behavior attention features in template user preference data to generate scrambled user preference data; then, a third behavior flow data cluster to be learned can be generated based on the template user session interaction behavior flow data and the disorder user preference data, and the remarkable relevance ranking is performed on the template user preference data, the estimated user preference data and the disorder user preference data based on the behavior attention feature in the first behavior flow data cluster to be learned, the behavior attention feature in the second behavior flow data cluster to be learned and the behavior attention feature in the third behavior flow data cluster to be learned by the attention mechanism unit, so that remarkable relevance ranking information is generated.

Wherein the timing pattern feature scrambling includes one or more of the following: 1. deleting partial target user preference knowledge points belonging to the target session interaction behavior data in the template user preference data, namely setting preference characteristic values corresponding to the partial target user preference knowledge points as failure values; 2. and adjusting the behavior node of the target session interaction behavior data in the template user preference data, namely replacing the preference characteristic values of partial target user preference knowledge points and the preference characteristic values of non-target user preference knowledge points belonging to the target session interaction behavior data in the template user preference data.

The saliency assessment index characterizes the saliency of the target session interaction behavior data in the template user preference data, and is larger than the saliency of the target session interaction behavior data in the estimated user preference data; and the significant correlation of the target session interaction behavior data in the estimated user preference data is greater than the significant correlation of the target session interaction behavior data in the cluttered user preference data. Illustratively, the saliency assessment index may include first training direction data (represented based on t 1) that obscures user preference data, second training direction data (represented based on t 2) that estimates user preference data, and third training direction data (represented based on t 3) that templates user preference data.

Illustratively, the salient preference features may also be extracted based on the attention mechanism unit, such that the user preference data positioning model may be optimized based on distinguishing distribution data between salient preference features between the user preference data, thereby improving user preference data positioning performance of the user preference data positioning model. Wherein the salient training error values may further comprise second training error values characterizing distinguishing distribution data between salient preference features in the template user preference data and salient preference features of the estimated user preference data. Then step a may further comprise the steps of: feature extraction is carried out on the template user preference data based on the attention mechanism unit, and salient preference features in the template user preference data are generated; performing feature extraction on the estimated user preference data based on the attention mechanism unit to generate salient preference features in the estimated user preference data; the second training error value may then be calculated based on the salient preference features in the template user preference data and the salient preference features in the estimated user preference data.

In an alternative embodiment, the second training error value may be calculated directly based on the salient preference features in the template user preference data and the salient preference features in the estimated user preference data. The attention mechanism unit can also be used for extracting features of the disorder user preference data to generate salient preference features in the disorder user preference data; the second training error value may then be calculated based on a triplet loss function based on the salient preference features in the template user preference data, the salient preference features in the estimated user preference data, and the salient preference features in the shuffled user preference data. In the triplet loss function, the salient preference features in the real template user preference data can be used as anchors, the salient preference features in the estimated user preference data can be used as positive examples, and the salient preference features in the disordered user preference data can be used as negative examples.

And B, calculating a target training error value of the user preference data positioning model based on the second user preference observation error value and the significant training error value.

For example, a first importance coefficient (expressed based on ef) corresponding to a second user preference observation error value and a second importance coefficient corresponding to a significant training error value may be obtained; then, a weight fusion calculation can be performed on the second user preference observation error value and the significant training error value based on the first importance coefficient and the second importance coefficient to generate a target training error value of the user preference data positioning model. For example, the second user preference observation error value may be weight fused based on the first importance coefficient, and a weight fused user preference observation error value may be generated; and performing weight fusion on the significant training error value based on the second importance coefficient to generate a significant training error value after weight fusion. And then, summing the user preference observation error value after the weight fusion and the significant training error value after the weight fusion to generate a target training error value of the user preference data positioning model. If the significant training error value comprises a first training error value and a second training error value, the number of the second importance coefficients is two; the importance coefficient corresponding to the first training error value may be denoted as eu and the importance coefficient corresponding to the second training error value may be denoted as et. Then, performing weight fusion on the significant training error value based on the second importance coefficient, and generating the significant training error value after weight fusion may specifically include: and performing weight fusion on the first training error value based on eu, generating a first training error value after weight fusion, performing weight fusion on the second training error value based on et, generating a second training error value after weight fusion, and taking the sum of the first training error value after weight fusion and the second training error value after weight fusion as a target training error value of the user preference data positioning model. Wherein, eu, et and ef are set to 0.01, 0.2 and 0.2, respectively.

The weight fusion calculation is performed on the second user preference observation error value and the significant training error value based on the first importance coefficient and the second importance coefficient, and the generating the target training error value of the user preference data positioning model specifically may include: and carrying out weight fusion calculation on the second user preference observation error value and the significant training error value based on the first importance coefficient and the second importance coefficient, and carrying out summation operation on the result of the weight fusion calculation and the training error value of the cross entropy loss function to generate a target training error value of the user preference data positioning model.

And C, optimizing weight information of the user preference data positioning model according to the direction of reducing the target training error value.

The initial weight information of the user preference data positioning model is R1, and the initial weight information R1 of the user preference data positioning model can be optimized to R2 by performing the above steps S201-S206 for the first time; then, R2 may be used as current weight information of the user preference data positioning model, and the user preference data positioning model is optimized for the second time, that is, the above steps S201 to S206 are executed again to optimize the current weight information R2 of the user preference data positioning model as R3; then, R3 may be used as current weight information of the user preference data positioning model, and the user preference data positioning model may be subjected to third weight information optimization, that is, the above step S201-step S206 is executed again to optimize the current weight information R3 of the user preference data positioning model as R4 …, and so on, the weight information optimization is continuously performed on the user preference data positioning model, so that the user preference data positioning model meets the training termination condition, and the user preference data positioning model meeting the training termination condition is used as the user preference data positioning model that may be deployed.

Based on the steps, the user preference range observation model can be subjected to weight parameter optimization by acquiring the user preference data with different saliency, so that the user preference range observation model can distinguish the user preference data with different saliency, the user preference range observation model is more sensitive to the saliency preference characteristics of the target session interaction behavior data in the user preference data, the obtained saliency sequencing information can accurately express the relation among the saliency correlations of each user preference data, the training error value in the estimated user preference data is determined, and the user preference data positioning model is subjected to weight parameter optimization based on the training error value in the estimated user preference data, so that the user preference data positioning model is more concerned with the saliency correlation of the session interaction behavior data.

Based on the steps, the user preference data positioning model can firstly position the user preference data of the template user session interaction behavior flow data based on the user preference data positioning model, and the undetermined user interest point data obtained in the user preference data positioning process is loaded into the user preference range observation model, so that the user preference range observation model can generate estimated interest point distribution data based on the undetermined user interest point data. Because the template user preference data is generated by calibrating preference data of target session interaction behavior data in the template user session interaction behavior flow data, the user preference range observation model can also generate reference interest point distribution data which can be used as a training learning target based on observation interest points of the target session interaction behavior data in the template user preference data. Since the estimated point of interest distribution data may characterize an estimated user preference range for the target session interaction behavior data, while the reference point of interest distribution data may characterize a reference user preference range (i.e., an actual user preference range) for the target session interaction behavior data; therefore, the difference distribution data between the estimated interest point distribution data and the reference interest point distribution data can express the difference distribution data between the estimated user preference range and the actual user preference range of the target session interaction behavior data, so that when the weight information of the user preference data positioning model is updated according to the difference distribution data, the user preference data positioning model can focus and learn the user preference range information of the target session interaction behavior data, the user preference data positioning model can be deployed to have the observation capability of the user preference range, and the user preference data positioning accuracy can be improved by combining the user preference range of the session interaction behavior data when the user preference data positioning is carried out, so that the accuracy of the classification of the subsequent user interaction images can be improved.

Fig. 2 illustrates a hardware structural intent of the AI-session interaction system 100 for implementing the artificial intelligence-based user-interaction image classification method as provided by an embodiment of the application, and as shown in fig. 2, the AI-session interaction system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

In some alternative embodiments, the AI session interaction system 100 may be a single server or a group of servers. The server set may be centralized or distributed (e.g., the AI session interaction system 100 may be a distributed system). In some alternative embodiments, the AI session interaction system 100 may be local or remote. For example, the AI-session interaction system 100 may access information and/or data stored in the machine-readable storage medium 120 via a network. As another example, the AI session interaction system 100 can be directly connected to the machine-readable storage medium 120 to access stored information and/or data. In some alternative embodiments, the AI-session interaction system 100 may be implemented on a cloud platform. For example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

The machine-readable storage medium 120 may store data and/or instructions. In some alternative implementations, the machine-readable storage medium 120 may store data acquired from an external terminal. In some alternative implementations, the machine-readable storage medium 120 may store data and/or instructions that are used by the AI session interaction system 100 to perform or use the exemplary methods described herein. In some alternative implementations, the machine-readable storage medium 120 may include mass storage, removable storage, volatile read-write memory, read-only memory (ROM), and the like, or any combination thereof. Exemplary mass storage devices may include magnetic disks, optical disks, solid state disks, and the like. Exemplary removable memory may include flash drives, floppy disks, optical disks, memory cards, compact disks, tape, and the like. Exemplary volatile read-write memory can include Random Access Memory (RAM). Exemplary RAM may include active random access memory (DRAM), double data rate synchronous active random access memory (DDR SDRAM), passive random access memory (SRAM), thyristor random access memory (T-RAM), zero capacitance random access memory (Z-RAM), and the like. Exemplary read-only memory may include mask read-only memory (MROM), programmable read-only memory (PROM), erasable programmable read-only memory (PEROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM), digital versatile disk read-only memory, and the like. In some alternative implementations, the machine-readable storage medium 120 may be implemented on a cloud platform. For example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc., or any combination thereof.

In a specific implementation, the plurality of processors 110 execute computer executable instructions stored by the machine-readable storage medium 120, so that the processors 110 may execute the user interaction image classification method based on artificial intelligence according to the above method embodiment, the processors 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processors 110 may be used to control the transceiving actions of the communication unit 140.

The specific implementation process of the processor 110 may refer to the above-mentioned embodiments of the method executed by the AI session interaction system 100, and the implementation principle and technical effects are similar, which is not repeated herein.

In addition, the embodiment of the application also provides a readable storage medium, wherein computer executable instructions are preset in the readable storage medium, and when a processor executes the computer executable instructions, the user interaction image classification method based on artificial intelligence is realized.

Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof. Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof.

Claims

1. The user interaction image classification method based on artificial intelligence is characterized by being realized through an AI session interaction system, and comprises the following steps:

And observing the user preference range according to the to-be-determined user interest point data based on the user preference range observation model, updating weight information of the user preference data positioning model based on an observation result, and classifying user interaction images after the updated user preference data positioning model is used for positioning the user preference data.

2. The method for classifying user interaction images based on artificial intelligence according to claim 1, wherein the step of observing the user preference range according to the to-be-determined user interest point data based on the user preference range observation model to update weight information of the user preference data positioning model based on the observation result, wherein the updated user preference data positioning model is used for classifying user interaction images after positioning user preference data comprises the following steps:

3. The artificial intelligence based user interaction image classification method according to claim 2, wherein the pending user interest point data comprises: the user preference data positioning model comprises a plurality of K coding-level conversation interaction behavior vectors generated by K encoders, wherein one encoder generates a coding-level conversation interaction behavior vector, and the data volume of the conversation interaction behavior vectors of any two coding levels is different;

4. The artificial intelligence based user interaction image classification method according to claim 2, wherein the template user preference data comprises one or more of the following: first user preference data, second user preference data, and third user preference data; the first user preference data is generated by calibrating preference data of the first session interaction behavior data in the template user session interaction behavior traffic data, the second user preference data is generated by calibrating preference data of the second session interaction behavior data in the template user session interaction behavior traffic data, and the third user preference data is generated by calibrating preference data of the first session interaction behavior data and the second session interaction behavior data in the template user session interaction behavior traffic data;

5. The artificial intelligence based user interaction image classification method according to claim 4, wherein if the target session interaction behavior data is the first session interaction behavior data or the second session interaction behavior data, the template user preference data includes the first user preference data or the second user preference data, and the reference point of interest distribution data includes the first reference point of interest distribution data or the second reference point of interest distribution data;

6. The artificial intelligence based user interaction image classification method according to claim 4, wherein if the target session interaction behavior data includes the first session interaction behavior data and the second session interaction behavior data, the template user preference data includes the third user preference data, and the reference point of interest distribution data includes the third reference point of interest distribution data;

7. The artificial intelligence based user interaction image classification method according to claim 2, wherein the updating the weight information of the user preference data positioning model based on the distinguishing distribution data between the estimated point of interest distribution data and the reference point of interest distribution data comprises:

8. The artificial intelligence based user interaction image classification method according to claim 7, wherein the cross entropy based cost calculation function performs weight parameter optimization error calculation based on the estimated point of interest distribution data and the reference point of interest distribution data, and generates a first user preference observation error value, comprising:

9. The artificial intelligence based user interaction image classification method of claim 7 or 8, wherein the calculating a second user preference observation error value based on the first user preference observation error value comprises:

10. An AI session interaction system comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the artificial intelligence based user interaction image classification method of any of claims 1-9.