WO2023197910A1

WO2023197910A1 - User behavior prediction method and related device thereof

Info

Publication number: WO2023197910A1
Application number: PCT/CN2023/086192
Authority: WO
Inventors: 刘卫文; 唐睿明; 张瑞; 傅凌玥; 林江浩; 张伟楠; 俞勇
Original assignee: 华为技术有限公司
Priority date: 2022-04-12
Filing date: 2023-04-04
Publication date: 2023-10-19
Also published as: CN114707070A

Abstract

Disclosed in the present application are a user behavior prediction method and a related device thereof, by means of which method the probability that a project obtained by a neural network model is clicked on by a user has a higher degree of accuracy, thereby facilitating subsequent accurate recommendation of projects of interest to the user. The method of the present application comprises: acquiring a first feature of a first project and a second feature of a second project, wherein the first project and the second project are located in different lists or the same list of a target page, and the second project is located before the first project; acquiring a second feature of the first project on the basis of the first feature of the first project and the second feature of the second project; and on the basis of the second feature of the first project, acquiring the probability that the first project is clicked on by a user.

Description

A user behavior prediction method and related equipment

This application claims priority to the Chinese patent application filed with the China Patent Office on April 12, 2022, with application number 202210379948.7 and the invention title "A user behavior prediction method and related equipment", the entire content of which is incorporated by reference. in this application.

Technical field

This application relates to the technical field of artificial intelligence (AI), and in particular to a user behavior prediction method and related equipment.

Background technique

With the rapid development of computer technology, in order to meet users' Internet needs, developers are increasingly inclined to display content that users are interested in on their pages. Based on this, for a certain page, it is often necessary to predict which item or items displayed on the page the user will click on, that is, predict the user's behavior on the page, and then modify the items to be displayed on the page to recommend them to the user. Projects of interest.

Generally, the arrangement of items on a certain page is often presented to the user in the form of multiple lists, that is, the page usually contains multiple lists, and each list contains multiple items. When predicting the user's behavior on this page, for any item on the page, the neural network model of AI technology can be used to determine the probability of the item being clicked by the user.

However, when the neural network model provided by related technologies predicts the probability of a certain item being clicked by a user, it usually only considers the impact of the remaining items in the list where the item is located on the item. It can be seen that the factors considered by the relevant technology are relatively single, resulting in the probability that the item is clicked by the user finally obtained by the model, which is often less accurate. Therefore, it cannot accurately recommend items of interest to the user in the future.

Contents of the invention

The embodiments of this application provide a user behavior prediction method and related equipment, which can make the probability of items being clicked by the user obtained by the neural network model have a higher accuracy, which is conducive to subsequent accurate recommendation of items of interest to the user. project.

The first aspect of the embodiments of this application provides a user behavior prediction method, which method includes:

When it is necessary to predict user behavior on the target page, that is, when it is necessary to obtain the probability that the first item in the target page is clicked by the user, the first feature of the first item and the second feature of the second item can be obtained first, and the first The first feature of the item and the second feature of the second item are input into the target model. Among them, the first item and the second item are located in different lists or the same list on the target page, and the second item is located before the first item, that is, the positional relationship between the first item and the second item exists in the following two situations: (1) The first item and the second item may be items in the same list, the second item is located before the first item and the second item is adjacent to the first item. (2) The first item and the second item can be items in different lists. The list where the second item is located is located before the list where the first item is located. The second item can be adjacent to the first item or not. One item is adjacent.

After inputting the first feature of the first item and the second feature of the second item into the target model, the first feature of the first item and the second feature of the second item can be processed by the target model to obtain the first feature of the first item. Second characteristic. It is worth noting that the first characteristic of the first item can be the attribute information of the first item itself, then the second characteristic of the first item Characteristics are information obtained by fusion based on the attribute information of the first item (that is, the first feature of the first item). Since the acquisition process of the second feature of the second item is the same as the acquisition process of the second feature of the first item, The second feature of the second item is also information obtained by fusion based on the attribute information of the second item (the first feature of the second item).

Finally, the second feature of the first item can be processed through the target model to obtain the probability that the first item is clicked by the user.

It can be seen from the above method that when it is necessary to predict the probability that the first item in the target page is clicked by the user, the first feature of the first item and the second feature of the second item can be input to the target model, where the first item and The second item is in a different list or in the same list on the target page, and the second item is before the first item. Then, the target model can obtain the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item, and then obtain the probability that the first item is clicked by the user based on the second characteristic of the first item. . In the aforementioned process, when obtaining the probability that the first item is clicked by the user, the target model considers the impact of the second item before the first item on the first item. Since the second item can not only be the list where the first item is located, The items in can also be items in other lists. Therefore, the factors considered by the target model are relatively comprehensive and can fit the actual situation when the user browses to the first item in the target page. Therefore, the first item finally obtained by the target model The probability of being clicked by the user has a high accuracy, which is conducive to accurately recommending items of interest to the user in the future.

In a possible implementation, the method further includes: obtaining the first characteristic of the third item, the first item and the third item are located in different lists or the same list on the target page, and the third item is related to the first item. Neighbor; based on the first feature of the first item and the first feature of the third item, obtaining the third feature of the first item; based on the second feature of the first item, obtaining the probability that the first item is clicked by the user includes: based on the The second feature of the item and the third feature of the first item are used to obtain the probability that the first item is clicked by the user. In the aforementioned implementation, the first feature of the third item can also be input to the target model, where the first item and the third item are located in different lists or the same list on the target page, and the third item is adjacent to the first item. , that is, the positional relationship between the first item and the third item exists in the following two situations: (1) The first item and the third item can be items in the same list, and the third item and the first item are adjacent. (2) The first item and the third item can be items in different lists, and the third item and the first item are adjacent. After the first feature of the third item is input into the target model, the first feature of the first item and the first feature of the third item can be processed by the target model, thereby obtaining the third feature of the first item. After obtaining the second feature of the first item and the third feature of the first item, the target model can calculate the second feature of the first item and the third feature of the first item, thereby obtaining the probability that the first item is clicked by the user. . Since the second characteristic of the first item can represent the impact of the second item on the first item, that is, when the user uses sequential browsing behavior and skipping behavior to browse to the first item, the user browses during these behaviors. The impact of the project on the first project. The third feature of the first project can represent the impact of the third project on the first project. That is, when the user uses the contrast behavior to browse to the first project, the user is in the process of performing this behavior. It can be seen that when predicting user behavior, the target model not only introduces conventional sequential browsing behavior, but also introduces browsing behaviors such as jump behavior and comparison behavior, that is, That is to say, the target model will consider the impact of the items browsed by the user on the first item when the user uses these complex and diverse browsing behaviors to browse the first item, which can further improve the final result of the target model. The accuracy of the probability that the first item is clicked by the user.

In a possible implementation, based on the first feature of the first item and the second feature of the second item, obtaining the second feature of the first item includes: mapping the first feature of the first item to obtain the second feature of the first item. The fourth feature of the first item; the second feature of the second item is processed based on the self-attention mechanism to obtain the fifth feature of the first item; the first item The fourth feature of the object and the fifth feature of the first item are subjected to a first fusion process to obtain the second feature of the first item. In the aforementioned implementation manner, after the first feature of the first item and the second feature of the second item are input into the target model, the target model can map the first feature of the first item on the latent space to obtain the first feature of the first item. Four features. At the same time, the target model can also process the second feature of the second item based on the self-attention mechanism to obtain the fifth feature of the first item. After obtaining the fourth feature of the first item and the fifth feature of the first item, the target model can use the recurrent neural unit to process the fourth feature of the first item and the fifth feature of the first item, thereby accurately obtaining the first item's fourth feature. Second characteristic.

In a possible implementation, the first feature of the first item is mapped to obtain the fourth feature of the first item: the first feature of the first item, the user's request for the target page, and the second item being The probability of the user clicking is mapped to obtain the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item; the sixth feature of the first item, the seventh feature of the first item, and The eighth feature of the first item is subjected to the second fusion process to obtain the fourth feature of the first item. In the foregoing implementation, before obtaining the fourth feature of the first item, the user's request for the target page and the probability that the second item is clicked by the user can also be input to the target model. Then, the target model can respectively obtain the third feature of the first item. The first feature, the user's request for the target page and the probability of the second item being clicked by the user are mapped on the latent space, and accordingly the sixth feature of the first item, the seventh feature of the first item and the eighth feature of the first item are obtained , and then splice the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item to obtain the fourth feature of the first item. It can be seen that when the target model analyzes the first item, it not only takes into account the influence of the attribute information of the first item itself, but also considers the influence of external factors such as the user's request for the target page and the probability of the second item being clicked by the user. The impact produced by the target model further improves the accuracy of the probability of the first item being clicked by the user.

In a possible implementation, based on the first feature of the first item and the first feature of the third item, obtaining the third feature of the first item includes: comparing the first feature of the first item and the third feature of the third item. Perform mapping processing on one feature to obtain the sixth feature of the first item and the ninth feature of the first item; perform a third fusion process on the sixth feature of the first item and the ninth feature of the first item to obtain the sixth feature of the first item. The tenth feature; performs the fourth fusion process on the sixth feature of the first item and the tenth feature of the first item to obtain the third feature of the first item. In the foregoing implementation, after the first feature of the third item is input into the target model, the target model can respectively map the first feature of the first item and the first feature of the third item on the latent space, and obtain the first item accordingly. The sixth characteristic of the first item and the ninth characteristic of the first item. Then, the target model can calculate the sixth feature of the first item and the ninth feature of the first item through the comparison function, and then perform a weighted sum based on the calculation results to obtain the tenth feature of the first item. Finally, the target model can perform an exclusive OR operation on the sixth feature of the first item and the tenth feature of the first item to accurately obtain the third feature of the first item.

In a possible implementation, performing a fourth fusion process on the sixth feature of the first item and the tenth feature of the first item, and obtaining the third feature of the first item includes: mapping the user's request for the target page , the seventh feature of the first item is obtained; the fourth fusion process is performed on the sixth feature of the first item, the seventh feature of the first item, and the tenth feature of the first item, to obtain the third feature of the first item. In the aforementioned implementation, when obtaining the third feature of the first item, the target model can also map the user's request for the target page on the latent space to obtain the seventh feature of the first item. Then, the target model can map the third feature of the first item. The sixth characteristic of one item, the seventh characteristic of the first item and the tenth characteristic of the first item are subjected to an exclusive OR operation to obtain the third characteristic of the first item. It can be seen that when the target model analyzes the first item, it not only takes into account the influence of the attribute information of the first item itself, but also considers the influence of external factors such as the user's request for the target page, thereby further improving the target model The final accuracy of the probability that the first item is clicked by the user.

In one possible implementation, if the first item is the first item in the target page, then the second characteristic of the second item is the preset value.

In a possible implementation, the target page contains multiple lists, multiple items located in the multiple lists form a directed acyclic graph, and the multiple items include a first item, a second item, and a third item.

The second aspect of the embodiment of the present application provides a method for constructing a directed acyclic graph. The method includes: obtaining the eye movement data of the user browsing the target page; based on the eye movement data, determining the user's browsing behavior for multiple items. Each item is located in multiple lists on the target page; based on browsing behavior, multiple items are connected to obtain a directed acyclic graph.

It can be seen from the above method that the user's browsing behavior for multiple items in the target page can be determined based on the eye movement data generated when the user browses the target page. Then, these browsing behaviors (for example, sequential browsing behavior and skipping behavior ), often determines the user's browsing order of items (for example, the user's browsing order in the same list and the user's browsing order between different lists), thereby connecting multiple items on the target page according to these browsing to obtain the target The directed acyclic graph of the page can be used in subsequent predictions of user behavior on the target page. Since the directed acyclic graph involves users’ complex and diverse browsing behaviors, it is helpful to improve users’ understanding of the target page. Accuracy of behavioral predictions.

In one possible implementation, based on browsing behavior, connecting multiple items to obtain a directed acyclic graph includes: connecting items in the same list that the user browsed in the first order, and The items in different lists browsed by the user in the second order are connected in the second order to obtain a directed acyclic graph. In the aforementioned implementation method, the user's browsing behavior includes two major types of browsing behavior. The first type of browsing behavior refers to the user browsing items in the same list, including the first type of sequential browsing behavior. Therefore, the user's browsing order in the same list can be called the first order, and the first order includes the first type of sequential browsing behavior. , the top-to-bottom and left-to-right order in which users browse all items in the same list. The second type of browsing behavior refers to the user browsing items between different lists, including the second type of sequential browsing behavior and comparison behavior. Therefore, the user's browsing order between different lists can be called the second order, and the second order includes the second type. In the sequential browsing behavior, the order in which the user browses several adjacent items in two adjacent lists, and in the comparison behavior, the jump order in which the user browses two items in two non-adjacent lists. . Then, all items in the target page can be connected according to the first order and the second order, thereby obtaining a directed acyclic graph for the target page.

In one possible implementation, obtaining the eye movement data of the user browsing the target page: collecting the eye movement data of the user browsing the target page through an eye tracker.

The third aspect of the embodiment of the present application provides a model training method. The method includes: obtaining the first feature of the first item and the second feature of the second item through the model to be trained, and the first item and the second item are located in the to-be-trained model. Process different lists on the page or the same list, and the second item is located before the first item; obtain the second feature of the first item based on the first feature of the first item and the second feature of the second item through the model to be trained, Among them, the first feature of the first item is the attribute information of the first item, the second feature of the first item is the information obtained by fusion based on the attribute information of the first item, and the second feature of the second item is the information based on the second item. Information obtained by fusing the attribute information of the second item (i.e., the first feature of the second item); using the model to be trained based on the second feature of the first item, the probability that the first item is clicked by the user is obtained; based on the probability that the first item is clicked by the user probability and the real probability that the first item is clicked by the user, and obtains the target loss. The target loss is used to indicate the difference between the probability that the first item is clicked by the user and the real probability that the first item is clicked by the user; based on the target loss, update the Train the parameters of the model until the model training conditions are met and the target model is obtained.

The target model obtained by the above method has the ability to predict user behavior on the page. When it is necessary to predict the probability that the first item in the target page is clicked by the user, the first feature of the first item and the first feature of the second item can be input to the target model. Second feature, wherein the first item and the second item are located in different lists or the same list on the target page, and the second item is located before the first item. Then, the target model can obtain the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item, and then obtain the probability that the first item is clicked by the user based on the second characteristic of the first item. . In the aforementioned process, when obtaining the probability that the first item is clicked by the user, the target model considers the impact of the second item before the first item on the first item. Since the second item can not only be the list where the first item is located, The items in can also be items in other lists. Therefore, the factors considered by the target model are relatively comprehensive and can fit the actual situation when the user browses to the first item in the target page. Therefore, the first item finally obtained by the target model The probability of being clicked by the user has a high accuracy, which is conducive to accurately recommending items of interest to the user in the future.

In a possible implementation, the method further includes: obtaining the first feature of the third item through the model to be trained, the first item and the third item are located in different lists or the same list of the page to be processed, and the third item adjacent to the first item; through the model to be trained based on the first feature of the first item and the first feature of the third item, the third feature of the first item is obtained; through the model to be trained based on the second feature of the first item, Obtaining the probability that the first item is clicked by the user includes: using the to-be-trained model to obtain the probability that the first item is clicked by the user based on the second feature of the first item and the third feature of the first item.

In a possible implementation, based on the first feature of the first item and the second feature of the second item, obtaining the second feature of the first item includes: mapping the first feature of the first item to obtain the second feature of the first item. The fourth feature of the first item; the second feature of the second item is processed based on the self-attention mechanism to obtain the fifth feature of the first item; the fourth feature of the first item and the fifth feature of the first item are processed The first fusion process is to obtain the second feature of the first item.

In a possible implementation, the first feature of the first item is mapped to obtain the fourth feature of the first item: the first feature of the first item, the user's request for the page to be processed, and the second item being processed. The probability of the user clicking is mapped to obtain the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item; the sixth feature of the first item, the seventh feature of the first item, and The eighth feature of the first item is subjected to the second fusion process to obtain the fourth feature of the first item.

In a possible implementation, based on the first feature of the first item and the first feature of the third item, obtaining the third feature of the first item includes: comparing the first feature of the first item and the third feature of the third item. Perform mapping processing on one feature to obtain the sixth feature of the first item and the ninth feature of the first item; perform a third fusion process on the sixth feature of the first item and the ninth feature of the first item to obtain the sixth feature of the first item. The tenth feature; performs the fourth fusion process on the sixth feature of the first item and the tenth feature of the first item to obtain the third feature of the first item.

In a possible implementation, performing a fourth fusion process on the sixth feature of the first item and the tenth feature of the first item to obtain the third feature of the first item includes: mapping the user's request for the page to be processed , the seventh feature of the first item is obtained; the fourth fusion process is performed on the sixth feature of the first item, the seventh feature of the first item, and the tenth feature of the first item, to obtain the third feature of the first item.

In a possible implementation, if the first item is the first item in the page to be processed, the second characteristic of the second item is a preset value.

In a possible implementation, the page to be processed includes multiple lists, multiple items located in the multiple lists form a directed acyclic graph, and the multiple items include a first item, a second item, and a third item.

The fourth aspect of the embodiment of the present application provides a user behavior prediction device. The device includes: a first acquisition module, configured to acquire the first feature of the first item and the second feature of the second item through the target model. The first project and second project is located in a different list or the same list on the target page, and the second item is located before the first item; the second acquisition module is used to obtain the second item based on the first feature of the first item and the second feature of the second item through the target model. The second characteristic of an item, wherein the first characteristic of the first item is the attribute information of the first item, the second characteristic of the first item is the information obtained by fusion based on the attribute information of the first item, and the second characteristic of the second item is The second feature is the information obtained by fusion based on the attribute information of the second item (i.e., the first feature of the second item); the third acquisition module is used to obtain the first item based on the second feature of the first item through the target model. The probability of a user clicking.

It can be seen from the above device that when it is necessary to predict the probability that the first item in the target page is clicked by the user, the first feature of the first item and the second feature of the second item can be input to the target model, where the first item and The second item is in a different list or in the same list on the target page, and the second item is before the first item. Then, the target model can obtain the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item, and then obtain the probability that the first item is clicked by the user based on the second characteristic of the first item. . In the aforementioned process, when obtaining the probability that the first item is clicked by the user, the target model considers the impact of the second item before the first item on the first item. Since the second item can not only be the list where the first item is located, The items in can also be items in other lists. Therefore, the factors considered by the target model are relatively comprehensive and can fit the actual situation when the user browses to the first item in the target page. Therefore, the first item finally obtained by the target model The probability of being clicked by the user has a high accuracy, which is conducive to accurately recommending items of interest to the user in the future.

In a possible implementation, the device further includes: a fourth acquisition module, configured to acquire the first feature of the third item through the target model, where the first item and the third item are located in different lists or the same list of the target page. , and the third item is adjacent to the first item; the fifth acquisition module is used to obtain the third feature of the first item based on the first feature of the first item and the first feature of the third item through the target model; the third The acquisition module is configured to obtain the probability that the first item is clicked by the user based on the second feature of the first item and the third feature of the first item through the target model.

In a possible implementation, the second acquisition module is configured to: map the first feature of the first item through the target model to obtain the fourth feature of the first item; map the third feature of the second item through the target model. The two features are processed based on the self-attention mechanism to obtain the fifth feature of the first item; the fourth feature of the first item and the fifth feature of the first item are first fused through the target model to obtain the first item's fifth feature. Second characteristic.

In a possible implementation, the second acquisition module is configured to: use the target model to map the first feature of the first item, the user's request for the target page, and the probability that the second item is clicked by the user, to obtain the third The sixth characteristic of the first item, the seventh characteristic of the first item and the eighth characteristic of the first item; through the target model, the sixth characteristic of the first item, the seventh characteristic of the first item and the eighth characteristic of the first item The second fusion process is performed to obtain the fourth feature of the first item.

In a possible implementation manner, the fifth acquisition module is used to map the first feature of the first item and the first feature of the third item through the target model to obtain the sixth feature of the first item and the first feature of the third item. The ninth feature of the first item; the third fusion process is performed on the sixth feature of the first item and the ninth feature of the first item through the target model to obtain the tenth feature of the first item; the third feature of the first item is obtained through the target model The six features and the tenth feature of the first item are subjected to the fourth fusion process to obtain the third feature of the first item.

In a possible implementation, the fifth acquisition module is used to: map the user's request for the target page through the target model to obtain the seventh feature of the first item; map the sixth feature of the first item through the target model Features, the seventh feature of the first item, and the tenth feature of the first item are subjected to a fourth fusion process to obtain the third feature of the first item.

The fifth aspect of the embodiment of the present application provides a device for constructing a directed acyclic graph. The device includes: an acquisition module, used to obtain the eye movement data of the user browsing the target page; and a determination module, used to determine based on the eye movement data. The user's browsing behavior for multiple items, multiple items are located in multiple lists on the target page; the connection module is used to connect multiple items based on the browsing behavior to obtain a directed acyclic graph.

The above device can determine the user's browsing behavior for multiple items in the target page based on the eye movement data generated when the user browses the target page. Then, these browsing behaviors (such as sequential browsing behavior and skipping behavior) often determine The order in which the user browses the items (for example, the order in which the user browses in the same list and the order in which the user browses between different lists) is used to connect multiple items of the target page according to these views, and a directed and undirected view of the target page is obtained. Ring graph, this directed acyclic graph can be used in the subsequent prediction of user behavior on the target page. Since this directed acyclic graph involves users' complex and diverse browsing behaviors, it is helpful to improve the accuracy of user behavior prediction on the target page. .

In a possible implementation, the connection module is used to connect items in the same list that the user browses in the first order, and connect items in different lists that the user browses in the second order. Connect in the second order to obtain a directed acyclic graph.

In a possible implementation, the acquisition module is used to collect eye movement data of the user browsing the target page through an eye tracker.

The sixth aspect of the embodiment of the present application provides a schematic structural diagram of a model training device. The device includes: a first acquisition module, configured to acquire the first feature of the first item and the second feature of the second item through the model to be trained. Features, the first item and the second item are located in different lists or the same list on the page to be processed, and the second item is located before the first item; the second acquisition module is used to use the model to be trained based on the first feature of the first item and the second feature of the second item, to obtain the second feature of the first item, where the first feature of the first item is the attribute information of the first item, and the second feature of the first item is the attribute information based on the first item The information obtained by fusion, the second feature of the second item is the information obtained by fusion based on the attribute information of the second item (ie, the first feature of the second item); the third acquisition module is used to use the model to be trained based on the first feature of the second item. The second feature of an item is to obtain the probability that the first item is clicked by the user; the fourth acquisition module is used to obtain the target loss based on the probability of the first item being clicked by the user and the real probability of the first item being clicked by the user. Used to indicate the difference between the probability of the first item being clicked by the user and the real probability of the first item being clicked by the user; the update module is used to update the parameters of the model to be trained based on the target loss until the model training conditions are met and the target is obtained Model.

The target model trained by the above device has the ability to predict user behavior on the page. When it is necessary to predict the probability that the first item in the target page is clicked by the user, the first feature of the first item and the second feature of the second item can be input to the target model, where the first item and the second item are located on the target page. Different lists or the same list, with the second item before the first item. Then, the target model can obtain the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item, and then obtain the probability that the first item is clicked by the user based on the second characteristic of the first item. . In the aforementioned process, when obtaining the probability that the first item is clicked by the user, the target model considers the impact of the second item before the first item on the first item. Since the second item can not only be the list where the first item is located, The items in can also be items in other lists, so the factors considered by the target model are relatively comprehensive and can fit the user's needs in the target page. Based on the actual situation when browsing to the first item, the probability that the first item is clicked by the user finally obtained by the target model has a high accuracy, which is conducive to accurately recommending items of interest to the user in the future.

In a possible implementation, the device includes: a fifth acquisition module, configured to acquire the first feature of the third item through the model to be trained, and the first item and the third item are located in different lists or the same list of the page to be processed. , and the third item is adjacent to the first item; the sixth acquisition module is used to obtain the third feature of the first item based on the first feature of the first item and the first feature of the third item through the model to be trained; The third acquisition module is used to obtain the probability that the first item is clicked by the user based on the second feature of the first item and the third feature of the first item through the model to be trained.

In a possible implementation, the second acquisition module is used to: perform mapping processing on the first feature of the first item to obtain the fourth feature of the first item; perform self-attention-based processing on the second feature of the second item. The force mechanism is processed to obtain the fifth characteristic of the first item; the fourth characteristic of the first item and the fifth characteristic of the first item are first fused to obtain the second characteristic of the first item.

In a possible implementation, the second acquisition module is used to map the first feature of the first item, the user's request for the page to be processed, and the probability that the second item is clicked by the user, and obtain the first item's first characteristic. The sixth feature, the seventh feature of the first item, and the eighth feature of the first item; performing a second fusion process on the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item, Get the fourth characteristic of the first item.

In a possible implementation, the sixth acquisition module is configured to map the first feature of the first item and the first feature of the third item to obtain the sixth feature of the first item and the first feature of the first item. The ninth characteristic; perform a third fusion process on the sixth characteristic of the first item and the ninth characteristic of the first item to obtain the tenth characteristic of the first item; perform the sixth characteristic of the first item and the tenth characteristic of the first item The features are subjected to the fourth fusion process to obtain the third feature of the first item.

In a possible implementation, the sixth acquisition module is used to: map the user's request for the page to be processed to obtain the seventh feature of the first item; map the sixth feature of the first item, the first item's The seventh feature and the tenth feature of the first item are subjected to the fourth fusion process to obtain the third feature of the first item.

The seventh aspect of the embodiment of the present application provides a user behavior prediction device, which includes a memory and a processor; the memory stores code, and the processor is configured to execute the code. When the code is executed, the user behavior prediction device executes as follows: The method described in the first aspect or any possible implementation manner of the first aspect.

The eighth aspect of the embodiment of the present application provides a device for constructing a directed acyclic graph, which includes a memory and a processor; the memory stores code, and the processor is configured to execute the code. When the code is executed, the directed acyclic graph is The ring graph construction device is as described in the second aspect or any possible implementation manner of the second aspect.

A ninth aspect of the embodiment of the present application provides a model training device, which includes a memory and a processor; the memory stores code, and the processor is configured to execute the code. When the code is executed, the model training device executes the third step The method described in any possible implementation manner of the aspect or the third aspect.

A tenth aspect of the embodiments of the present application provides a circuit system. The circuit system includes a processing circuit configured to perform any of the possible implementations of the first aspect, the second aspect, Any possible implementation manner in the second aspect or the method described in any possible implementation manner in the third aspect or the third aspect.

An eleventh aspect of the embodiments of the present application provides a chip system. The chip system includes a processor for calling a computer program or computer instructions stored in a memory, so that the processor executes the first aspect as described in the first aspect. Any one of the possible implementations in the second aspect, any one of the possible implementations of the second aspect, or the third aspect, the method described in any one of the possible implementations of the third aspect.

In one possible implementation, the processor is coupled to the memory through an interface.

In a possible implementation, the chip system further includes a memory, and computer programs or computer instructions are stored in the memory.

A twelfth aspect of the embodiments of the present application provides a computer storage medium. The computer storage medium stores a computer program. When the program is executed by a computer, the computer implements any one of the first aspect and the first aspect. Possible implementations, the second aspect, any one possible implementation of the second aspect, or the third aspect, the method described in any one of the possible implementations of the third aspect.

A thirteenth aspect of the embodiments of the present application provides a computer program product. The computer program product stores instructions. When executed by a computer, the instructions make it possible for the computer to implement any one of the first aspect and the first aspect. The method described in the implementation, the second aspect, any one possible implementation of the second aspect, or the third aspect, any one possible implementation of the third aspect.

In the embodiment of the present application, when it is necessary to predict the probability that the first item in the target page is clicked by the user, the first feature of the first item and the second feature of the second item can be input to the target model, where the first item and the second feature can be input to the target model. The two items are in different lists or in the same list on the target page, and the second item is before the first item. Then, the target model can obtain the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item, and then obtain the probability that the first item is clicked by the user based on the second characteristic of the first item. . In the aforementioned process, when obtaining the probability that the first item is clicked by the user, the target model considers the impact of the second item before the first item on the first item. Since the second item can not only be the list where the first item is located, The items in can also be items in other lists. Therefore, the factors considered by the target model are relatively comprehensive and can fit the actual situation when the user browses to the first item in the target page. Therefore, the first item finally obtained by the target model The probability of being clicked by the user has a high accuracy, which is conducive to accurately recommending items of interest to the user in the future.

Description of the drawings

Figure 1 is a structural schematic diagram of the main framework of artificial intelligence;

Figure 2a is a schematic structural diagram of the user behavior prediction system provided by the embodiment of the present application;

Figure 2b is another structural schematic diagram of the user behavior prediction system provided by the embodiment of the present application;

Figure 2c is a schematic diagram of related equipment for data sequence processing provided by the embodiment of the present application;

Figure 3 is a schematic diagram of the architecture of the system 100 provided by the embodiment of the present application;

Figure 4 is a schematic flow chart of a directed acyclic graph construction method provided by an embodiment of the present application;

Figure 5 is a schematic diagram of the target page provided by the embodiment of the present application;

Figure 6 is a schematic diagram of an eye tracker provided by an embodiment of the present application;

Figure 7 is a schematic diagram of a directed acyclic graph provided by an embodiment of the present application;

Figure 8 is a schematic flow chart of the user behavior prediction method provided by the embodiment of the present application;

Figure 9 is a schematic structural diagram of the target model provided by the embodiment of the present application;

Figure 10 is a schematic flow chart of the model training method provided by the embodiment of the present application;

Figure 11 is a schematic structural diagram of a user behavior prediction device provided by an embodiment of the present application;

Figure 12 is a schematic structural diagram of a directed acyclic graph construction device provided by an embodiment of the present application;

Figure 13 is a schematic structural diagram of the model training device provided by the embodiment of the present application;

Figure 14 is a schematic structural diagram of an execution device provided by an embodiment of the present application;

Figure 15 is a schematic structural diagram of the training equipment provided by the embodiment of the present application;

Figure 16 is a schematic structural diagram of a chip provided by an embodiment of the present application.

Detailed ways

The terms "first", "second", etc. in the description and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms so used are interchangeable under appropriate circumstances, and are merely a way of distinguishing objects with the same attributes in describing the embodiments of the present application. Furthermore, the terms "include" and "having" and any variations thereof, are intended to cover non-exclusive inclusions, such that a process, method, system, product or apparatus comprising a series of elements need not be limited to those elements, but may include not explicitly other elements specifically listed or inherent to such processes, methods, products or equipment.

Generally, the arrangement of items on a certain page is often presented to the user in the form of multiple lists, that is, the page usually contains multiple lists, and each list contains multiple items. When predicting the user's target When the page behaves, for any item on the page, the neural network model in the field of AI technology can be used to determine the probability of the item being clicked by the user. For example, on a page of an app mall, multiple horizontal lists and multiple vertical arrangements are displayed. These multiple horizontal lists and multiple vertical arrangements are staggered. Multiple applications in the horizontal list are arranged in a row, and in the vertical arrangement, Multiple applications are arranged in a row, so that the page can display the introduction information of various applications to the user in the form of a staggered list. In order to predict the user's clicking behavior on this page, the neural network model can be used to analyze each application one by one, thereby obtaining the probability that the user clicks on each application on the page.

Furthermore, the user's browsing behavior on this page is often complicated. For example, when the user is browsing the items in the current list, he directly jumps to the items in another list (the other list and the current list are two non-adjacent lists). to browse. Related technologies often fail to take into account the impact of multiple complex browsing behaviors, and will also reduce the accuracy of the probability of an item being clicked by the user finally obtained by the model.

Furthermore, when models of related technologies analyze a certain project, they often only focus on the relevant information of the project itself. (For example, assuming a project is an application, the relevant information of the application includes the developer, type, size, etc. of the application.) The analysis does not take into account the impact of external factors such as users, which will also degrade the model. The accuracy of the final probability of the item being clicked by the user.

In order to solve this problem, embodiments of the present application provide a user behavior prediction method, which can be implemented in conjunction with artificial intelligence (artificial intelligence, AI) technology. AI technology is a technical discipline that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence. AI technology obtains the best results by perceiving the environment, acquiring knowledge and using knowledge. In other words, artificial intelligence technology is a branch of computer science that attempts to understand the nature of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Using artificial intelligence for data processing is a common application method of artificial intelligence.

First, the overall workflow of the artificial intelligence system is described. Please refer to Figure 1. Figure 1 is a structural schematic diagram of the main framework of artificial intelligence. The following is from the "intelligent information chain" (horizontal axis) and "IT value chain" (vertical axis) The above artificial intelligence theme framework is elaborated on in two dimensions. Among them, the "intelligent information chain" reflects a series of processes from data acquisition to processing. For example, it can be the general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, intelligent execution and output. In this process, the data has gone through the condensation process of "data-information-knowledge-wisdom". The "IT value chain" reflects the value that artificial intelligence brings to the information technology industry, from the underlying infrastructure of human intelligence and information (providing and processing technology implementation) to the systematic industrial ecological process.

(1)Infrastructure

Infrastructure provides computing power support for artificial intelligence systems, enables communication with the external world, and supports it through basic platforms. Communicate with the outside through sensors; computing power is provided by smart chips (hardware acceleration chips such as CPU, NPU, GPU, ASIC, FPGA, etc.); the basic platform includes distributed computing framework and network and other related platform guarantees and support, which can include cloud storage and Computing, interconnection networks, etc. For example, sensors communicate with the outside world to obtain data, which are provided to smart chips in the distributed computing system provided by the basic platform for calculation.

(2)Data

Data from the upper layer of the infrastructure is used to represent data sources in the field of artificial intelligence. The data involves graphics, images, voice, and text, as well as IoT data of traditional devices, including business data of existing systems and sensory data such as force, displacement, liquid level, temperature, and humidity.

(3)Data processing

Data processing usually includes data training, machine learning, deep learning, search, reasoning, decision-making and other methods.

Among them, machine learning and deep learning can perform symbolic and formal intelligent information modeling, extraction, preprocessing, training, etc. on data.

Reasoning refers to the process of simulating human intelligent reasoning in computers or intelligent systems, using formal information to perform machine thinking and problem solving based on reasoning control strategies. Typical functions are search and matching.

Decision-making refers to the process of decision-making after intelligent information is reasoned, and usually provides functions such as classification, sorting, and prediction.

(4) General ability

After the data is processed as mentioned above, some general capabilities can be formed based on the results of further data processing, such as algorithms or a general system, such as translation, text analysis, computer vision processing, speech recognition, and image processing. identification, etc.

(5) Intelligent products and industry applications

Intelligent products and industry applications refer to the products and applications of artificial intelligence systems in various fields. They are the encapsulation of overall artificial intelligence solutions, productizing intelligent information decision-making and realizing practical applications. Its application fields mainly include: intelligent terminals, intelligent transportation, Smart healthcare, autonomous driving, smart cities, etc.

Next, several application scenarios of this application will be introduced.

Figure 2a is a schematic structural diagram of a user behavior prediction system provided by an embodiment of the present application. The user behavior prediction system includes user equipment and data processing equipment. Among them, user equipment includes smart terminals such as mobile phones, personal computers, or information processing centers. The user device is the initiator of user behavior prediction for the page. As the initiator of the user behavior prediction request, the user usually initiates the request through the user device.

The above-mentioned data processing equipment may be a cloud server, a network server, an application server, a management server, and other equipment or servers with data processing functions. The data processing device receives the user behavior prediction request from the smart terminal for the page through the interactive interface, and then performs page processing through machine learning, deep learning, search, reasoning, decision-making and other methods through the memory for storing data and the processor for data processing. The memory in the data processing device can be a general term, including local storage and a database that stores historical data. The database can be on the data processing device or on other network servers.

In the user behavior prediction system shown in Figure 2a, the user device can receive instructions from the user. For example, the user device can obtain a page input/selected by the user, and then initiate a request to the data processing device, so that the data processing device can respond to the information obtained by the user device. This page executes the user behavior prediction application to obtain the processing results for this page. For example, the user device can obtain a page input by the user, and then initiate a user behavior prediction request for the page to the data processing device, so that the data processing device processes the characteristics of each item in the page, thereby obtaining the processing result of the page. That is, the probability that each item on the page is clicked by the user.

In Figure 2a, the data processing device can execute the directed acyclic graph construction method and the user behavior prediction method of the embodiment of the present application.

Figure 2b is another schematic structural diagram of a user behavior prediction system provided by an embodiment of the present application. In Figure 2b, the user equipment itself can execute the user behavior prediction application. The user equipment can directly obtain input from the user and directly obtain input from the user equipment. The hardware itself is used for processing. The specific process is similar to Figure 2a. Please refer to the above description and will not be repeated here.

In the user behavior prediction system shown in Figure 2b, the user device can receive instructions from the user. For example, the user device can obtain a page selected by the user on the user device, and then the user device itself can target the characteristics of each item in the page. Perform processing to obtain the processing result of the page, that is, the probability of each item on the page being clicked by the user.

In Figure 2b, the user equipment itself can execute the directed acyclic graph construction method and user behavior prediction method in the embodiment of the present application.

Figure 2c is a schematic diagram of related equipment for user behavior prediction processing provided by the embodiment of the present application.

The user equipment in Figure 2a and Figure 2b can be the local device 301 or the local device 302 in Figure 2c, and the data processing device in Figure 2a can be the execution device 210 in Figure 2c, where the data storage system 250 can To store the data to be processed by the execution device 210, the data storage system 250 can be integrated on the execution device 210, or can be set up on the cloud or other network servers.

The processors in Figure 2a and Figure 2b can perform data training/machine learning/deep learning through neural network models or other models (for example, models based on support vector machines), and use the data to ultimately train or learn the model to execute on the page User behavior prediction application to obtain corresponding processing results.

Figure 3 is a schematic diagram of the architecture of the system 100 provided by the embodiment of the present application. In Figure 3, the execution device 110 is configured with an input/output (I/O) interface 112 for data interaction with external devices. The user Data can be input to the I/O interface 112 through the client device 140. In this embodiment of the present application, the input data may include: various to-be-scheduled tasks, callable resources, and other parameters.

When the execution device 110 preprocesses the input data, or when the calculation module 111 of the execution device 110 performs calculation and other related processing (such as implementing the function of the neural network in this application), the execution device 110 can call the data storage system 150 The data, codes, etc. in the system can be used for corresponding processing, and the data, instructions, etc. obtained by corresponding processing can also be stored in the data storage system 150 .

Finally, the I/O interface 112 returns the processing results to the client device 140, thereby providing them to the user.

It is worth mentioning that the training device 120 can generate corresponding target models/rules based on different training data for different goals or different tasks, and the corresponding target models/rules can be used to achieve the above goals or complete the above tasks. , thereby providing users with the desired results. The training data may be stored in the database 130 and come from training samples collected by the data collection device 160 .

In the case shown in FIG. 3 , the user can manually enter the input data, and the manual input can be operated through the interface provided by the I/O interface 112 . In another case, the client device 140 can automatically send input data to the I/O interface 112. If requiring the client device 140 to automatically send input data requires the user's authorization, the user can set corresponding permissions in the client device 140. The user can view the results output by the execution device 110 on the client device 140, and the specific presentation form may be display, sound, action, etc. The client device 140 can also be used as a data collection end to collect the input data of the input I/O interface 112 and the output results of the output I/O interface 112 as new sample data, and store them in the database 130 . Of course, it is also possible to collect without going through the client device 140. Instead, the I/O interface 112 directly uses the input data input to the I/O interface 112 and the output result of the output I/O interface 112 as a new sample as shown in the figure. The data is stored in database 130.

It is worth noting that Figure 3 is only a schematic diagram of a system architecture provided by an embodiment of the present application. The positional relationship between the devices, devices, modules, etc. shown in the figure does not constitute any limitation. For example, in Figure 3, the data The storage system 150 is an external memory relative to the execution device 110. In other cases, the data storage system 150 can also be placed in the execution device 110. As shown in Figure 3, the neural network can be trained according to the training device 120.

An embodiment of the present application also provides a chip, which includes a neural network processor NPU. The chip can be disposed in the execution device 110 as shown in FIG. 3 to complete the calculation work of the calculation module 111. The chip can also be installed in the training device 120 as shown in Figure 3 to complete the training work of the training device 120 and output the target model/rules.

Neural network processor NPU, NPU is mounted on the main central processing unit (CPU) (host CPU) as a co-processor, and the main CPU allocates tasks. The core part of the NPU is the arithmetic circuit. The controller controls the arithmetic circuit to extract the data in the memory (weight memory or input memory) and perform operations.

In some implementations, the computing circuit includes multiple processing units (PE). In some implementations, the arithmetic circuit is a two-dimensional systolic array. The arithmetic circuit may also be a one-dimensional systolic array or other electronic circuit capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit is a general-purpose matrix processor.

For example, assume there is an input matrix A, a weight matrix B, and an output matrix C. The arithmetic circuit fetches the corresponding data of matrix B from the weight memory and caches it on each PE in the arithmetic circuit. The operation circuit takes matrix A data and matrix B from the input memory to perform matrix operations, and the partial result or final result of the obtained matrix is stored in the accumulator (accumulator).

The vector calculation unit can further process the output of the arithmetic circuit, such as vector multiplication, vector addition, exponential operation, etc. Numerical operations, size comparison, etc. For example, the vector computing unit can be used for network calculations in non-convolutional/non-FC layers in neural networks, such as pooling, batch normalization, local response normalization, etc.

In some implementations, the vector computation unit can store the processed output vectors into a unified buffer. For example, the vector calculation unit may apply a nonlinear function to the output of the arithmetic circuit, such as a vector of accumulated values, to generate activation values. In some implementations, the vector computation unit generates normalized values, merged values, or both. In some implementations, the processed output vector can be used as an activation input to an arithmetic circuit, such as for use in a subsequent layer in a neural network.

Unified memory is used to store input data and output data.

The weight data directly transfers the input data in the external memory to the input memory and/or the unified memory through the storage unit access controller (direct memory access controller, DMAC), stores the weight data in the external memory into the weight memory, and transfers the weight data to the unified memory. The data in is stored in external memory.

The bus interface unit (BIU) is used to realize the interaction between the main CPU, DMAC and instruction memory through the bus.

The instruction fetch buffer connected to the controller is used to store instructions used by the controller;

The controller is used to call instructions cached in the memory to control the working process of the computing accelerator.

Generally, the unified memory, input memory, weight memory and instruction memory are all on-chip memories, and the external memory is the memory outside the NPU. The external memory can be double data rate synchronous dynamic random access memory (double data). rate synchronous dynamic random access memory (DDR SDRAM), high bandwidth memory (high bandwidth memory (HBM)) or other readable and writable memory.

Since the embodiments of the present application involve the application of a large number of neural networks, in order to facilitate understanding, the relevant terms involved in the embodiments of the present application and related concepts such as neural networks are first introduced below.

(1)Neural network

The neural network can be composed of neural units. The neural unit can refer to an arithmetic unit that takes xs and intercept 1 as input. The output of the arithmetic unit can be:

Among them, s=1, 2,...n, n is a natural number greater than 1, Ws is the weight of xs, and b is the bias of the neural unit. f is the activation function of the neural unit, which is used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into an output signal. The output signal of this activation function can be used as the input of the next convolutional layer. The activation function can be a sigmoid function. A neural network is a network formed by connecting many of the above-mentioned single neural units together, that is, the output of one neural unit can be the input of another neural unit. The input of each neural unit can be connected to the local receptive field of the previous layer to extract the features of the local receptive field. The local receptive field can be an area composed of several neural units.

The work of each layer in the neural network can be described by the mathematical expression y=a(Wx+b): From the physical level, the work of each layer in the neural network can be understood as five pairs of input spaces (input vectors) Set) operations to complete the transformation from input space to output space (i.e., row space to column space of the matrix). These five operations include: 1. Dimension raising/dimension reduction; 2. Zoom in/out; 3. Rotate; 4. Translate; 5. "Bend". Among them, the operations of 1, 2, and 3 are completed by Wx, the operation of 4 is completed by +b, and the operation of 5 is implemented by a(). The reason why the word "space" is used here is because the object to be classified is not a single thing, but a class of things. Space refers to the collection of all individuals of this type of thing. Among them, W is a weight vector, and each value in the vector represents the weight value of a neuron in the neural network of this layer. This vector W determines the spatial transformation from the input space to the output space described above, that is, the weight W of each layer controls how to transform the space. The purpose of training a neural network is to finally obtain the weight matrix of all layers of the trained neural network (a weight matrix formed by the vector W of many layers). Therefore, the training process of neural network is essentially to learn how to control spatial transformation, and more specifically, to learn the weight matrix.

Because you want the output of the neural network to be as close as possible to the value you really want to predict, you can compare the predicted value of the current network with the really desired target value, and then update each layer of the neural network based on the difference between the two. weight vector (of course, there is usually an initialization process before the first update, that is, pre-configuring parameters for each layer in the neural network). For example, if the predicted value of the network is high, adjust the weight vector to make it predict lower Some, constant adjustments are made until the neural network can predict the truly desired target value. Therefore, it is necessary to define in advance "how to compare the difference between the predicted value and the target value". This is the loss function (loss function) or objective function (objective function), which is used to measure the difference between the predicted value and the target value. Important equations. Among them, taking the loss function as an example, the higher the output value (loss) of the loss function, the greater the difference. Then the training of the neural network becomes a process of reducing this loss as much as possible.

(2)Back propagation algorithm

The neural network can use the error back propagation (BP) algorithm to modify the size of the parameters in the initial neural network model during the training process, so that the reconstruction error loss of the neural network model becomes smaller and smaller. Specifically, forward propagation of the input signal until the output will produce an error loss, and the parameters in the initial neural network model are updated by backpropagating the error loss information, so that the error loss converges. The backpropagation algorithm is a backpropagation movement dominated by error loss, aiming to obtain the optimal parameters of the neural network model, such as the weight matrix.

The method provided by this application is described below from the training side of the neural network and the application side of the neural network.

The model training method provided by the embodiment of this application involves the processing of data sequences, and can be specifically applied to data training, machine learning, deep learning and other methods. For training data (for example, the first item of the page to be processed in this application is (features, etc.) to perform symbolic and formalized intelligent information modeling, extraction, preprocessing, training, etc., and finally obtain a trained neural network (such as the target model in this application); and, provided by the embodiments of this application The user behavior prediction method can use the above-mentioned trained neural network to input input data (for example, the first feature of the first item of the target page in this application, etc.) into the trained neural network to obtain output data. (For example, in the user behavior prediction method provided by this application, the probability of the first item being clicked by the user, etc.). It should be noted that the model training method and the user behavior prediction method provided in the embodiments of this application are inventions based on the same concept, and can also be understood as two parts of a system, or two stages of an overall process: such as Model training phase and model application phase.

It is worth noting that before predicting user behavior on the target page, a directed acyclic graph for the target page can be constructed first. The construction process of the directed acyclic graph is introduced below. Figure 4 is a schematic flow chart of a directed acyclic graph construction method provided by an embodiment of the present application. As shown in Figure 4, the method includes:

401. Obtain the eye movement data of the user browsing the target page.

In this embodiment, when it is necessary to predict user behavior on the target page, the target page can be obtained first. The target page contains multiple lists. For any list, the list contains multiple items, and the multiple items are arranged according to a certain Order Arrange (for example, if multiple items are arranged in a row, the list is a horizontal list, if multiple items are arranged in a column, the list is a vertical list). For example, as shown in Figure 5 (Figure 5 is a schematic diagram of a target page provided by an embodiment of the present application), assume that the target page is a display page of an application mall. This page can display information about multiple applications to the user, so that the user can Browse and download the applications you need on this page. This page contains vertical list B ₁ , vertical list B ₃ , vertical list B ₅ , horizontal list B ₂ and horizontal list B ₄ . These three vertical lists and two horizontal lists are staggered, that is, according to the vertical list B ₁ and the horizontal list B _2, vertical list B ₃ , horizontal list B ₄ , and vertical list B ₅ are arranged in this order. Among them, horizontal lists B ₂ and B ₄ contain 5 applications, and vertical lists B ₁ and B ₃ contain 3 applications. In this way, the page can display information about 19 applications to users.

After obtaining the target page, you can invite at least one user to browse the target page, and obtain the eye movement data generated by these users when browsing the target page.

Specifically, the eye movement data of users browsing the target page can be obtained through the following methods:

As shown in Figure 6 (Figure 6 is a schematic diagram of an eye tracker provided by an embodiment of the present application), a user-oriented eye tracker can be deployed near the user equipment used to display the target page, and the eye tracker can be connected to the user equipment. Electrical connection. In addition, auxiliary tools can also be deployed. This auxiliary tool is used to stabilize the user's head so that the eye tracker can accurately track the user's line of sight. In order to better simulate the interactive environment between the user and the user device, it can be set The distance between the user and the screen of the user equipment (for example, the distance can be set to 30-40cm, etc.), and the tilt angle of the user equipment (for example, the tilt angle can be set to 65°-70°, etc.).

When the user starts browsing the target page displayed on the user device with the help of assistive tools, the eye tracker can track and record the user's gaze position and gaze movement on the target page, generate eye movement data of the user browsing the target page, and send it to the user device . In this way, the user device successfully obtains the eye movement data of the user browsing the target page.

It should be understood that the user equipment here can be the user equipment in the system shown in Figure 2a or Figure 2b. Then, after obtaining the eye movement data of the user browsing the target page, the user equipment can analyze and construct the eye movement data on its own. For the directed acyclic graph of the target page, the eye movement data can also be sent to the data processing device in the system shown in Figure 2a or Figure 2b, so that the data processing equipment can analyze the eye movement data and build an effective target page. Towards acyclic graph, we will not go into details later.

402. Based on the eye movement data, determine the user's browsing behavior for multiple items, and the multiple items are located in multiple lists on the target page.

After obtaining the eye movement data of the user browsing the target page, the eye movement data can be analyzed to obtain the user's browsing behavior of multiple items in the target page. These multiple items are located in multiple lists on the target page, so there are many items can be expressed as a set i=[i _1,1 ,i _1,2 ,..., _io,q ], where i _t,j represents the j-th item in the t-th list in the target page, t=1,...,o, j=1,...,q.

Specifically, after obtaining the eye movement data of the user browsing the target page, the following analysis can be performed based on the eye movement data:

(1) Take any three items with consecutive positions in set i, and call them item A, item B and item C respectively. In the target page, item A is before item B and item A is adjacent to item B (item A and item B can be two items located in two adjacent lists, that is, t _A < t _B , for example, item A is the last item of the first list on the target page, item B is the first item of the second list on the target page, etc. Item A and item B can also be two items in the same list, and The position of item A is higher, that is, t _A = t _B , and j _A < j _B . For example, item A is the first item of the first list on the target page, and item B is the first list on the target page. the 2nd item and so on), item B is before item C and item B is adjacent to item C.

Based on the eye movement data, the user's browsing order of Item A, Item B and Item C can be counted. The statistical results are shown in Table 1:

Table 1

Based on Table 1, it can be seen that the browsing order of item A→item B→item C accounts for the largest proportion, indicating that when users browse multiple items on the target page, they mainly follow the order (sorting) of these multiple items on the target page. To browse, this browsing behavior can be called sequential examination. Sequential browsing behavior includes two categories, which will be introduced separately below:

(1.1) The first type of sequential browsing behavior refers to any list on the target page. If the list is a horizontal list, all items in the list will be browsed in order from left to right. If the list is vertical list, browse all items in the list in order from top to bottom. Still using the example shown in Figure 5, for list B ₂ , the first type of sequential browsing behavior is: in the order of i _2,1 →i _2,2 →i _2,3 →i _2,4 →i _2,5 , Let's browse the five items i _2,1 , i _2,2 , i _2,3 , i _2,4 and i _2,5 in list B ₂ . For list B ₁ , the first type of sequential browsing behavior is: browsing i _1,1 , i _1,2 and i ₁ in list B ₁ in the order of i 1,1 → _i _1,2 →i _1,3 ₃ these 3 items.

(1.2) The second type of sequential browsing behavior means that for two adjacent lists on the target page, the adjacent items in the two lists are browsed in the order of front and back. It should be noted that if among the two lists, the former list is a vertical list and the latter list is a horizontal list, the adjacent items in the two lists include the last item in the vertical list and all the items in the horizontal list, that is, horizontal All items of a list can be considered items adjacent to the last item of the vertical list. If among the two lists, the former list is a horizontal list and the latter list is a vertical list, the adjacent items in the two lists include the first item of the vertical list and all the items of the horizontal list, that is, all the items of the horizontal list. Both can be considered as items adjacent to the first item in the vertical list. Still using the example shown in Figure 5, for lists B ₁ and B ₂ , the second type of sequential browsing behavior is: i _1,3 →i _2,1 , i _1,3 →i _2,2 , i _1,3 →i _2,3 , i _1,3 →i _2,4 , i _1,3 →i _2,5 in the order to browse i _1,3 , i _2,1 , i ₂ in lists B ₁ and B ₂ , ₂ , i _2,3 , i _2,4 and i _2,5 are six adjacent items. For lists B ₂ and B ₃ , the second type of sequential browsing behavior is: i _2,1 →i _3,1 , i _2,2 →i _3,1 , i _2,3 →i _3,1 , i _{2, 4} →i _3,1 , i _2,5 →i _3,1 in the order to browse i _2,1 , i _2,2 , i _2,3 , i _2,4 , i ₂ in lists B ₂ and B ₃ _,5 and i _3,1 are six adjacent items.

It should be understood that if the two adjacent lists are both horizontal lists, then for any item in the former list, all items in the latter list can be regarded as items adjacent to the item, and similarly , for any item in the latter list, all items in the previous list can be regarded as items adjacent to the item.

(2) When browsing the target page, the user may jump directly from the current list to another list for browsing, and the current list and the other list are separated by at least one list. For ease of introduction, the current list is called list D below, and the other list is called list E. List D and list E are separated by at least one list. The list skip length between list D and list E is defined. )l=t _D -t _E , for example, when list D is the second list in the target page and list D is the fourth list in the target page, then l=4-2=2. Based on eye movement data, different list skip lengths can be counted. The statistical results are shown in Table 2:

Table 2

Based on Table 2, it can be seen that the browsing method with a list skip length of 2 accounts for the largest proportion, indicating that in addition to sequential browsing behavior, users also often send behaviors of skipping an entire list and browsing the next list directly. This browsing behavior can This is called block skip. It is worth noting that if the target page is a page with multiple horizontal lists (also called horizontal blocks) and multiple vertical lists (also called vertical blocks) staggered (i.e. F-type page), the skipping behavior with a list skip length of 2 includes two major types of behavior. The first type of skipping behavior refers to jumping from a horizontal list to a horizontal list, and the second type of skipping behavior refers to jumping from a vertical list to a vertical list, where , almost all skipping behaviors are from vertical lists to vertical lists, accounting for 94.5%, while jumping from horizontal lists to horizontal lists only account for 5.5%, indicating that users are more inclined to jump from vertical lists to vertical lists. .

Based on this, we can further calculate which item of the two lists the skipping behavior of jumping from the vertical list to the vertical list mainly occurs on, that is, count the starting item and ending item of this type of skipping behavior, and count The results are shown in Table 3:

table 3

Based on Table 3, it can be seen that the maximum probability of the starting item of the user's skipping behavior is the last item in a certain list, and the maximum probability of the ending item of the user's skipping behavior is the first item of another list.

Then, based on the above analysis, the user's skipping behavior can be summarized as: in two non-adjacent lists (these two lists are separated by one list), the user jumps from the last item of the previous list to the next one. Browsing continues with the first item in the list, and the browsing sequence between these two items can be called a jump sequence. Still using the example shown in Figure 5, when the user browses i _1,3 , he directly skips B ₂ and browses i _3,1 .

(3) Based on Table 1, it can be seen that in addition to sequential browsing behaviors, the largest proportion of non-sequential browsing behaviors are browsing behaviors in the order of B→A→B and browsing in the order of A→B→A. Behavior indicates that users tend to browse two adjacent items repeatedly to compare between items. This browsing behavior can be called comparison behaviors.

It can be seen that after analyzing the statistical data, the user's browsing behavior for multiple items in the target page can be determined, including sequential browsing behavior, skipping behavior, comparison behavior, etc.

403. Based on browsing behavior, connect multiple items to obtain a directed acyclic graph.

After obtaining the user's browsing behavior for multiple items in the target page, based on these browsing behaviors, multiple items in the target page can be connected to obtain a directed acyclic graph for the target page.

Specifically, these browsing behaviors include two major categories of browsing behaviors. The first type of browsing behavior refers to the user browsing items in the same list, including the aforementioned first type of sequential browsing behavior. Therefore, the user's browsing order in the same list can be called the first order, and the first order includes the first type of sequential browsing. Behavior, the order in which the user browses all items in the same list, from top to bottom and from left to right. The second type of browsing behavior refers to the user browsing items between different lists, including the aforementioned second type of sequential browsing behavior and comparison behavior. Therefore, the user's browsing order between different lists can be called the second order, and the second order includes In the second type of sequential browsing behavior, the user browses several adjacent items in two adjacent lists in the order in which they browse, and in contrast behavior, the user browses two items in two non-adjacent lists according to the jump. Turn order. Then, the directed acyclic graph for the target page can be obtained in the following way:

(1) In the target page, connect the items in the same list that the user browsed in the first order, that is, for any list in the target page, if the list is a horizontal list, press from left to right Connect all the items in the list in order from right to right. If the list is a vertical list, connect all the items in the list in order from top to bottom. In this way, you can complete the internal content of each list on the target page. connect. For example, as shown in Figure 7 (Figure 7 is a schematic diagram of a directed acyclic graph provided by the embodiment of the present application, and Figure 7 is drawn based on Figure 5), for list B ₁ , it can be calculated according to i _1, The order ₁ →i _1,2 →i _1,3 is used to connect the three items i _1,1 , i _1,2 and i _1,3 in list B ₁ . For list B ₂ , i 2,1 , _i 2,2 , in list B ₂ can be connected in the order _of i _2,1 →i _2,2 →i _2,3 →i _2,4 →i _2,5 There are five items: i _2,3 , i _2,4 and i _2,5 . The same is true for lists B ₃ , B ₄ and B ₅ , which will not be repeated here. In this way, the internal connections of these five lists in the target page are completed.

(2) Connect the items in different lists browsed by the user in the second order to obtain a directed acyclic graph. That is, for two adjacent lists in the target page, the items in the two lists can be connected according to the second order. The order of adjacent items is used to connect these adjacent items, and for two non-adjacent lists in the target page (one list is separated between the two lists), the user can browse the two lists according to the Jump sequence to connect the last item of the previous list and the first item of the next list in the two lists. In this way, the connection between the lists in the target page can be completed. Understand, directed acyclic picture. Still using the example shown in Figure 7, for lists B ₁ and B ₂ , i _1,3 →i _2,1 , i _1,3 →i _2,2 , i _1,3 →i _2,3 , i In the order of _1,3 →i _2,4 and i _1,3 →i _2,5 , connect i _1,3 to i _2,1 , connect i _1,3 to i _2,2 , connect i _1,3 Connect with i _2,3 , connect i _1,3 with i _2,4 , connect i _1,3 with i _2,5 , for lists B ₂ and B ₃ , lists B ₃ and B ₄ , lists B ₄ and The same is true for B ₅ , which will not be described again here. Furthermore, for lists B ₁ and B ₃ , i _1,3 can be connected to i _3,1 in the order of i _1,3 →i _3,1 , and the same is true for lists B ₃ and B ₅ , this No further details will be given. In this way, a directed acyclic graph for the target page can be obtained.

In the embodiment of the present application, the user's interest in the target page can be determined based on the eye movement data generated when the user browses the target page. The browsing behavior of multiple items in the list, then these browsing behaviors (for example, sequential browsing behavior and skipping behavior) often determine the user's browsing order of items (for example, the user's browsing order in the same list and the user's browsing order in different items). Browsing order between lists), thereby connecting multiple items of the target page according to these browses, and obtaining a directed acyclic graph for the target page. This directed acyclic graph can be used in subsequent user behavior predictions for the target page. Since this directed acyclic graph involves users' complex and diverse browsing behaviors, it is helpful to improve the accuracy of predicting user behavior on the target page.

The above is a detailed description of the directed acyclic graph construction method provided by the embodiment of the present application. The user behavior prediction method provided by the embodiment of the present application will be introduced below. Figure 8 is a schematic flow chart of a user behavior prediction method provided by an embodiment of the present application. As shown in Figure 8, the method includes:

801. Obtain the first characteristic of the first item and the second characteristic of the second item through the target model. The first item and the second item are located in different lists or the same list on the target page, and the second item is located before the first item.

In this embodiment, when the target page is required to predict user behavior, that is, to obtain the probability that each item in the target page is clicked by the user, the first feature of each item in the target page can be extracted first. It should be noted that for any item , the first characteristic of the project refers to the attribute information of the project itself. For example, when the project is an application on the page of the application mall, the first characteristic of the application may include the developer of the application, the size of the application , the type of the application, the icon of the application, etc., when the item is a product in the shopping mall page, the first feature of the product may include the price of the product, the type of the product, the color of the product etc.

It is worth noting that since the target page contains multiple items, multiple rounds of operations can be performed on the target page. One round operates on one item in the target page (that is, steps 801 to 801 are performed once in each round). Step 805, that is, steps 801 to 805 will be executed for each item. Since one round of operation can obtain the probability of an item being clicked by the user, after completing all rounds, the probability of all items in the target page being clicked can be obtained. The probability of a user clicking. Based on this, this embodiment makes a schematic introduction using one of the items in the target page, and calls this item the first item.

Then, when it is necessary to estimate the probability that the first item is clicked by the user, the first feature of the first item and the second feature of the second item can be input to the target model (trained neural network model). Among them, the first item and the second item are located in different lists or the same list on the target page, and the second item is located before the first item, that is, the positional relationship between the first item and the second item exists in the following two situations: (1) The first item and the second item may be items in the same list, the second item is located before the first item and the second item is adjacent to the first item. (2) The first item and the second item can be items in different lists. The list where the second item is located is located before the list where the first item is located. The second item can be adjacent to the first item or not. One item is adjacent.

Since all the items in the target page can form a directed acyclic graph, any one of these items can be regarded as a node in the directed acyclic graph, and there is a one-way connection between the nodes in the directed acyclic graph. connection relationship. In this way, the directed acyclic graph has parent nodes and child nodes, and the connection direction between the parent node and the child node is from the parent node to the child node. Then, the aforementioned first item can be regarded as a child node in the directed acyclic graph for the target page, and the second item is all the parent nodes of the child node. Still as shown in Figure 7, when the first When the item is i _1,2 , the second item is i _1,1 . When the first item is i _2,1 , the second item is i _1,3 . When the first item is i _2,2 , the second item is i _2,1 and i _1,3 . When the first item is i _3,1 , the second item is i _2,1 , i _2,2 , i _2,3 , i _2,4 , i _2,5 , i _1,3 and so on.

It should be understood that the acquisition process of the second feature of the second item may refer to the subsequent acquisition process of the second feature of the first item, and will not be described again here.

It should also be understood that when the first item is the first item in the target page (for example, i _1,1 in the example shown in Figure 7), then There is no second item. At this time, the second characteristic of the second item can be understood as a preset value (the size of the preset value can be set according to actual needs, and there is no limit here). Therefore, the first item can also be The first characteristic of the project and this preset value are input into the target model.

802. Obtain the second feature of the first item based on the first feature of the first item and the second feature of the second item through the target model.

After inputting the first feature of the first item and the second feature of the second item into the target model, the first feature of the first item and the second feature of the second item can be processed by the target model to obtain the first feature of the first item. Second characteristic.

Specifically, the second feature of the first item can be obtained in the following way:

(1) Extract the user's request for the target page and the probability that the second item is clicked by the user. The user's request for the target page can include keywords entered by the user on the target page to search for certain items, etc., second The probability that an item is clicked by the user can be understood as the probability that the item targeted in the previous round (that is, the item before the first item) is clicked by the user.

(2) After the first feature of the first item and the second feature of the second item have been input into the target model, the user's request for the target page and the probability that the second item is clicked by the user can also be input into the target model. , so that the target model maps the first feature of the first item, the user's request for the target page, and the probability of the second item being clicked by the user on the latent space (i.e., the aforementioned mapping process), and accordingly obtains the first item's The sixth characteristic, the seventh characteristic of the first item and the eighth characteristic of the first item are then spliced together with the sixth characteristic of the first item, the seventh characteristic of the first item and the eighth characteristic of the first item (that is, the aforementioned second fusion process) to obtain the fourth feature of the first item. At the same time, the target model can also process the second feature of the second item based on the self-attention mechanism to obtain the fifth feature of the first item.

For example, in the example shown in Figure 9 (Figure 9 is a schematic structural diagram of the target model provided by the embodiment of the present application), assume that the first item is i _t,j , and the set of second items is P _t,j , and the set The k-th second item in is i _k (k=1,...,n), the first feature of the first item is I, the second feature of the k-th second item is h _k , the user’s target page The request is Q, and the probability of the second item being clicked by the user is C.

After inputting the first feature I of the first item, the second feature h ₁ ,..., h _n of the second item, the user's request Q for the target page, and the probability C of the second item being clicked by the user into the target model, the target The model can map the first feature I of the first item on the latent space to obtain the sixth feature V _I of the first item, and map the user's request Q for the target page on the latent space to obtain the seventh feature of the first item. V _Q , map the probability C of the second item clicked by the user on the latent space, and obtain the eighth feature V _C of the first item. Then, the target model can splice the sixth feature V _I of the first item, the seventh feature V _Q of the first item, and the eighth feature V _C of the first item to obtain the fourth feature x _t,j of the first item. .

At the same time, the target model can also use the self-attention mechanism to calculate the second features h ₁ ,..., h _n of the second item, and obtain the fifth feature e _t,j of the first item. Among them, the calculation based on the self-attention mechanism is as shown in the following formula:

(3) Obtain the fourth feature of the first item and the fifth feature of the first item. The target model can use recurrent neural units (GRUcell) processes the fourth feature of the first item and the fifth feature of the first item (ie, the aforementioned first fusion process) to obtain the second feature of the first item.

Still using the example shown in Figure 9, the fourth feature x _t,j of the first item and the fifth feature e _t,j of the first item are obtained. The target model can input these two features into the recurrent neural unit for processing. Get the second feature h _t,j of the first item. Among them, the processing implemented by the recurrent neural unit is as shown in the following formula:
h _t,j =GRUcell(x _t,j ,e _t,j ) (3)

It can be understood that the second characteristic of the first item can represent the impact of the second item on the first item (it can also be understood as the relationship between the second item and the first item), that is, the user's sequential browsing behavior and When skipping behaviors and browsing to the first item, the impact of the items browsed by the user on the first item while performing these behaviors.

It should be understood that in the process of obtaining the fourth feature of the first feature, the user's request for the target page and the probability of the second item being clicked by the user may not be input to the target model, so that the target model directly obtains the first feature of the first item. After continuing the Ning mapping process, the fourth feature of the first feature is obtained.

803. Obtain the first characteristic of the third item through the target model. The first item and the third item are located in different lists or the same list on the target page, and the third item is adjacent to the first item.

In addition, the first feature of the third item can also be input to the target model, wherein the first item and the third item are located in different lists or the same list on the target page, and the third item is adjacent to the first item, that is, the third item is adjacent to the first item. There are two situations in the positional relationship between the first item and the third item: (1) The first item and the third item can be items in the same list, and the third item and the first item are adjacent. (2) The first item and the third item can be items in different lists, and the third item and the first item are adjacent.

Still as shown in the example in Figure 7, when the first item is i _1,2 , the third item is i _1,1 . When the first item is i _1,3 , the third item is i _1,2 and i _2,1 . When the first item is i _2,1 , the second item is i _1,3 , i _2,2 , i _3,1 and so on.

804. Obtain the third feature of the first item based on the first feature of the first item and the first feature of the third item through the target model.

After the first feature of the third item is input into the target model, the first feature of the first item and the first feature of the third item can be processed by the target model, thereby obtaining the third feature of the first item.

Specifically, the third feature of the first item can be obtained in the following way:

(1) The target model maps the first feature of the first item, the user's request for the target page, and the first feature of the third item on the latent space respectively (i.e., the aforementioned mapping process), and accordingly obtains the first feature of the first item. Six characteristics, the seventh characteristic of the first item and the ninth characteristic of the first item.

Still using the example shown in Figure 9, let the set of third items be N _t,j , the f-th third item in the set is if ( _f =1,...,m), the f-th third item in the set is The first characteristic of the project is _If .

After inputting the first feature I ₁ ,..., I _m of the third item into the target model, the target model can map the first feature I of the first item on the latent space to obtain the sixth feature V _I of the first item. , map the user's request Q for the target page on the latent space, and obtain the seventh feature V _Q of the first item, and map the first feature I ₁ ,..., I _m of the third item on the latent space, and obtain The ninth characteristic V ₁ , ..., V _m of the first item.

(2) Then, the target model can perform the comparison function on the sixth feature of the first item and the ninth feature of the first item. Calculate, and then perform weighted summation based on the calculation results (the calculation of the comparison function and the weighted summation calculation are the aforementioned third fusion process) to obtain the tenth feature of the first item.

Still using the example shown in Figure 9, after obtaining the sixth feature V _I of the first item and the ninth feature V ₁ ,..., V _m of the first item, these features can be calculated and summed based on the contrast function g. Weighted summation calculation is performed to obtain the tenth feature d _t,j of the first item. Among them, the calculation process is as shown in the following formula:

Among them, the comparison function g can be one of the following three functions: inner product function neural network function kernel function

(3) Finally, the target model can perform an exclusive OR operation (i.e., the aforementioned fourth fusion process) on the sixth feature of the first item, the seventh feature of the first item, and the tenth feature of the first item to obtain the first item The third characteristic.

Still using the example shown in Figure 9, to obtain the tenth feature d _t,j of the first item, the sixth feature V _I of the first item, the seventh feature V _Q of the first item, and the tenth feature of the first item can be The features d _t,j are subjected to an exclusive OR operation to obtain the third feature cp _t,j of the first item. Among them, the operation process is shown in the following formula:
cp _{t, j} = d _{t, j} ⊙V _I ⊙V _Q (5)

It can be understood that the third characteristic of the first item can represent the impact of the third item on the first item (it can also be understood as the relationship between the third item and the first item), that is, the user uses the comparison behavior to browse to The first item refers to the impact of the items browsed by the user on the first item during the behavior.

It should be understood that when obtaining the third feature of the first item, the user's request for the target page may not be input to the target model. Therefore, the target model may only perform the third feature on the sixth feature of the first item and the tenth feature of the first item. Four fusion processes are performed to obtain the third feature of the first item.

805. Use the target model to obtain the probability that the first item is clicked by the user based on the second feature of the first item and the third feature of the first item.

After obtaining the second feature of the first item and the third feature of the first item, the target model can calculate the second feature of the first item and the third feature of the first item, thereby obtaining the probability that the first item is clicked by the user. .

Still using the example shown in Figure 9, after obtaining the second feature h _t,j of the first item and the third feature cp _t,j of the first item, the two features can be calculated to obtain the first item that is used by the user. Probability of click C _t,j . Among them, the calculation process is as shown in the following formula:

In the same way, for the remaining items in the target page other than the first item, the same operations as those performed on the first item can also be performed. Therefore, the probability of all items in the target page being clicked by the user can be obtained, thereby completing the analysis of the target page. User behavior prediction.

It should be understood that when obtaining the probability that the first item is clicked by the user, steps 803 and 804 may not be performed, so that the target model directly calculates the second feature of the first item to obtain the probability that the first item is clicked by the user.

In addition, the prediction results of the target model provided by the embodiment of the present application can also be compared with the prediction results of the model of related technologies. The comparison results are shown in Table 4:

Table 4

Based on Table 4, it can be seen that the prediction ability displayed by the target model provided by the embodiment of the present application is significantly improved in both indicators compared with the model provided by the related technology.

In the embodiment of the present application, when it is necessary to predict the probability that the first item in the target page is clicked by the user, the first feature of the first item and the second feature of the second item can be input to the target model, where the first item and the second feature can be input to the target model. The two items are in different lists or in the same list on the target page, and the second item is before the first item. Then, the target model can obtain the second feature of the first item based on the first feature of the first item and the second feature of the second item, and then based on the second feature of the first item Feature, obtain the probability that the first item is clicked by the user. In the aforementioned process, when obtaining the probability that the first item is clicked by the user, the target model considers the impact of the second item before the first item on the first item. Since the second item can not only be the list where the first item is located, The items in can also be items in other lists. Therefore, the factors considered by the target model are relatively comprehensive and can fit the actual situation when the user browses to the first item in the target page. Therefore, the first item finally obtained by the target model The probability of being clicked by the user has a high accuracy, which is conducive to accurately recommending items of interest to the user in the future.

Furthermore, the target model provided by the embodiments of this application not only introduces conventional sequential browsing behaviors, but also introduces browsing behaviors such as jump behaviors and comparison behaviors. In other words, the target model will consider the user's use of these complex and diverse browsing behaviors. When the user browses to the first item, the impact of the item browsed by the user on the first item during the behavior can further improve the accuracy of the probability of the first item being clicked by the user finally obtained by the target model.

Furthermore, when analyzing the first item, the target model provided by the embodiment of the present application not only takes into account the influence of the attribute information of the first item itself, but also considers the user's request for the target page and the second item clicked by the user. The influence of external factors such as probability can further improve the accuracy of the probability of the first item being clicked by the user finally obtained by the target model.

The above is a detailed description of the user behavior prediction method provided by the embodiment of the present application. The model training method provided by the embodiment of the present application will be introduced below. Figure 10 is a schematic flow chart of the model training method provided by the embodiment of the present application. As shown in Figure 10, the method includes:

1001. Obtain the first feature of the first item and the second feature of the second item through the model to be trained. The first item and the second item are located in different lists or the same list of the page to be processed, and the second item is located in the first item. Before.

In this embodiment, when the model to be trained needs to be trained, a batch of training data can be obtained first. The batch of training data includes pages to be processed. The pages to be processed include multiple lists, and each list contains at least one item. It is worth noting that in the page to be processed, the true probability of any item being clicked by the user is known.

It should be noted that, regarding the first item, the second item, the first feature of the first item, and the second feature of the second item of the page to be processed, reference may be made to the third item of the target page in step 801 in the embodiment shown in FIG. 8 The relevant descriptions of the first item, the second item, the first feature of the first item, and the second feature of the second item will not be described again here.

It can be understood that the first feature of the first item is the attribute information of the first item, the second feature of the first item is the information obtained by fusion based on the attribute information of the first item, and the second feature of the second item is the information obtained based on the fusion of the attribute information of the first item. Information obtained by fusing the attribute information of the second item (that is, the first feature of the second item).

1002. Obtain the second feature of the first item based on the first feature of the first item and the second feature of the second item through the model to be trained.

After inputting the first feature of the first item and the second feature of the second item into the model to be trained, the first feature of the first item and the second feature of the second item can be processed by the model to be trained, thereby obtaining the first Secondary characteristics of the project.

In a possible implementation, the first feature of the first item is mapped to obtain the fourth feature of the first item: the first feature of the first item, the user's request for the page to be processed, and the second item being processed. The probability of a user clicking on Perform mapping processing to obtain the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item;

Perform a second fusion process on the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item to obtain the fourth feature of the first item.

It should be noted that for the introduction of step 1002, reference may be made to the relevant description of step 802 in the embodiment shown in FIG. 8 , which will not be described again here.

1003. Obtain the first feature of the third item through the model to be trained. The first item and the third item are located in different lists or the same list on the page to be processed, and the third item is adjacent to the first item.

In this embodiment, the first feature of the third item can also be input to the target model. It should be noted that, regarding the third item of the page to be processed and the first feature of the third item, reference can be made to the embodiment shown in Figure 8 The relevant description of the third item of the target page and the first feature of the third item in step 803 will not be described again here.

1004. Obtain the third feature of the first item based on the first feature of the first item and the first feature of the third item through the model to be trained.

After the first feature of the third item is input into the model to be trained, the first feature of the first item and the first feature of the third item can be processed by the model to be trained, thereby obtaining the third feature of the first item.

It should be noted that for the introduction of step 1004, reference may be made to the relevant description of step 804 in the embodiment shown in FIG. 8 and will not be described again here.

1005. Obtain the probability that the first item is clicked by the user based on the second feature of the first item and the third feature of the first item through the model to be trained.

After obtaining the second feature of the first item and the third feature of the first item, the second feature of the first item and the third feature of the first item can be processed by the model to be trained, thereby obtaining that the first item was clicked by the user. The probability (can also be called the predicted probability that the first item is clicked by the user).

It should be noted that for the introduction of step 1005, reference may be made to the relevant description of step 805 in the embodiment shown in FIG. 8 , which will not be described again here.

1006. Based on the probability of the first item being clicked by the user and the real probability of the first item being clicked by the user, obtain the target loss. The target loss is used to indicate the probability of the first item being clicked by the user and the real probability of the first item being clicked by the user. difference between.

After obtaining the predicted probability of the first item being clicked by the user, since the real probability of the first item being clicked by the user is known, the predicted probability of the first item being clicked by the user and the predicted probability of the first item being clicked by the user can be calculated through the preset target loss function The true probability of clicking is calculated to obtain the target loss. The target loss is used to indicate the predicted probability of the first item being clicked by the user and the predicted probability of the first item being clicked. The difference between the true probability of an item being clicked by the user.

1007. Based on the target loss, update the parameters of the model to be trained until the model training conditions are met and the target model is obtained.

After obtaining the target loss, the parameters of the model to be trained can be updated based on the target loss, and the next batch of training data can be obtained, and the next batch of training data can be used to continue training the model to be trained after the updated parameters (i.e., re-execute steps 1001 to 1001). 1007), until the model training conditions are met (for example, the target loss reaches convergence, etc.), the target model in the embodiment shown in Figure 8 can be obtained.

The target model trained in the embodiment of this application has the ability to predict user behavior on the page. When it is necessary to predict the probability that the first item in the target page is clicked by the user, the first feature of the first item and the second feature of the second item can be input to the target model, where the first item and the second item are located on the target page. Different lists or the same list, with the second item before the first item. Then, the target model can obtain the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item, and then obtain the probability that the first item is clicked by the user based on the second characteristic of the first item. . In the aforementioned process, when obtaining the probability that the first item is clicked by the user, the target model considers the impact of the second item before the first item on the first item. Since the second item can not only be the list where the first item is located, The items in can also be items in other lists. Therefore, the factors considered by the target model are relatively comprehensive and can fit the actual situation when the user browses to the first item in the target page. Therefore, the first item finally obtained by the target model The probability of being clicked by the user has a high accuracy, which is conducive to accurately recommending items of interest to the user in the future.

The above is a detailed description of the model training method provided by the embodiment of the present application. The device and equipment provided by the embodiment of the present application will be introduced below. Figure 11 is a schematic structural diagram of a user behavior prediction device provided by an embodiment of the present application. As shown in Figure 11, the device includes:

The first acquisition module 1101 is used to acquire the first characteristics of the first item and the second characteristics of the second item through the target model. The first item and the second item are located in different lists or the same list of the target page, and the second item Located before the first item;

The second acquisition module 1102 is configured to acquire the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item through the target model, where the first characteristic of the first item is the first item. Attribute information of Information obtained by fusion;

The third acquisition module 1103 is configured to acquire the probability that the first item is clicked by the user based on the second feature of the first item through the target model.

In a possible implementation, the device further includes: a fourth acquisition module, configured to acquire the first feature of the third item through the target model, where the first item and the third item are located in different lists or the same list of the target page. , and the third item is adjacent to the first item; the fifth acquisition module is used to obtain the third feature of the first item based on the first feature of the first item and the first feature of the third item through the target model; the third The acquisition module 1103 is configured to obtain the probability that the first item is clicked by the user based on the second feature of the first item and the third feature of the first item through the target model.

In a possible implementation, the second acquisition module 1102 is configured to: map the first feature of the first item through the target model to obtain the fourth feature of the first item; map the first feature of the second item through the target model. The second feature is processed based on the self-attention mechanism to obtain the fifth feature of the first item; the fourth feature of the first item and the fifth feature of the first item are first fused through the target model to obtain the first item the second characteristic.

In a possible implementation, the second acquisition module 1102 is configured to perform mapping processing on the first feature of the first item, the user's request for the target page, and the probability that the second item is clicked by the user through the target model, to obtain The sixth characteristic of the first item, the seventh characteristic of the first item, and the eighth characteristic of the first item; through the target model, the sixth characteristic of the first item, the seventh characteristic of the first item, and the eighth characteristic of the first item are The features undergo a second fusion process to obtain the fourth feature of the first item.

In a possible implementation, if the first item is the first item in the target page, the second characteristic of the second item is a preset value.

Figure 12 is a schematic structural diagram of a directed acyclic graph construction device provided by an embodiment of the present application. As shown in Figure 12, the device includes:

The acquisition module 1201 is used to obtain the eye movement data of the user browsing the target page;

The determination module 1202 is used to determine the user's browsing behavior for multiple items based on eye movement data, and the multiple items are located in multiple lists on the target page;

The connection module 1203 is used to connect multiple items based on browsing behavior to obtain a directed acyclic graph.

In the embodiment of the present application, the user's browsing behavior for multiple items in the target page can be determined based on the eye movement data generated when the user browses the target page. Then, these browsing behaviors (for example, sequential browsing behavior and skipping behavior) , often determines the user's browsing order of items (for example, the user's browsing order in the same list and the user's browsing order between different lists), thereby connecting multiple items of the target page according to these browsing, and obtaining the target page A directed acyclic graph, which can be used in the subsequent prediction of user behavior on the target page. Since the directed acyclic graph involves users' complex and diverse browsing behaviors, it is conducive to improving user behavior on the target page. Prediction accuracy.

In a possible implementation, the connection module 1203 is used to connect items in the same list that the user browses in the first order, and connect items in different lists that the user browses in the second order. , connect in the second order to obtain a directed acyclic graph.

In a possible implementation, the acquisition module 1201 is configured to collect eye movement data of the user browsing the target page through an eye tracker.

Figure 13 is a schematic structural diagram of a model training device provided by an embodiment of the present application. As shown in Figure 13, the device includes:

The first acquisition module 1301 is used to acquire the first feature of the first item and the second feature of the second item through the model to be trained. The first item and the second item are located in different lists or the same list of the page to be processed, and the first item The second item precedes the first item;

The second acquisition module 1302 is configured to acquire the second feature of the first item based on the first feature of the first item and the second feature of the second item through the model to be trained, where the first feature of the first item is the first Attribute information of the item, the second feature of the first item is the information obtained by fusion based on the attribute information of the first item, and the second feature of the second item is based on the attribute information of the second item (i.e., the first feature of the second item ) information obtained by fusion;

The third acquisition module 1303 is used to obtain the probability that the first item is clicked by the user based on the second feature of the first item through the model to be trained;

The fourth acquisition module 1304 is used to obtain the target loss based on the probability that the first item is clicked by the user and the real probability that the first item is clicked by the user. The target loss is used to indicate the probability that the first item is clicked by the user and the real probability that the first item is clicked by the user. The difference between the true probability of a user clicking;

The update module 1305 is used to update the parameters of the model to be trained based on the target loss until the model training conditions are met and the target model is obtained.

In a possible implementation, the device includes: a fifth acquisition module, configured to acquire the first feature of the third item through the model to be trained, and the first item and the third item are located in different lists or the same list of the page to be processed. , and the third item is adjacent to the first item; the sixth acquisition module is used to obtain the third feature of the first item based on the first feature of the first item and the first feature of the third item through the model to be trained; The third acquisition module 1303 is configured to obtain the probability that the first item is clicked by the user based on the second feature of the first item and the third feature of the first item through the model to be trained.

In a possible implementation, the second acquisition module 1302 is configured to: perform mapping processing on the first feature of the first item to obtain the fourth feature of the first item; perform self-based mapping on the second feature of the second item. Attention mechanism processing, got to the fifth feature of the first item; perform a first fusion process on the fourth feature of the first item and the fifth feature of the first item to obtain the second feature of the first item.

In a possible implementation, the second acquisition module 1302 is configured to perform mapping processing on the first feature of the first item, the user's request for the page to be processed, and the probability that the second item is clicked by the user, to obtain the first item. The sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item; performing a second fusion process on the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item , get the fourth feature of the first item.

It should be noted that the information interaction, execution process, etc. between the modules/units of the above-mentioned device are based on the same concept as the method embodiments of the present application, and the technical effects they bring are the same as those of the method embodiments of the present application. The specific content can be Refer to the description in the method embodiments shown above in the embodiments of the present application, which will not be described again here.

The embodiment of the present application also relates to an execution device. Figure 14 is a schematic structural diagram of the execution device provided by the embodiment of the present application. As shown in Figure 14, the execution device 1400 can be embodied as a mobile phone, a tablet, a laptop, a smart wearable device, a server, etc., and is not limited here. Among them, the user behavior prediction device described in the corresponding embodiment of FIG. 11 and the directed acyclic graph construction device described in the corresponding embodiment of FIG. 12 may be deployed on the execution device 1400 to implement the user behavior prediction device described in the corresponding embodiment of FIG. 4 The function of constructing an acyclic graph and the function of predicting user behavior in the corresponding embodiment of FIG. 8 . Specifically, the execution device 1400 includes: a receiver 1401, a transmitter 1402, a processor 1403 and a memory 1404 (the number of processors 1403 in the execution device 1400 can be one or more, one processor is taken as an example in Figure 14) , wherein the processor 1403 may include an application processor 14031 and a communication processor 14032. In some embodiments of the present application, the receiver 1401, the transmitter 1402, the processor 1403, and the memory 1404 may be connected by a bus or other means.

Memory 1404 may include read-only memory and random access memory and provides instructions and data to processor 1403 . A portion of memory 1404 may also include non-volatile random access memory (NVRAM). The memory 1404 stores processor and operating instructions, executable modules or data structures, or a subset thereof, or an extended set thereof, where the operating instructions may include various operating instructions for implementing various operations.

The processor 1403 controls the execution of operations of the device. In specific applications, various components of the execution device are coupled together through a bus system. In addition to the data bus, the bus system may also include a power bus, a control bus, a status signal bus, etc. However, for the sake of clarity, various buses are called bus systems in the figure.

The methods disclosed in the above embodiments of the present application can be applied to the processor 1403 or implemented by the processor 1403. The processor 1403 may be an integrated circuit chip with signal processing capabilities. During the implementation process, each step of the above method The steps may be completed by instructions in the form of hardware integrated logic circuits or software in the processor 1403 . The above-mentioned processor 1403 can be a general-purpose processor, a digital signal processor (DSP), a microprocessor or a microcontroller, and can further include an application specific integrated circuit (ASIC), a field programmable Gate array (field-programmable gate array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The processor 1403 can implement or execute each method, step and logical block diagram disclosed in the embodiment of this application. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the method disclosed in conjunction with the embodiments of the present application can be directly implemented by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other mature storage media in this field. The storage medium is located in the memory 1404. The processor 1403 reads the information in the memory 1404 and completes the steps of the above method in combination with its hardware.

The receiver 1401 may be configured to receive input numeric or character information and generate signal inputs related to performing relevant settings and functional controls of the device. The transmitter 1402 can be used to output numeric or character information through the first interface; the transmitter 1402 can also be used to send instructions to the disk group through the first interface to modify the data in the disk group; the transmitter 1402 can also include a display device such as a display screen .

In the embodiment of the present application, in one case, the processor 1403 is used to predict user behavior for the target page through the target model in the corresponding embodiment of FIG. 8 .

The embodiment of the present application also relates to a training device. Figure 15 is a schematic structural diagram of the training device provided by the embodiment of the present application. As shown in Figure 15, the training device 1500 is implemented by one or more servers. The training device 1500 can vary greatly due to different configurations or performance, and can include one or more central processing units (CPU) 1514 (eg, one or more processors) and memory 1532, one or more storage media 1530 (eg, one or more mass storage devices) storing applications 1542 or data 1544. Among them, the memory 1532 and the storage medium 1530 may be short-term storage or persistent storage. The program stored in the storage medium 1530 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the training device. Furthermore, the central processor 1514 may be configured to communicate with the storage medium 1530 and execute a series of instruction operations in the storage medium 1530 on the training device 1500 .

The training device 1500 may also include one or more power supplies 1526, one or more wired or wireless network interfaces 1550, one or more input and output interfaces 1558; or, one or more operating systems 1541, such as Windows ServerTM, Mac OS XTM , UnixTM, LinuxTM, FreeBSDTM and so on.

Specifically, the training device can execute the model training method in the corresponding embodiment of Figure 10.

Embodiments of the present application also relate to a computer storage medium. The computer-readable storage medium stores a program for performing signal processing. When the program is run on a computer, it causes the computer to perform the steps performed by the aforementioned execution device, or, The computer is caused to perform the steps performed by the aforementioned training device.

Embodiments of the present application also relate to a computer program product that stores instructions that, when executed by a computer, cause the computer to perform the steps performed by the foregoing execution device, or cause the computer to perform the steps performed by the foregoing training device. A step of.

The execution device, training device or terminal device provided by the embodiment of the present application may specifically be a chip. The chip includes: a processing unit and a communication unit. The processing unit may be, for example, a processor. The communication unit may be, for example, an input/output interface. Pins or circuits, etc. The processing unit can execute the computer execution instructions stored in the storage unit, so that the chip in the execution device executes the data processing method described in the above embodiment, or so that the chip in the training device executes the data processing method described in the above embodiment. Optionally, the storage unit is a storage unit within the chip, such as a register, cache, etc. The storage unit may also be a storage unit located outside the chip in the wireless access device, such as Read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (random access memory, RAM), etc.

Specifically, please refer to Figure 16. Figure 16 is a schematic structural diagram of a chip provided by an embodiment of the present application. The chip can be represented as a neural network processor NPU 1600. The NPU 1600 serves as a co-processor and is mounted to the host CPU (Host CPU). ), tasks are allocated by the Host CPU. The core part of the NPU is the arithmetic circuit 1603. The arithmetic circuit 1603 is controlled by the controller 1604 to extract the matrix data in the memory and perform multiplication operations.

In some implementations, the computing circuit 1603 includes multiple processing units (Process Engine, PE). In some implementations, arithmetic circuit 1603 is a two-dimensional systolic array. The arithmetic circuit 1603 may also be a one-dimensional systolic array or other electronic circuit capable of performing mathematical operations such as multiplication and addition. In some implementations, arithmetic circuit 1603 is a general-purpose matrix processor.

For example, assume there is an input matrix A, a weight matrix B, and an output matrix C. The arithmetic circuit obtains the corresponding data of matrix B from the weight memory 1602 and caches it on each PE in the arithmetic circuit. The operation circuit takes matrix A data and matrix B from the input memory 1601 to perform matrix operations, and the partial result or final result of the matrix is stored in an accumulator (accumulator) 1608 .

The unified memory 1606 is used to store input data and output data. The weight data directly passes through the storage unit access controller (Direct Memory Access Controller, DMAC) 1605, and the DMAC is transferred to the weight memory 1602. Input data is also transferred to unified memory 1606 via DMAC.

BIU is the Bus Interface Unit, that is, the bus interface unit 1613, which is used for the interaction between the AXI bus and the DMAC and the Instruction Fetch Buffer (IFB) 1609.

The bus interface unit 1613 (Bus Interface Unit, BIU for short) is used to fetch the memory 1609 to obtain instructions from the external memory, and is also used for the storage unit access controller 1605 to obtain the original data of the input matrix A or the weight matrix B from the external memory.

DMAC is mainly used to transfer the input data in the external memory DDR to the unified memory 1606 or the weight data to the weight memory 1602 or the input data to the input memory 1601 .

The vector calculation unit 1607 includes multiple arithmetic processing units, and if necessary, further processes the output of the arithmetic circuit 1603, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc. It is mainly used for non-convolutional/fully connected layer network calculations in neural networks, such as Batch Normalization, pixel-level summation, upsampling of predicted label planes, etc.

In some implementations, vector calculation unit 1607 can store the processed output vectors to unified memory 1606 . For example, the vector calculation unit 1607 can apply a linear function; or a nonlinear function to the output of the operation circuit 1603, such as linear interpolation on the prediction label plane extracted by the convolution layer, or a vector of accumulated values, to generate an activation value. . In some implementations, vector calculation unit 1607 generates normalized values, pixel-wise summed values, or both. In some implementations, the processed output vector can be used as an activation input to the arithmetic circuit 1603, such as for use in a subsequent layer in a neural network.

The instruction fetch buffer 1609 connected to the controller 1604 is used to store instructions used by the controller 1604;

The unified memory 1606, the input memory 1601, the weight memory 1602 and the fetch memory 1609 are all On-Chip memories. External memory is private to the NPU hardware architecture.

The processor mentioned in any of the above places can be a general central processing unit, a microprocessor, an ASIC, or one or more integrated circuits used to control the execution of the above programs.

In addition, it should be noted that the device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physically separate. The physical unit can be located in one place, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the drawings of the device embodiments provided in this application, the connection relationship between modules indicates that there are communication connections between them, which can be specifically implemented as one or more communication buses or signal lines.

Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by software plus necessary general hardware. Of course, it can also be implemented by dedicated hardware including dedicated integrated circuits, dedicated CPUs, dedicated memories, Special components, etc. to achieve. In general, all functions performed by computer programs can be easily implemented with corresponding hardware. Moreover, the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits or special-purpose circuits. circuit etc. However, for this application, software program implementation is a better implementation in most cases. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product in essence or that contributes to the existing technology. The computer software product is stored in a readable storage medium, such as a computer floppy disk. , U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk, etc., including several instructions to cause a computer device (which can be a personal computer, training device, or network device, etc.) to execute the steps described in various embodiments of this application. method.

In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product.

The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another, for example, the computer instructions may be transferred from a website, computer, training device, or data The center transmits to another website site, computer, training equipment or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store, or a data storage device such as a training device or a data center integrated with one or more available media. The available media may be magnetic media (eg, floppy disk, hard disk, magnetic tape), optical media (eg, DVD), or semiconductor media (eg, solid state disk (Solid State Disk, SSD)), etc.

Claims

A user behavior prediction method, characterized in that the method is implemented through a target model, and the method includes:

Obtain the first characteristic of the first item and the second characteristic of the second item, the first item and the second item are located in different lists or the same list of the target page, and the second item is located in the first Before the project;

Based on the first feature of the first item and the second feature of the second item, the second feature of the first item is obtained. The first feature is attribute information, and the second feature is fusion based on the attribute information. information obtained later;

Based on the second characteristic of the first item, a probability that the first item is clicked by the user is obtained.
The method of claim 1, further comprising:

Obtain the first characteristic of the third item, the first item and the third item are located in different lists or the same list of the target page, and the third item is adjacent to the first item;

Obtaining a third characteristic of the first item based on the first characteristic of the first item and the first characteristic of the third item;

Obtaining the probability that the first item is clicked by the user based on the second characteristic of the first item includes:

Based on the second characteristic of the first item and the third characteristic of the first item, a probability that the first item is clicked by the user is obtained.
The method according to claim 1, characterized in that, based on the first characteristic of the first item and the second characteristic of the second item, obtaining the second characteristic of the first item includes:

Perform mapping processing on the first feature of the first item to obtain a fourth feature of the first item;

Perform processing based on the self-attention mechanism on the second feature of the second item to obtain the fifth feature of the first item;

Perform a first fusion process on the fourth feature of the first item and the fifth feature of the first item to obtain the second feature of the first item.
The method according to claim 3, characterized in that the mapping process is performed on the first feature of the first item to obtain a fourth feature of the first item:

Perform mapping processing on the first feature of the first item, the user's request for the target page, and the probability that the second item is clicked by the user, to obtain the sixth feature of the first item, the first item The seventh characteristic and the eighth characteristic of the first item;

Perform a second fusion process on the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item to obtain the fourth feature of the first item.
The method of claim 2, wherein obtaining the third characteristic of the first item based on the first characteristic of the first item and the first characteristic of the third item includes:

Perform mapping processing on the first feature of the first item and the first feature of the third item to obtain the sixth feature of the first item and the ninth feature of the first item;

Perform a third fusion process on the sixth feature of the first item and the ninth feature of the first item to obtain the tenth feature of the first item;

Perform a fourth fusion process on the sixth feature of the first item and the tenth feature of the first item to obtain the third feature of the first item.
The method of claim 5, wherein the sixth characteristic of the first item and the first item The tenth feature of the object is subjected to the fourth fusion process to obtain the third feature of the first item including:

Perform mapping processing on the user's request for the target page to obtain the seventh feature of the first item;

Perform a fourth fusion process on the sixth feature of the first item, the seventh feature of the first item, and the tenth feature of the first item to obtain the third feature of the first item.
The method according to any one of claims 1 to 6, characterized in that if the first item is the first item in the target page, the second characteristic of the second item is a preset value.
The method according to any one of claims 1 to 7, characterized in that the target page contains multiple lists, multiple items located in the multiple lists constitute a directed acyclic graph, and the multiple items include all the first item, the second item and the third item.
A method for constructing a directed acyclic graph, characterized in that the method includes:

Obtain the eye movement data of users browsing the target page;

Based on the eye movement data, determining the user's browsing behavior for a plurality of items located in a plurality of lists on the target page;

Based on the browsing behavior, the multiple items are connected to obtain a directed acyclic graph.
The method according to claim 9, characterized in that, based on the browsing behavior, connecting the plurality of items to obtain a directed acyclic graph includes:

The items in the same list browsed by the user in the first order are connected in the first order, and the items in different lists browsed by the user in the second order are connected in the second order. , get the directed acyclic graph.
The method according to claim 9 or 10, characterized in that said obtaining the eye movement data of the user browsing the target page:

Use an eye tracker to collect eye movement data of users browsing the target page.
A model training method, characterized in that the method includes:

The first feature of the first item and the second feature of the second item are obtained through the model to be trained, the first item and the second item are located in different lists or the same list of the page to be processed, and the second item located before said first item;

The model to be trained obtains the second feature of the first item based on the first feature of the first item and the second feature of the second item. The first feature is attribute information, and the second feature is based on Information obtained after fusion of the attribute information;

Obtain the probability that the first item is clicked by the user based on the second feature of the first item through the model to be trained;

Based on the probability that the first item is clicked by the user and the real probability that the first item is clicked by the user, a target loss is obtained, the target loss is used to indicate the probability that the first item is clicked by the user and the first item is clicked by the user The difference between the true probabilities;

Based on the target loss, the parameters of the model to be trained are updated until the model training conditions are met, and the target model is obtained.
The method of claim 12, further comprising:

The first feature of the third item is obtained through the model to be trained, the first item and the third item are located in different lists or the same list of the page to be processed, and the third item is the same as the third item. The first item mentioned above is adjacent;

Obtain the third feature of the first item based on the first feature of the first item and the first feature of the third item by the to-be-trained model;

The method of obtaining the probability that the first item is clicked by the user based on the second feature of the first item through the model to be trained includes:

The probability that the first item is clicked by the user is obtained by using the to-be-trained model based on the second feature of the first item and the third feature of the first item.
A user behavior prediction device, characterized in that the device includes:

A first acquisition module configured to acquire the first feature of the first item and the second feature of the second item through the target model, where the first item and the second item are located in different lists or the same list on the target page, and The second item is located before the first item;

The second acquisition module is configured to acquire the second characteristics of the first item based on the first characteristics of the first item and the second characteristics of the second item through the target model. The first characteristics are attribute information. The second feature is information obtained after fusion based on the attribute information;

The third acquisition module is configured to acquire the probability that the first item is clicked by the user based on the second feature of the first item through the target model.
The device according to claim 14, characterized in that the device further includes:

The fourth acquisition module is used to acquire the first characteristics of the third item through the target model. The first item and the third item are located in different lists or the same list of the target page, and the third item an item adjacent to said first item;

A fifth acquisition module, configured to acquire the third feature of the first item based on the first feature of the first item and the first feature of the third item through the target model;

The third acquisition module is configured to acquire the probability that the first item is clicked by the user based on the second characteristic of the first item and the third characteristic of the first item through a target model.
The device according to claim 14, characterized in that the second acquisition module is used for:

Perform mapping processing on the first feature of the first item through the target model to obtain the fourth feature of the first item;

The second feature of the second item is processed based on the self-attention mechanism through the target model to obtain the fifth feature of the first item;

The fourth feature of the first item and the fifth feature of the first item are subjected to a first fusion process through the target model to obtain the second feature of the first item.
The device according to claim 16, characterized in that the second acquisition module is used for:

The first feature of the first item, the user's request for the target page, and the probability that the second item is clicked by the user are mapped through the target model to obtain the sixth feature of the first item, the The seventh feature of the first item and the eighth feature of said first item;

The fourth feature of the first item is obtained by performing a second fusion process on the sixth feature of the first item, the seventh feature of the first item, and the eighth feature of the first item through the target model.
The device according to claim 17, characterized in that the fifth acquisition module is used for:

The first feature of the first item and the first feature of the third item are mapped through the target model to obtain the sixth feature of the first item and the ninth feature of the first item;

Perform a third fusion process on the sixth feature of the first item and the ninth feature of the first item through the target model, Obtain the tenth characteristic of the first item;

The target model performs a fourth fusion process on the sixth feature of the first item and the tenth feature of the first item to obtain the third feature of the first item.
The device according to claim 18, characterized in that the fifth acquisition module is used for:

Perform mapping processing on the user's request for the target page through a target model to obtain the seventh feature of the first item;

The third feature of the first item is obtained by performing a fourth fusion process on the sixth feature of the first item, the seventh feature of the first item, and the tenth feature of the first item through the target model.
The device according to any one of claims 14 to 19, characterized in that if the first item is the first item in the target page, the second characteristic of the second item is a preset value.
The device according to any one of claims 14 to 20, wherein the target page includes multiple lists, multiple items located in the multiple lists constitute a directed acyclic graph, and the multiple items include all the first item, the second item and the third item.
A directed acyclic graph construction device, characterized in that the device includes:

The acquisition module is used to obtain eye movement data of users browsing the target page;

A determining module, configured to determine the user's browsing behavior for multiple items based on the eye movement data, where the multiple items are located in multiple lists of the target page;

The connection module is used to connect the multiple items based on the browsing behavior to obtain a directed acyclic graph.
The device according to claim 22, characterized in that the connection module is used to connect items in the same list browsed by the user in the first order in the first order, and connect the user Items in different lists browsed in the second order are connected according to the second order to obtain a directed acyclic graph.
The device according to claim 22 or 23, characterized in that the acquisition module is used to collect the eye movement data of the user browsing the target page through an eye tracker.
A model training device, characterized in that the device includes:

A first acquisition module, configured to acquire the first feature of the first item and the second feature of the second item through the model to be trained, where the first item and the second item are located in different lists or the same list of the page to be processed. , and the second item is located before the first item;

The second acquisition module is used to acquire the second characteristic of the first item based on the first characteristic of the first item and the second characteristic of the second item through the to-be-trained model, where the first characteristic is an attribute. Information, the second feature is information obtained after fusion based on the attribute information;

A third acquisition module, configured to acquire the probability that the first item is clicked by the user based on the second feature of the first item through the model to be trained;

The fourth acquisition module is used to obtain a target loss based on the probability that the first item is clicked by the user and the real probability that the first item is clicked by the user. The target loss is used to indicate the probability that the first item is clicked by the user. and the difference between the real probability of the first item being clicked by the user;

An update module, configured to update the parameters of the model to be trained based on the target loss until the model training conditions are met to obtain the target model.
The device according to claim 25, characterized in that the device includes:

A fifth acquisition module, configured to acquire the first feature of the third item through the model to be trained, where the first item and the third item are located in different lists or the same list of the page to be processed, and The third item is adjacent to the first item;

A sixth acquisition module, configured to acquire the third feature of the first item based on the first feature of the first item and the first feature of the third item through the to-be-trained model;

The third acquisition module is configured to acquire the probability that the first item is clicked by the user based on the second characteristic of the first item and the third characteristic of the first item through the model to be trained.
A user behavior prediction device, characterized in that the device includes a memory and a processor;

The memory stores code, the processor is configured to execute the code, and when the code is executed, the device performs the method of any one of claims 1 to 13.
A computer storage medium, characterized in that the computer storage medium stores a computer program, which when executed by a computer causes the computer to implement the method described in any one of claims 1 to 13.
A computer program product, characterized in that the computer program product stores instructions, which when executed by a computer, cause the computer to implement the method described in any one of claims 1 to 13.