Behavior prediction method and device, electronic equipment and storage medium

Abstract

The invention relates to a behavior prediction method and device, electronic equipment and a storage medium. The method comprises the steps that historical interaction information of a user in a target application in multiple historical time periods is acquired; the historical interaction information comprises intention, time information and interaction behavior information of the user for opening the target application in each historical time period; inputting the historical interaction information of each historical time period into an encoding and decoding module in a behavior prediction model for processing to obtain a short-term intention and a long-term intention of the user; the behavior prediction model is a model constructed on the basis of a neuro-hox process; and inputting the short-term intention and the long-term intention of the user into an intention prediction module in the behavior prediction model for processing to obtain an intention of the user for opening the target application at a prediction time point. The method can accurately identify the specific reason for the user to open the application for the first time.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a behavior prediction method, device, electronic device, storage medium and program product. Background Art

[0002] In order to cope with information overload, integrating search and recommendation services into a single application is an important feature configuration operation, enabling users to access information in both active and passive ways. For example, in a short video application, users can perform information search and also explore the content recommended by the application to interact, which leads to users opening such applications with different intentions. Different intentions for opening applications correspond to different needs and behavior patterns. Predicting the intention of users to open applications can help improve user engagement and enhance the performance of various downstream tasks.

[0003] In the related art, predictions for user interactions are generally user retention prediction and user intent prediction. User retention prediction is used to predict the probability of a user using an application again; user intent prediction is used to predict the user's possible intentions when using an application, such as making payments, buying movie tickets, ordering coffee, buying snacks, etc. However, both methods are predictions after the user uses the application, and cannot identify the user's intention to open the application. Summary of the invention

[0004] The present disclosure provides an intent recognition method, device, electronic device, storage medium and program product to at least solve the problem that the related art can only widely predict multiple intents within an application but cannot accurately identify the specific reason why a user opens an application for the first time. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, there is provided a method for identifying intent, including:

[0006] Acquire historical interaction information of the user in the target application in multiple historical time periods; the historical interaction information includes the user's intention to open the target application in each of the historical time periods, time information, and interaction behavior information;

[0007] Input the historical interaction information of each historical time period into the encoding and decoding module in the intention recognition model for processing to obtain the short-term intention and long-term intention of the user; the short-term intention represents the user's intention to open the target application in the time period with the latest start time among the multiple historical time periods, and the long-term intention represents the user's comprehensive intention to open the target application in the multiple historical time periods;

[0008] The short-term intention and the long-term intention of the user are input into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point.

[0009] In an exemplary embodiment, the encoding and decoding module includes an encoding submodule and a decoding submodule; the inputting of the historical interaction information of each historical time period into the encoding and decoding module in the intention recognition model for processing to obtain the short-term intention and long-term intention of the user includes:

[0010] Inputting the historical interaction information of each historical time period into the encoding submodule for encoding processing respectively to obtain encoding information of each historical time period;

[0011] Input the encoding information of each historical time period, the starting time point of each historical time period, and the intention of opening the target application in each historical time period into the decoding submodule for decoding processing to obtain decoding information of each historical time period;

[0012] The short-term intention of the user is obtained based on the encoded information of the latest time period having the latest start time among the multiple historical time periods, and the long-term intention of the user is obtained based on the decoded information of the latest time period.

[0013] In an exemplary embodiment, the inputting the short-term intention and the long-term intention of the user into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point includes:

[0014] Obtaining user identification information of the user;

[0015] The short-term intention, the long-term intention and the user identification information of the user are input into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point.

[0016] In an exemplary embodiment, the inputting the short-term intention, the long-term intention, and the user identification information of the user into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point includes:

[0017] In the intention prediction module in the intention recognition model, the short-term intention and the long-term intention are fused to obtain fused representation information of the user's intention;

[0018] Based on the fused representation information and the user identification information, the user's intention to open the target application at the predicted time point is determined.

[0019] In an exemplary embodiment, the fusing the short-term intention and the long-term intention to obtain the fused representation information of the user intention includes:

[0020] Obtaining a time difference between a latest time point corresponding to a latest time period among the multiple time periods and the predicted time point;

[0021] By using the time difference, mapping and converting the short-term intention and the long-term intention to obtain a weight coefficient; the weight coefficient is used to characterize the influence of the short-term intention on the fusion representation information of the user intention;

[0022] The short-term intention and the long-term intention are fused by using the weight coefficient to obtain fused representation information of the user intention.

[0023] In an exemplary embodiment, the intention prediction module includes a multilayer perceptron; the determining the user's intention to open the target application at the predicted time point based on the fused representation information and the user identification information includes:

[0024] Inputting the fusion representation information and the user identification information into the multi-layer perceptron for processing to obtain perception information;

[0025] Obtain an intention relevance score corresponding to the latest time period among the multiple time periods, and determine the user's intention to open the target application at the predicted time point based on the intention relevance score and the perception information; the intention relevance score represents the ratio between the number of occurrences of different intentions in the same time period.

[0026] In an exemplary embodiment, the intent recognition model is trained in the following manner:

[0027] Acquire sample interaction information of sample users in a target application in multiple historical time periods; the historical interaction information includes the user's real intention, time information, and interaction behavior information for opening the target application in each of the historical time periods;

[0028] The sample interaction information of each historical time period is input into the encoding and decoding module in the intention recognition model to be trained for processing, so as to obtain the sample short-term intention of the sample user and the sample long-term intention corresponding to each historical time period.

[0029] Input the sample short-term intention of the sample user and the sample long-term intention corresponding to the latest time period among the multiple historical time periods into the intention prediction module in the intention recognition model to be trained for processing, obtain the predicted intention of the sample user to open the target application at the predicted time point, and obtain a first loss value based on the predicted intention and the true intention; and obtain a second loss value of the sample user's intention to open the target application based on the sample long-term intention corresponding to each of the historical time periods;

[0030] The intention recognition model to be trained is trained according to the loss value and the second loss value to obtain the intention recognition model.

[0031] In an exemplary embodiment, obtaining the second loss value of the sample user's intention to open the target application based on the sample long-term intention corresponding to each of the historical time periods includes:

[0032] Obtaining the intention relevance score of the sample user corresponding to each of the historical time periods;

[0033] Based on the intention relevance score corresponding to each of the historical time periods and the sample long-term intention, a second loss value of the sample user's intention to open the target application is obtained.

[0034] According to a second aspect of an embodiment of the present disclosure, there is provided an intention recognition device, including:

[0035] An information acquisition unit is configured to acquire historical interaction information of a user in a target application in multiple historical time periods; the historical interaction information includes the user's intention to open the target application in each of the historical time periods, time information, and interaction behavior information;

[0036] an information processing unit configured to execute inputting the historical interaction information of each historical time period into a coding and decoding module in an intention recognition model for processing, so as to obtain a short-term intention and a long-term intention of the user; the short-term intention represents the intention of the user to open the target application in the time period with the latest start time among the multiple historical time periods, and the long-term intention represents the comprehensive intention of the user to open the target application in the multiple historical time periods;

[0037] The intention prediction unit is configured to execute the input of the short-term intention and the long-term intention of the user into the intention prediction module in the intention recognition model for processing, so as to obtain the intention of the user to open the target application at the predicted time point.

[0038] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0039] processor;

[0040] a memory for storing instructions executable by the processor;

[0041] The processor is configured to execute the instructions to implement the method as described in any one of the above items.

[0042] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute any of the methods described above.

[0043] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, wherein the computer program product includes instructions, and when the instructions are executed by a processor of an electronic device, the electronic device is enabled to execute any of the methods described above.

[0044] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0045] Obtain the historical interaction information of the user in the target application in multiple historical time periods; input the historical interaction information of each time period into the encoding and decoding module in the intent recognition model for processing to obtain the user's short-term intention and long-term intention; input the user's short-term intention and long-term intention into the intention prediction module in the intent recognition model for processing to obtain the user's intention to open the target application at the predicted time point. Unlike traditional user behavior intention prediction, this method adopts a new direction of behavior prediction for opening applications, focusing more on the time characteristics of user behavior rather than just predicting multiple intentions within the application. Traditional methods often ignore the periodic characteristics of user behavior, for example, users may drink coffee at a specific time (such as in the afternoon) or buy movie tickets on weekends. The goal of the present disclosure is to accurately identify the specific reasons why users open the application for the first time, such as searching or recommending, rather than broadly predicting various intentions. This refined insight is crucial to the effective integration of search and recommendation applications, and provides the possibility of achieving a more personalized user experience.

[0046] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0048] Fig. 1 is a schematic flow chart of a behavior prediction method according to an exemplary embodiment.

[0049] Fig. 2 is a schematic diagram of the structure of an intention recognition model according to an exemplary embodiment.

[0050] Fig. 3 is a schematic flow chart of an intention prediction step according to an exemplary embodiment.

[0051] Fig. 4 is a schematic diagram of a training process of an intent recognition model according to an exemplary embodiment.

[0052] Fig. 5 is a structural block diagram of a behavior prediction device according to an exemplary embodiment.

[0053] Fig. 6 is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0054] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0055] It should be noted that the embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0056] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure 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 data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in the present disclosure are all information and data authorized by the user or fully authorized by all parties.

[0057] In an exemplary embodiment, as shown in FIG1 , a behavior prediction method is provided. This embodiment is illustrated by applying the method to a terminal. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. Among them, the terminal can be but is not limited to various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented as an independent server or a server cluster consisting of multiple servers. In this embodiment, the method includes the following steps:

[0058] In step S110, historical interaction information of the user in the target application in multiple historical time periods is obtained; the historical interaction information includes the user's intention to open the target application in each historical time period, time information, and interaction behavior information.

[0059] Among them, the target application refers to a computer program that integrates both search and recommendation services, such as short video applications, e-commerce applications, etc.

[0060] Each historical time period may represent a time period from when a user opens a target application to when the user closes the target application, and the historical interaction information corresponding to each historical time period may include all interactions of the user from when the user opens the target application to when the user closes the target application.

[0061] Specifically, the historical interaction information of each historical time period may include the user's intention when opening the target application (such as searching for information and browsing recommended content), time information (including the time points of each interactive behavior performed in the time period), and interactive behavior information (such as item interaction or query behavior). Among them, the intention when opening the application can be determined based on the user's behavior within a preset time after opening the target application. For example, if the user searches within 30 seconds of opening the target application, the intention is determined to be "searching for information", otherwise, the intention is determined to be "browsing recommended content".

[0062] For example, referring to the schematic diagram of the structure of the intent recognition model shown in FIG2, , , …, Can indicate time periods separately ~ , Time period ~ , …, time period ~ Corresponding historical interaction information. Taking the first time period as an example, Can indicate time points , and The corresponding interactive operations, where the dashed box and Indicates at a point in time and The interactive operation is to search for information. Indicates at a point in time The interactive operation of is browsing the recommended information. The historical interactive information of other time periods is similar to that of the first time period and will not be repeated here.

[0063] In step S120, the historical interaction information of each historical time period is input into the encoding and decoding module in the intention recognition model for processing to obtain the user's short-term intention and long-term intention.

[0064] Among them, the short-term intention represents the user's intention to open the target application in the time period with the latest start time among multiple historical time periods, and the long-term intention represents the user's comprehensive intention to open the target application in multiple historical time periods.

[0065] Among them, the intent recognition model is the Neural Hawkes Process-based Open-App Motivation prediction model (NHP-OAM), which is a novel neural Hawkes process model that can capture the temporal dependency between users' historical browsing and query behaviors.

[0066] The Hawkes process is a mathematical model used to model the interdependence between events in an event sequence. It is a self-exciting point process, the main feature of which is that the occurrence of an event will increase the probability of other events occurring in a short period of time in the future. This model is very suitable for scenarios where events have a triggering relationship with each other, such as transactions in financial markets, information dissemination in social networks, and aftershock sequences of earthquakes. In the Hawkes process, the occurrence of each event may cause the probability density function of subsequent events to increase, forming a "self-exciting" pattern. In this way, the Hawkes process can very effectively capture the time dynamics and structural dependencies in event data, making it of great application value in predicting the time of future events and analyzing the impact mechanism between events.

[0067] The user's intention to open an app is affected by user-specific factors, search queries, clicked items, and the time of occurrence, and these activities occur sequentially, showing a complex time dependency. For example, a user may open an app multiple times to search for the same content, that is, there is repeated query behavior. If a certain content was searched for the previous time, there is a high probability that the same content will be searched again the next time the app is opened. Based on these situations, it can be seen that the user's intention to open an app shows periodic and repeated query characteristics, and the time interval between users opening apps is not fixed. Therefore, in view of the characteristics of this problem, the Neural Hawkes Process can effectively capture the obvious periodic and repeated query characteristics in the past intention to open apps without assuming a consistent sampling time interval like time series modeling.

[0068] Among them, short-term intent is used to represent the user's immediate preference, which can be obtained based on the user's historical interaction information in the historical time period closest to the current time.

[0069] Among them, long-term intent is used to represent the user's interests and preferences over a longer period of time, and can be obtained based on all historical interaction information of the user in each historical time period.

[0070] Specifically, the structure of the hierarchical transformation module is a hierarchical architecture, for example, a two-layer architecture, one layer is a coding submodule, and the other layer is a decoding submodule. When applied, the historical interaction information of each historical time period is first input into the coding submodule of the first layer for coding processing to obtain the coding information of each historical time period; then the coding information of each historical time period is input into the decoding submodule of the second layer for decoding processing to obtain the decoding information of each historical time period.

[0071] Among them, the encoding submodule is used to encode the historical interaction information of each historical time period respectively, that is, the encoding process corresponding to each historical time period is independent of each other, so the encoding information of the latest time period among multiple historical time periods can be used to represent the user's short-term intention. The decoding submodule is used to generate a long-term embedded representation of each historical time period. The decoding process corresponding to each historical time period combines the information of other previous historical time periods. Therefore, the decoding information of the latest time period among each historical time period can represent the user's long-term intention.

[0072] In this step, the encoding and decoding module can not only learn the sequence of historical intentions to open applications, but also learn the session-level and historical-level contextual representations of different intention behaviors.

[0073] In step S130, the user's short-term intention and long-term intention are input into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point.

[0074] Among them, the intentions for opening the target application include "search" and "recommendation", namely "searching for information" and "browsing recommended content".

[0075] Specifically, as shown in Figure 2, the intent prediction module (i.e., the open application intent prediction layer in the figure) can be added on top of the neural Hawkes process part, and the user's historical information (including short-term intentions and long-term intentions) is integrated through the intent prediction module to obtain the user's intention to open the target application at the predicted time point.

[0076] Furthermore, in one embodiment, the user's historical information may also include the user's personal information (recorded as user identification information). When identifying the user's intention to open the target application at the predicted time point, in addition to the user's short-term intention and long-term intention, the user identification information may also be introduced. The user's short-term intention, long-term intention and user identification information are jointly input into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point.

[0077] This step adds an additional intent prediction module on top of the neural Hawkes process model to better predict the user's intention to open the target app.

[0078] In the above behavior prediction method, the historical interaction information of the user in the target application in multiple historical time periods is obtained; the historical interaction information of each time period is input into the encoding and decoding module in the intent recognition model for processing to obtain the user's short-term intention and long-term intention; the user's short-term intention and long-term intention are input into the intention prediction module in the intent recognition model for processing to obtain the user's intention to open the target application at the predicted time point. Unlike traditional user intention prediction, this method adopts a new direction of behavior prediction for opening applications, focusing more on the time characteristics of user behavior rather than just predicting multiple intentions within the application. Traditional methods often ignore the periodic characteristics of user behavior, for example, users may drink coffee at a specific time (such as in the afternoon) or buy movie tickets on weekends. The goal of the present disclosure is to accurately identify the specific reasons why users open the application for the first time, such as searching or recommending, rather than broadly predicting various intentions. This refined insight is crucial for the effective integration of search and recommendation applications, and provides the possibility of achieving a more personalized user experience.

[0079] The encoding and decoding module in the intention recognition model includes an encoding submodule and a decoding submodule; correspondingly, in an exemplary embodiment, in step S120, the historical interaction information of each historical time period is input into the encoding and decoding module in the intention recognition model for processing to obtain the user's short-term intention and long-term intention, which can be specifically achieved by the following steps:

[0080] In step S121, the historical interaction information of each historical time period is input into the encoding submodule for encoding processing to obtain the encoding information of each historical time period.

[0081] In step S122, the coding information of each historical time period, the starting time point of each historical time period, and the intention to open the target application in each historical time period are input into the decoding submodule for decoding processing to obtain the decoding information of each historical time period.

[0082] In step S123, the short-term intention of the user is obtained based on the encoded information of the latest time period among the multiple historical time periods, and the long-term intention of the user is obtained based on the decoded information of the latest time period.

[0083] In a specific implementation, if the historical interaction information of each historical time period is regarded as a historical session, the encoding submodule of the first layer represents a session-level encoder, which is used to generate embedded information of the user's intention in each time period. Specifically, the encoding submodule of the first layer may include multiple encoders, and the encoder may be a Transformer encoder. Each encoder corresponds to a time period and is used to encode the historical interaction information of the time period. The encoded information obtained by encoding the historical interaction information of the last time period can represent the user's short-term intention.

[0084] The decoding submodule of the second layer can be regarded as a historical encoder, which is used to generate the long-term embedding of the intention prediction of opening the application and the historical representation of the intensity function. The Transformer decoder can be used. Specifically, the input information of the decoding submodule of the second layer includes the output of the encoding submodule of the first layer (i.e., the encoding information of each historical time period), the starting time point of each historical time period (i.e., the time point of opening the target application), and the intention of opening the target application in each historical time period. The encoding information, starting time point, and intention of opening the target application corresponding to each historical time period can constitute a time series. The time series input information is input into the decoding submodule of the second layer for processing. The generated time-aware hidden representation of each time period (i.e., the output decoded information) contains the historical interaction information of other time periods before the time period. Therefore, the decoded information of the last time period contains all the historical interaction information from the starting time point of the first time period to the ending time point of the last time period, so it can be used to represent the user's long-term intention.

[0085] Furthermore, since the time point, intent type, interaction item, query vocabulary and other information in the historical interaction information belong to different types, each type has its corresponding embedded representation. Therefore, before the historical interaction information of each time period is input into the encoding submodule of the first layer for encoding processing, it is also necessary to pre-process the various information in the historical interaction information to convert it into a vector of fixed dimension for subsequent processing. Specifically, an embedding layer can be added to the encoding and decoding module, and the embedding layer is set before the encoding submodule of the first layer. After the historical interaction information of each time period is input into the embedding layer for dimensional conversion processing, the processed historical interaction information of each time period output by the embedding layer is used as an input variable and input into the corresponding encoder for encoding processing.

[0086] For example, referring to the structural diagram of the intent recognition model shown in FIG2, the encoding submodule of the first layer may correspond to each Transformer encoder in the figure, and the decoding submodule of the second layer may correspond to the Transformer decoder in the figure. The process of obtaining the user's short-term and long-term intentions includes: collecting the user's historical interaction information in each historical time period; , ,… Input the embedding layer of the encoding and decoding module to obtain the embedded representation of the historical interaction information of each historical time period; the embedded representation of each historical time period output by the embedding layer corresponds to each Transformer encoder input in the first layer, and output the encoded information of each historical time period , ,… ,in Indicates the user's short-term intention; encodes information from each historical time period , ,… , Start time , ,… and an intent to open said target app , ,… , and are jointly input into the Transformer decoder of the second layer to obtain the decoding information of each historical time period , ,… ,in Indicates the user's long-term intent.

[0087] It can be understood that the user's intention to open an application is affected by searches, queries, and clicks in the historical logs. In order to capture the rich contextual information in the user session, the present disclosure uses a hierarchical encoding and decoding module as shown in Figure 4 to replace the historical level encoder of the traditional Hawkes process. First, the user's interactive behaviors such as queries and clicks in the session are encoded into session-level representations using the encoding submodule. Subsequently, the session-level information of the historical session is combined with the intention and time information of opening the application and input into the decoding submodule to ensure that future data will not be leaked, so that it can be used to calculate the historical level representation of the intensity function.

[0088] In this embodiment, the encoding and decoding module can be used as a historical encoder in the Hawkes process, which can not only learn the sequence of historical intentions to open applications, but also learn the session-level and historical-level context representations of different intention behaviors. The encoding and decoding module encodes multiple factors that affect the user's intention to open the application, and obtains the user's short-term intention and long-term intention, so as to improve the accuracy of subsequent prediction of the user's intention to open the target application based on the short-term intention and long-term intention.

[0089] In an exemplary embodiment, in step S130, the user's short-term intention and long-term intention are input into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point, which can be specifically achieved by the following steps:

[0090] Step S131, obtaining user identification information of the user;

[0091] Step S132, inputting the user's short-term intention, long-term intention and user identification information into the intention prediction module in the intention recognition model for processing, and obtaining the user's intention to open the target application at the predicted time point.

[0092] Among them, the intention prediction module is used to predict the user's intention to open the target application next time.

[0093] The user identification information refers to the user's personal information, such as a user identification (ID).

[0094] Specifically, when predicting the user's intention to open the target application, in addition to the user's historical short-term intentions and long-term intentions, this embodiment also introduces the user's personal information, and inputs the three types of information together as input information of the intention prediction module for processing to improve the accuracy of the predicted user's intention to open the target application.

[0095] In the specific implementation, the user's historical short-term intentions and long-term intentions can be fused first, and the fused information can be regarded as a comprehensive embedding of historical information. When further fusing the user identification information, the user identification information can be converted into a vector, and this vector is processed together with the comprehensive embedding of historical information to obtain the user's intention to open the target application at the predicted time point.

[0096] In this embodiment, when predicting the user's intention to open the target application, in addition to the user's historical short-term and long-term intentions, the user's personal information is also introduced. By considering multi-dimensional influencing factors, the credibility and accuracy of the prediction results of the user's next intention to open the target application can be improved.

[0097] In an exemplary embodiment, as shown in FIG. 3 , step S132 inputs the user's short-term intention, long-term intention, and user identification information into the intention prediction module in the intention recognition model for processing, and obtains the user's intention to open the target application at the predicted time point, specifically including:

[0098] Step S310, in the intention prediction module in the intention recognition model, the short-term intention and the long-term intention are fused to obtain fused representation information of the user's intention;

[0099] Step S320 , determining the user's intention to open the target application at the predicted time point based on the fused representation information and the user identification information.

[0100] In a specific implementation, in the intent prediction module in the intent recognition model, the fusion processing of short-term intent and long-term intent can be achieved through the time gate method. Specifically, the short-term intent and long-term intent of the user can be fused through the time gate method, and the obtained result is used as the fusion representation information of the user's intent. When applying the time gate method, the relative time is obtained by using the difference between the predicted time point and the time point of the user's last interaction. Combined with the relative time, the user's short-term intent and long-term intent are mapped and transformed to obtain the weight coefficient. The user's short-term intent and long-term intent are fused through the weight coefficient to obtain the fusion representation information of the user's intent.

[0101] Furthermore, the user's intention to open the target application at the predicted time point can be determined based on the fused representation information of the user's intention and the user identification information. This process can be implemented through a multi-layer perceptron, that is, the fused representation information of the user's intention and the user identification information are input into the multi-layer perceptron, and the user's intention to open the target application at the predicted time point is obtained based on the output result of the multi-layer perceptron.

[0102] In this embodiment, the user's short-term intention and long-term intention are combined, and the comprehensive representation information of the user's intention is obtained through fusion processing. The user's identification information is also considered, and then the user's intention to open the target application at the predicted time point is determined based on this information. This method comprehensively considers the user's immediate preferences and long-term goals, and combines the user's identity information, which helps to understand and infer the user's behavioral intention more comprehensively and accurately.

[0103] In an exemplary embodiment, in step S310, the short-term intention and the long-term intention are fused to obtain fused representation information of the user's intention, including:

[0104] Step S311, obtaining the time difference between the latest time point corresponding to the latest time period among the multiple time periods and the predicted time point;

[0105] Step S312, mapping and converting the short-term intention and the long-term intention by using the activation function and the time difference to obtain a weight coefficient; the weight coefficient is used to characterize the influence of the short-term intention on the fusion representation information of the user intention;

[0106] Step S313: The short-term intention and the long-term intention are fused by using the weight coefficient to obtain fused representation information of the user's intention.

[0107] It can be understood that, considering the significant impact of time factors and different user habits on the user's intention to open the application, the time gate method is used to fuse short-term intentions and long-term intentions. When using the time gate method, time information must also be introduced. Time information refers to the time difference between the predicted time point and the latest time point in the user's historical interaction information. The predicted time point can be regarded as the current time, and the time information can be understood as the time difference between the current time and the time when the user last opened the target application to interact. By introducing this time difference, time-related parameters can be adjusted, so that the intent recognition model can better understand the impact of time changes on user behavior, thereby improving the credibility of the prediction results.

[0108] In the specific implementation, as shown in Figure 2, Corresponding to the user's short-term intention, Corresponding to the user's long-term intention, when the short-term intention and long-term intention are integrated through the time gate, the current time is introduced at the same time. (corresponding to the predicted time point) the latest time point in the user's historical interaction information The time difference between The time difference is encoded using an embedding layer to obtain a time code Then, through the activation function (such as Sigmoid function) and the time encoding, the short-term intention and long-term intentions The mapping conversion process is performed to obtain the weight coefficient g. The mapping conversion process can be expressed by the formula:

[0109]

[0110] in, and is a learnable weight matrix. The weight coefficients are used to integrate the influence of long-term intention, short-term intention, and time encoding, helping the model decide which information should be considered more in the final prediction.

[0111] Furthermore, the weight coefficient is used as the control weight to fuse the short-term intention and the long-term intention to obtain the fusion representation information of the user's intention. , the fusion process can be expressed as:

[0112]

[0113] in, Represents element-wise product.

[0114] In this embodiment, by calculating the time difference between the latest time point corresponding to the latest time period and the predicted time point, the time difference is introduced and integrated with the hierarchical user history representation embedding (i.e., the user's short-term intention and long-term intention), the fusion representation information of the user's intention is obtained, and the time correlation of the user's behavior is considered, which can better understand the temporal characteristics of the user's behavior and improve the accuracy of subsequent intention prediction. Further, through the mapping conversion processing of the time difference, a weight coefficient is generated, and the trade-off of the time factor can be introduced between the short-term intention and the long-term intention, so as to better capture the dynamic behavior characteristics of the user.

[0115] In an exemplary embodiment, the intention prediction module includes a multi-layer perceptron; correspondingly, determining the user's intention to open the target application at the predicted time point based on the fusion representation information and the user identification information in step S320 includes:

[0116] Step S321, inputting the fusion representation information and the user identification information into a multi-layer perceptron for processing to obtain perception information;

[0117] Step S322, obtaining the intention relevance score corresponding to the latest time period among the multiple time periods, and determining the user's intention to open the target application at the predicted time point based on the intention relevance score and the perception information.

[0118] The intent relevance score represents the ratio of the number of occurrences of different intents in the same time period. ~ During the time period, the number of information search operations was and , the operations for browsing recommended information are , therefore, the intention correlation score corresponding to this time period can be 2:1.

[0119] In the specific implementation, as shown in Figure 2, The fused representation information corresponding to the user’s intention, Corresponding to the user identification information, based on the fusion representation information and the user identification information, when determining the user's intention to open the target application at the predicted time point, the fusion representation information of the user's intention is first and user identification information The multilayer perceptron (MLP) is input together to model feature interaction to obtain perception information, wherein the multilayer perceptron can be specifically a double-layer MLP. The perception information is further processed through an activation function (Sigmoid function) to obtain the user's intention to open the target application at the predicted time point.

[0120] In addition, considering the monotonically increasing characteristics of the Sigmoid function, in addition to the perception information, the intention relevance score corresponding to the latest time period in the user's historical interaction information is also introduced ,Based on the intention relevance score and perception information, the user’s intention to open the target application at the predicted time point is determined.,The process can be expressed by the formula:

[0121]

[0122] in, Indicates a connection operation. The prediction score representing the user's intent to open the target app at the predicted time.

[0123] In this embodiment, based on the fusion representation information of the user's intention and the user identification information, the relevance score of the user's intention is further combined to predict the user's intention to open the target application, which can improve the accuracy of the prediction result.

[0124] In an exemplary embodiment, as shown in FIG4 , the training method of the intent recognition model includes the following steps:

[0125] Step S410, obtaining sample interaction information of sample users in the target application in multiple historical time periods; the historical interaction information includes the user's real intention to open the target application in each historical time period, time information, and interaction behavior information;

[0126] Step S420, inputting the sample interaction information of each historical time period into the encoding and decoding module in the intention recognition model to be trained for processing, and obtaining the sample short-term intention of the sample user and the sample long-term intention corresponding to each historical time period;

[0127] Step S430, inputting the sample short-term intention of the sample user and the sample long-term intention corresponding to the latest time period in multiple historical time periods into the intention prediction module in the intention recognition model to be trained for processing, obtaining the predicted intention of the sample user to open the target application at the predicted time point, and obtaining a first loss value based on the predicted intention and the real intention;

[0128] Step S440, obtaining a second loss value of the sample user's intention to open the target application based on the sample long-term intention corresponding to each historical time period;

[0129] Step S450: Train the intent recognition model to be trained according to the first loss value and the second loss value to obtain the intent recognition model.

[0130] In a specific implementation, before training the intent recognition model to be trained, a data set can be obtained first, and the sample interaction information of the sample users in the data set in the target application in multiple historical time periods is used as the input variable. The input variable is first processed by the encoding and decoding module in the intent recognition model to obtain the short-term intention of the sample user and the long-term intention corresponding to each historical time period. Then, on the one hand, the sample short-term intention of the sample user and the sample long-term intention corresponding to the latest time period in multiple historical time periods are processed by the intention prediction module to predict the predicted intention of the sample user to open the target application at the predicted time point, so that the first loss value can be calculated based on the difference between the predicted intention and the true intention of the sample user, wherein the loss function can adopt the binary cross entropy loss function. On the other hand, the Hawkes process module can be used to perform Hawkes process conversion on the sample long-term intention corresponding to each historical time period to obtain the log-likelihood value of the sample user's intention to open the target application as the second loss value. As shown in FIG. 2, it can be understood that in the Hawkes process, the maximum likelihood is used to estimate the parameters of the model, thereby fitting the data and inferring the influence relationship between events. Therefore, the log-likelihood value can be calculated by the maximum likelihood estimation (MLE). Finally, the first loss value and the second loss value are added to obtain the total loss, and the intent recognition model to be trained is trained based on the total loss to obtain the intent recognition model.

[0131] The intent recognition model trained in this embodiment is different from the traditional user intent prediction model. It adopts a new direction of open application intent prediction, focusing more on the time characteristics of user behavior rather than just predicting multiple intents within the application. Traditional methods often ignore the cyclical characteristics of user behavior, while this method accurately identifies the specific reasons why users open the application for the first time, such as searching or recommending, rather than broadly predicting various intents. This refined insight is crucial for effectively integrating search and recommendation applications, and provides the possibility of achieving a more personalized user experience.

[0132] In an exemplary embodiment, step S440 obtains a second loss value of the sample user's intention to open the target application based on the sample long-term intention corresponding to each historical time period, including: obtaining the intention correlation score of the sample user corresponding to each historical time period; based on the intention correlation score corresponding to each historical time period and the sample long-term intention, obtaining the second loss value of the sample user's intention to open the target application.

[0133] Specifically, the intent relevance score corresponding to each historical time period and the sample long-term intention can be used as input variables, and the input variables are transformed into a Hawkes process through the constructed intensity function to obtain the log-likelihood value of the sample user's intention to open the target application as the second loss value.

[0134] As shown in Figure 2, the factors involved in the maximum likelihood estimation include the long-term intention of the samples corresponding to each historical time period. , ,… In addition, it also includes the intention relevance score of the sample users in each historical time period , ,… It can be understood that in the intention recognition model based on the neural Hawkes process, the intensity function is the core component based on the Hawkes process, which is mainly used to predict the possibility of future events based on the user's historical behavior information. The intensity function constructed in the present disclosure combines the result of time processing, the output of nonlinear transformation and the intention relevance score.

[0135] Among them, time processing means that in order to cope with the discontinuity of time information and the complexity of user behavior, the model uses a linear interpolation method to smooth the time variable, helping the model to better adapt to changes in time. Nonlinear transformation means that the model processes the hidden state through a specific type of nonlinear function, converting it into an actual numerical value to enhance the model's ability to capture complex user behavior patterns. The intention relevance score takes into account that the user's intention to open the app may be related to their search or interaction with recommended content. Therefore, a score reflecting the proportion of this interaction is added to the intensity function. At the same time, considering that the search ratio is usually low and may introduce noise, this score is batch normalized to ensure a consistent scale of the ratio, reduce the impact of noise and outliers, and make the model prediction more stable and accurate.

[0136] By combining the above elements to construct an intensity function, we can not only recognize the proportion of different intention behaviors, but also use the information learned from the encoding and decoding module, including the user's previous query and click behaviors, which are all encoded in the hidden state, realizing the use of historical information. Through this intensity function, the intent recognition model can predict the user's intention to open the app at a certain point in the future.

[0137] In this embodiment, considering that the intensity function of the Hawkes process is history-dependent, this embodiment uses the long-term intention h learned by the historical level encoding module (i.e., the decoding submodule of the second layer in the encoding and decoding module, corresponding to the Transformer decoder in Figure 2) containing rich historical interaction information to calculate a specific type of intensity function. At the same time, considering the time efficiency and the complexity of user behavior, a linear interpolation time module is used to determine the continuity of the conditional intensity function. In addition, a specific type of nonlinear transformation is also used to convert the hidden state The conversion to scalar increases the model’s ability to capture complex behavior patterns. In addition, considering the relevance characteristics of open application intent, an intent relevance score is added, which captures the proportion of user search and recommendation interactions in each time period, allowing for better intent prediction.

[0138] In an exemplary embodiment, in order to facilitate those skilled in the art to understand the embodiment of the present disclosure, the behavior prediction principle of the present disclosure will be further described below in conjunction with FIG. 2 .

[0139] As shown in FIG2 , the intention recognition model disclosed in the present invention is a model based on the neural Hawkes process, which can effectively capture the obvious periodicity and repeated query features in the intention of users to open applications in the past without assuming a fixed sampling time interval like time series modeling, so that it can better adapt to the prediction scenario of the intention to open applications in this application. In addition, unlike the previous Hawkes model, considering that the intention of users to open applications is affected by multiple factors, the present invention uses a hierarchical transformation module (corresponding to the Transformer encoder and Transformer decoder in FIG2 ) as a history encoder, which can not only learn the past sequence of intentions to open applications, but also learn the contextual representation of search and recommendation behaviors at the session level and history level. At the same time, when designing the intensity function, in order to capture the correlation characteristics, an intensity function that can be aware of the search and recommendation behavior ratio and the representation learned from the hierarchical transformer model is proposed. Finally, considering the significant impact of time factors and different user habits on the user's intention to open applications, a time gate is used to fuse time information with hierarchical user history representations and specific user embeddings to predict the user's intention to open applications. The intention recognition model specifically includes the following parts:

[0140] (1) Hierarchical transformation module

[0141] The hierarchical transformation module is used to implement the encoding of historical sequence behaviors in order to capture rich contextual information in user sessions. The specific implementation is as follows:

[0142] Embedding layer: Use embedding vectors to represent historical interaction information such as users, items, query terms, intent types for opening apps, and interaction types. Each type has its corresponding embedding representation, which is used to convert information such as users, items, query terms, etc. into vectors of fixed dimensions.

[0143] Session-level encoder: For each historical session of a user, the session-level encoder first integrates the embedding vectors of the items or queries in the session into a session embedding matrix. Then, through the Transformer encoder module, these session data are encoded into session-level representations that capture the user's intentions and behaviors in the current session.

[0144] History-level encoder: The history-level encoder is used to further encode all session-level representations to generate a representation of the entire history sequence. This process involves combining the session-level representation of each session with the corresponding open app intent and time information, and then processing it through the Transformer decoder module to generate a time-sensitive hidden representation of the history information.

[0145] Among them, the Transformer architecture used for historical behavior modeling can also be implemented using other architectures, such as a recurrent neural network (RNN) model, and the present disclosure does not make specific limitations on this.

[0146] This hierarchical transformer-based historical sequence encoding method can effectively combine time series data and user behavior data, thereby accurately capturing the complex patterns and periodic characteristics behind users' intentions to open apps without the need for fixed sampling time intervals.

[0147] (2) Intensity function

[0148] The intensity function is the core component of the Hawkes process, which is mainly used to predict the possibility of future events based on the user's historical behavior information. The principle of implementing the intensity function is explained as follows:

[0149] Utilization of historical information: The model uses information obtained from the history-level encoder, which includes the user’s previous query and click behaviors, etc., which are encoded in the hidden state.

[0150] Time processing: In order to cope with the discontinuity of time information and the complexity of user behavior, the model uses a linear interpolation method to smoothly process time variables, helping the model better adapt to changes in time.

[0151] Non-linear transformation: The model processes the hidden state through a specific type of non-linear function to transform it into an actual numerical value, which helps enhance the model’s ability to capture complex user behavior patterns.

[0152] Intent relevance score: Considering that the user's intention to open the app may be related to their search or interaction with the recommended content, the model adds a score that reflects the proportion of this interaction. This score is processed by batch normalization to reduce the impact of noise and outliers, making the model prediction more stable and accurate.

[0153] Combining the above elements, the model constructs an intensity function that combines the results of time processing, the output of nonlinear transformation, and the intent relevance score. Through this function, the model can predict the user's intention to open the app at a certain point in the future. Through this design, the intent recognition model can effectively combine historical data and real-time user behavior to provide support for predicting users' application usage intentions.

[0154] (3) Open application intent prediction layer

[0155] The prediction layer is used to predict the user's intention to open the app next time. This layer uses the Neural Hawkes Process model to improve performance by adding an additional prediction layer. This prediction layer incorporates the user's historical information, which includes the user's short-term and long-term intentions, as well as time information and user identification information. The specific implementation method is as follows:

[0156] Time gate method: The prediction layer uses the time gate method to integrate short-term and long-term intent. This method adjusts time-related parameters by comparing the current time with the time difference between the user's last interaction, so that the model can better understand the impact of time changes on user behavior.

[0157] Weight coefficient: A weight coefficient is calculated to combine the influence of long-term intention, short-term intention and time encoding. This vector helps the model decide which information should be considered more in the final prediction.

[0158] Fusion of user-specific information: In addition to historical intent and temporal information, the model also incorporates the user’s personal information. This is achieved by converting the user ID into a vector and processing this vector together with the comprehensive embedding of historical information.

[0159] Output prediction: Finally, the model uses an activation function (such as sigmoid) to combine all input information to predict the user's intention to open the app next time. This process also adds scores related to user intention to improve the accuracy of the prediction.

[0160] Through this method of combining information from multiple levels, the intent recognition model can accurately predict users' future behavioral intentions, effectively helping to improve the application's user experience and service quality.

[0161] The new neural Hawkes process-based model provided in this embodiment effectively utilizes the properties of the Hawkes process to capture the behavioral patterns behind the user's intention to open an application. This includes using periodic and repeated query features to solve the time-user bias problem. The model uses hierarchical transformers and novel intensity functions to encode multiple factors, and integrates time and user-specific information through an open application intention prediction layer to predict the user's intention to open an application, which can not only improve user interaction, but also improve performance in various downstream tasks, such as application open advertising and message pop-up recommendations. In addition, in order to demonstrate the superiority of the model and build a benchmark for the open application intention prediction task, the present disclosure not only expands the public search and recommendation dataset, but also constructs a new real-world open application intention dataset (OAMD). Through experiments on these two datasets, the superiority of the NHP-OAM model over the baseline model is verified, and the effectiveness of NHP-OAM in predicting user open application intentions is demonstrated through further downstream application experiments.

[0162] In addition, the effectiveness of this method can also be verified through experiments, as shown below:

[0163] (1) In terms of evaluation indicators. By collecting an offline dataset of 25,000 users’ behavioral data when using a short video application for verification, this method has achieved a significant improvement in the accuracy of the model’s prediction results compared to the current best baseline model. Specifically, for several core indicators, the accuracy rate increased by 0.58%, the precision increased by 3.01%, and the F1 score increased by 1.88%.

[0164] (2) In terms of improving user engagement. The intention of opening an app is crucial to optimizing user engagement. An experiment on a short video application showed that adjusting the search module from an inconspicuous position to a prominent position led to a significant increase in user interaction, increasing the proportion of opening the app for search from 0.1% to 0.8%. However, this change triggered competition between search and recommendation, with 87% of the gain coming from search and the return from recommended videos decreasing by 13%. By predicting the user's intention to open the app, the search and recommendation modules in the app can be dynamically adjusted, thereby alleviating the competition between search and recommendation and improving user engagement.

[0165] (3) In terms of improving the performance in downstream tasks. The intention of opening an app plays a key role in various downstream applications, including: (1) Based on the prediction of the intention of opening an app, more accurate guidance on the source of user interest (search or recommendation history) can be obtained in recommendation tasks that need to stimulate user interest in a very short time, such as advertisements and pop-up message recommendations when the app is opened. (2) Learning better user representations as powerful user features to enhance the performance of the recommendation system.

[0166] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0167] It can be understood that the same / similar parts between the various embodiments of the above method in this specification can refer to each other, and each embodiment focuses on the differences from other embodiments. For related points, please refer to the description of other method embodiments.

[0168] Based on the same inventive concept, the embodiment of the present disclosure also provides a behavior prediction device for implementing the above-mentioned behavior prediction method.

[0169] FIG5 is a structural block diagram of a behavior prediction device according to an exemplary embodiment. Referring to FIG5 , the device includes:

[0170] The information acquisition unit 510 is configured to acquire historical interaction information of the user in the target application in multiple historical time periods; the historical interaction information includes the user's intention to open the target application in each historical time period, time information, and interaction behavior information;

[0171] The information processing unit 520 is configured to execute the input of the historical interaction information of each historical time period into the encoding and decoding module in the intention recognition model for processing, and obtain the user's short-term intention and long-term intention; the short-term intention represents the user's intention to open the target application in the time period with the latest start time among multiple historical time periods, and the long-term intention represents the user's comprehensive intention to open the target application in multiple historical time periods;

[0172] The intention prediction unit 530 is configured to execute the input of the user's short-term intention and long-term intention into the intention prediction module in the intention recognition model for processing, and obtain the user's intention to open the target application at the predicted time point.

[0173] In an exemplary embodiment, the encoding and decoding module includes an encoding submodule and a decoding submodule; the information processing unit 520 is further configured to execute encoding processing of inputting historical interaction information of each historical time period into the encoding submodule respectively to obtain encoding information of each historical time period; inputting the encoding information of each historical time period, the starting time point of each historical time period and the intention of opening the target application in each historical time period into the decoding submodule for decoding processing to obtain decoding information of each historical time period; obtaining the user's short-term intention based on the encoding information of the latest time period with the latest start time among multiple historical time periods, and obtaining the user's long-term intention based on the decoding information of the latest time period.

[0174] In an exemplary embodiment, the intention prediction unit 530 is also configured to execute the acquisition of the user's user identification information; input the user's short-term intention, long-term intention and user identification information into the intention prediction module in the intention recognition model for processing, and obtain the user's intention to open the target application at the predicted time point.

[0175] In an exemplary embodiment, the intention prediction unit 530 is also configured to execute in the intention prediction module in the intention recognition model, fuse the short-term intention and the long-term intention to obtain the fused representation information of the user's intention; based on the fused representation information and the user identification information, determine the user's intention to open the target application at the predicted time point.

[0176] In an exemplary embodiment, the intention prediction unit 530 is further configured to execute the time difference between the latest time point corresponding to the latest time period in multiple time periods and the predicted time point; through the time difference, the short-term intention and the long-term intention are mapped and converted to obtain a weight coefficient; through the weight coefficient, the short-term intention and the long-term intention are fused to obtain fused representation information of the user's intention; the weight coefficient is used to characterize the degree of influence of the short-term intention on the fused representation information of the user's intention.

[0177] In an exemplary embodiment, the intent prediction module includes a multi-layer perceptron; the intent prediction unit 530 is further configured to execute the input of the fused representation information and the user identification information into the multi-layer perceptron for processing to obtain perception information; obtain the intent relevance score corresponding to the latest time period among multiple time periods, and determine the user's intention to open the target application at the predicted time point based on the intent relevance score and the perception information; the intent relevance score represents the ratio between the number of occurrences of different intentions in the same time period.

[0178] In an exemplary embodiment, the above-mentioned device also includes a model training module, which is configured to execute acquisition of sample interaction information of sample users in the target application in multiple historical time periods; the historical interaction information includes the user's real intention, time information and interaction behavior information to open the target application in each historical time period; the sample interaction information of each historical time period is input into the encoding and decoding module in the intention recognition model to be trained for processing, so as to obtain the sample short-term intention of the sample user and the sample long-term intention corresponding to each historical time period; the sample short-term intention of the sample user and the sample long-term intention corresponding to the latest time period in multiple historical time periods are input into the intention prediction module in the intention recognition model to be trained for processing, so as to obtain the predicted intention of the sample user to open the target application at the predicted time point, and obtain a first loss value based on the predicted intention and the real intention; and, based on the sample long-term intention corresponding to each historical time period, obtain a second loss value of the sample user's intention to open the target application; according to the first loss value and the second loss value, the intention recognition model to be trained is trained to obtain the intention recognition model.

[0179] In an exemplary embodiment, the model training module is also configured to obtain the intent relevance score corresponding to the sample user in each historical time period; based on the intent relevance score corresponding to each historical time period and the sample long-term intention, obtain a second loss value of the sample user's intention to open the target application.

[0180] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0181] Fig. 6 is a block diagram of an electronic device 600 for implementing an intention recognition method according to an exemplary embodiment. For example, the electronic device 600 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0182] 6 , electronic device 600 may include one or more of the following components: a processing component 602 , a memory 604 , a power component 606 , a multimedia component 608 , an audio component 610 , an input / output (I / O) interface 612 , a sensor component 614 , and a communication component 616 .

[0183] The processing component 602 generally controls the overall operation of the electronic device 600, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 602 may include one or more processors 620 to execute instructions to complete all or part of the steps of the above-mentioned method. In addition, the processing component 602 may include one or more modules to facilitate the interaction between the processing component 602 and other components. For example, the processing component 602 may include a multimedia module to facilitate the interaction between the multimedia component 608 and the processing component 602.

[0184] The memory 604 is configured to store various types of data to support operations on the electronic device 600. Examples of such data include instructions for any application or method operating on the electronic device 600, contact data, phone book data, messages, pictures, videos, etc. The memory 604 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, an optical disk, or a graphene memory.

[0185] The power supply component 606 provides power to the various components of the electronic device 600. The power supply component 606 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device 600.

[0186] The multimedia component 608 includes a screen that provides an output interface between the electronic device 600 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 608 includes a front camera and / or a rear camera. When the electronic device 600 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

[0187] The audio component 610 is configured to output and / or input audio signals. For example, the audio component 610 includes a microphone (MIC), and when the electronic device 600 is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 604 or sent via the communication component 616. In some embodiments, the audio component 610 also includes a speaker for outputting an audio signal.

[0188] I / O interface 612 provides an interface between processing component 602 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: a home button, a volume button, a start button, and a lock button.

[0189] The sensor assembly 614 includes one or more sensors for providing various aspects of status assessment for the electronic device 600. For example, the sensor assembly 614 can detect the open / closed state of the electronic device 600, the relative positioning of the components, such as the display and keypad of the electronic device 600, and the sensor assembly 614 can also detect the position change of the electronic device 600 or the electronic device 600 component, the presence or absence of contact between the user and the electronic device 600, the device 600 orientation or acceleration / deceleration and the temperature change of the electronic device 600. The sensor assembly 614 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly 614 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 614 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0190] The communication component 616 is configured to facilitate wired or wireless communication between the electronic device 600 and other devices. The electronic device 600 can access a wireless network based on a communication standard, such as WiFi, a carrier network (such as 2G, 3G, 4G or 5G), or a combination thereof. In an exemplary embodiment, the communication component 616 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 616 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0191] In an exemplary embodiment, the electronic device 600 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above methods.

[0192] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 604 including instructions, and the above instructions can be executed by the processor 620 of the electronic device 600 to complete the above method. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0193] In an exemplary embodiment, a computer program product is further provided. The computer program product includes instructions. The instructions can be executed by the processor 620 of the electronic device 600 to complete the above method.

[0194] It should be noted that the above-mentioned devices, electronic devices, computer-readable storage media, computer program products, etc. may also include other implementation methods according to the description of the method embodiments. The specific implementation methods can refer to the description of the relevant method embodiments, which will not be described one by one here.

[0195] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0196] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A behavior prediction method, characterized in that: include: Acquire historical interaction information of the user in the target application in multiple historical time periods; the historical interaction information includes the user's intention to open the target application in each of the historical time periods, time information, and interaction behavior information; Input the historical interaction information of each historical time period into the encoding and decoding module in the intention recognition model for processing to obtain the short-term intention and long-term intention of the user; the short-term intention represents the user's intention to open the target application in the time period with the latest start time among the multiple historical time periods, and the long-term intention represents the user's comprehensive intention to open the target application in the multiple historical time periods; The short-term intention and the long-term intention of the user are input into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point.

2. The method according to claim 1, characterized in that The encoding and decoding module includes an encoding submodule and a decoding submodule; the inputting of the historical interaction information of each historical time period into the encoding and decoding module in the intention recognition model for processing to obtain the short-term intention and long-term intention of the user includes: Inputting the historical interaction information of each historical time period into the encoding submodule for encoding processing respectively to obtain encoding information of each historical time period; Input the encoding information of each historical time period, the starting time point of each historical time period, and the intention of opening the target application in each historical time period into the decoding submodule for decoding processing to obtain decoding information of each historical time period; The short-term intention of the user is obtained based on the encoded information of the latest time period having the latest start time among the multiple historical time periods, and the long-term intention of the user is obtained based on the decoded information of the latest time period.

3. The method according to claim 1, characterized in that The inputting the short-term intention and the long-term intention of the user into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point includes: Obtaining user identification information of the user; The short-term intention, the long-term intention and the user identification information of the user are input into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point.

4. The method according to claim 3, characterized in that The inputting the short-term intention, the long-term intention and the user identification information of the user into the intention prediction module in the intention recognition model for processing to obtain the user's intention to open the target application at the predicted time point, comprises: In the intention prediction module in the intention recognition model, the short-term intention and the long-term intention are fused to obtain fused representation information of the user's intention; Based on the fused representation information and the user identification information, the user's intention to open the target application at the predicted time point is determined.

5. The method according to claim 4, characterized in that The fusing the short-term intention and the long-term intention to obtain fused representation information of the user intention includes: Obtaining a time difference between a latest time point corresponding to a latest time period among the multiple time periods and the predicted time point; By using the time difference, mapping and converting the short-term intention and the long-term intention to obtain a weight coefficient; the weight coefficient is used to characterize the influence of the short-term intention on the fusion representation information of the user intention; The short-term intention and the long-term intention are fused by using the weight coefficient to obtain fused representation information of the user intention.

6. The method according to claim 4, characterized in that The intention prediction module includes a multi-layer perceptron; the determining the user's intention to open the target application at the predicted time point based on the fused representation information and the user identification information includes: Inputting the fusion representation information and the user identification information into the multi-layer perceptron for processing to obtain perception information; Obtain an intention relevance score corresponding to the latest time period among the multiple time periods, and determine the user's intention to open the target application at the predicted time point based on the intention relevance score and the perception information; the intention relevance score represents the ratio between the number of occurrences of different intentions in the same time period.

7. The method according to claim 1, characterized in that The intent recognition model is trained in the following way: Acquire sample interaction information of sample users in a target application in multiple historical time periods; the historical interaction information includes the user's real intention, time information, and interaction behavior information for opening the target application in each of the historical time periods; Input the sample interaction information of each historical time period into the encoding and decoding module of the intention recognition model to be trained for processing, so as to obtain the sample short-term intention of the sample user and the sample long-term intention corresponding to each historical time period; Input the sample short-term intention of the sample user and the sample long-term intention corresponding to the latest time period among the multiple historical time periods into the intention prediction module in the intention recognition model to be trained for processing, obtain the predicted intention of the sample user to open the target application at the predicted time point, and obtain a first loss value based on the predicted intention and the true intention; and obtain a second loss value of the sample user's intention to open the target application based on the sample long-term intention corresponding to each of the historical time periods; The intention recognition model to be trained is trained according to the first loss value and the second loss value to obtain the intention recognition model.

8. The method according to claim 7, characterized in that The obtaining, based on the sample long-term intention corresponding to each of the historical time periods, a second loss value of the sample user's intention to open the target application includes: Obtaining the intention relevance score of the sample user corresponding to each of the historical time periods; Based on the intention relevance score corresponding to each of the historical time periods and the sample long-term intention, a second loss value of the sample user's intention to open the target application is obtained.

9. A behavior prediction device, characterized in that: include: An information acquisition unit is configured to acquire historical interaction information of a user in a target application in multiple historical time periods; the historical interaction information includes the user's intention to open the target application in each of the historical time periods, time information, and interaction behavior information; an information processing unit configured to execute inputting the historical interaction information of each historical time period into a coding and decoding module in an intention recognition model for processing, so as to obtain a short-term intention and a long-term intention of the user; the short-term intention represents the intention of the user to open the target application in the time period with the latest start time among the multiple historical time periods, and the long-term intention represents the comprehensive intention of the user to open the target application in the multiple historical time periods; The intention prediction unit is configured to execute the input of the short-term intention and the long-term intention of the user into the intention prediction module in the intention recognition model for processing, so as to obtain the intention of the user to open the target application at the predicted time point.

10. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the behavior prediction method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the behavior prediction method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the behavior prediction method according to any one of claims 1 to 8 are implemented.

Images