Training method and device of intention recognition model, electronic equipment and storage medium

Abstract

The invention discloses an intention recognition model training method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining historical behavior data which comprises interaction data, user attributes and application attributes, and the interaction data is generated when a target user operates a target terminal containing a target application; the target user corresponds to the user attribute, and the target application corresponds to the application attribute; performing filtering processing on the interaction data to obtain processed interaction data, and performing feature extraction on the processed interaction data to obtain user behavior representation; performing feature extraction on the user attribute, the application attribute and the terminal attribute of the target terminal, and splicing the obtained feature representation with the user behavior representation to obtain a historical behavior representation; and training the intention recognition model according to the historical behavior representation to obtain a trained intention recognition model. The intention recognition is performed based on the interaction behaviors, and the intention recognition accuracy is improved.

Description

Technical Field

[0001] The present application relates to the technical field of intent recognition processing, and in particular to a training method, device, electronic device and storage medium for an intent recognition model. Background Art

[0002] In the era of intelligent interconnection of all things, with the continuous emergence of billions of connected services, how to achieve accurate service distribution has become a core issue. In the field of mobile phones, intent recognition technology has become an increasingly popular solution to this problem, which mainly helps machines understand and recognize users' intentions and needs, and then provide corresponding feedback and responses.

[0003] When performing intent recognition, the main problem is how to accurately provide users with the services they need. For example, at 7:03 p.m., a white-collar worker in his 20s boarded the subway. At this time, when recognizing the user's intention, it is usually predicted that the user's needs at this time are to listen to music or watch videos. At the same time, when determining the user's actual intention, the user is required to perform corresponding operations, such as voice input or behavioral operation input, and then analyze and match to determine the current required services based on the rule engine. This makes intent recognition relatively passive, and the recognition effect and accuracy of intent are low. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a training method, device, electronic device and storage medium for an intent recognition model to solve the technical problems of low effect and accuracy in user intent recognition in related technologies.

[0005] In a first aspect, an embodiment of the present application provides a method for training an intent recognition model, comprising:

[0006] Acquire historical behavior data, wherein the historical behavior data includes interaction data, user attributes, and application attributes, and the interaction data is generated by a target user operating a target terminal including a target application, the target user corresponds to the user attributes, and the target application corresponds to the application attributes;

[0007] Performing filtering processing on the interaction data to obtain processed interaction data, and performing feature extraction on the processed interaction data to obtain user behavior representation;

[0008] Extracting features from the user attributes, the application attributes, and the terminal attributes of the target terminal, and combining the obtained feature representation with the user behavior representation to obtain a historical behavior representation;

[0009] The intent recognition model is trained according to the historical behavior representation to obtain a trained intent recognition model.

[0010] In a second aspect, an embodiment of the present application provides a training device for an intent recognition model, comprising:

[0011] A data acquisition module, used to acquire historical behavior data, wherein the historical behavior data includes interaction data, user attributes and application attributes, and the interaction data is generated by a target user operating a target terminal including a target application, the target user corresponds to the user attributes, and the target application corresponds to the application attributes;

[0012] A first processing module is used to filter the interaction data to obtain processed interaction data, and extract features from the processed interaction data to obtain user behavior representation;

[0013] A second processing module is used to extract features from the user attributes, the application attributes and the terminal attributes of the target terminal, and to combine the obtained feature representation with the user behavior representation to obtain a historical behavior table;

[0014] The model training module is used to train the intent recognition model according to the historical behavior representation to obtain a trained intent recognition model.

[0015] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in the training method of the intent recognition model described in any one of the above items when executing the computer program.

[0016] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps in the training method of the intent recognition model described in any one of the above.

[0017] The embodiment of the present application provides a training method, device, electronic device and storage medium for an intent recognition model, which obtains the user's actual interaction data, that is, historical behavior data, including interaction data, user attributes and application data, and extracts features after filtering the historical behavior data to obtain user behavior representation, and then trains the intent recognition model in combination with user attributes, terminal attributes and application attributes. It reduces the noise in the historical behavior data through filtering during the training process, improves the accuracy of model training, and uses user behavior data as the basis for user intent recognition, making intent recognition more accurate and closer to user needs, effectively improving the accuracy of intent recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1It is a flowchart of a training method for an intent recognition model provided in an embodiment of the present application;

[0019] Figure 2 is a flowchart of steps for obtaining user behavior representation provided in an embodiment of the present application;

[0020] Figure 3 is a flowchart of steps for obtaining user behavior representation provided in an embodiment of the present application;

[0021] Figure 4 It is a schematic block diagram of a training deployment system provided in an embodiment of the present application;

[0022] Figure 5 It is a block diagram schematic diagram of a method for generating a training set provided in an embodiment of the present application;

[0023] Figure 6 It is a schematic diagram of the structural block diagram of the intent recognition model and the data processing flow block diagram provided in the embodiment of the present application;

[0024] Figure 7 It is a schematic block diagram of an extractor for extracting features from user behavior data provided in an embodiment of the present application;

[0025] Figure 8 It is a schematic diagram of a structural block diagram of a multi-layer perceptron provided in an embodiment of the present application;

[0026] Fig. 9 It is a schematic diagram of the block diagram of the deployment and use of the model provided in the embodiment of the present application;

[0027] Fig.10 It is a schematic diagram of a block diagram of an Android terminal model reasoning process provided in an embodiment of the present application;

[0028] Fig.11 It is a structural schematic diagram of a training device for an intent recognition model provided in an embodiment of the present application;

[0029] Fig.12 is a structural schematic diagram of an electronic device provided in an embodiment of the present application;

[0030] Fig.13 This is another structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0032] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0033] The term "including" and its variations used herein are open inclusions, i.e., "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description.

[0034] In related technologies, when performing intent recognition, a pre-trained related model is usually used to recognize and predict intent. When obtaining a related intent recognition model, a related rule engine is usually set up, and then based on the user's active input, such as voice input, the active input information is analyzed and recognized, combined with the set rule engine, to determine the user's current intent, and then make service recommendations. This method is limited to the setting of the rule engine, and the user actively initiates and needs to input relevant demand information, which makes the intent recognition not accurate enough, and the user's active participation will lead to a poor user experience.

[0035] Therefore, there is an urgent need for a method that can proactively determine user needs and perform intent recognition, while also being able to better improve the accuracy of intent recognition.

[0036] In order to solve the technical problems existing in the related art, the present application embodiment provides a training method for an intent recognition model, see Figure 1 , Figure 1 It is a flow chart of a training method for an intent recognition model provided in an embodiment of the present application, and the method includes steps 101 to 104.

[0037] Step 101, obtaining historical behavior data, wherein the historical behavior data includes interaction data, user attributes and application attributes, and the interaction data is generated by a target user operating a target terminal including a target application, the target user corresponds to the user attributes, and the target application corresponds to the application attributes.

[0038] In one embodiment, when training a model, sample data for model training is first determined. Specifically, the sample data for training includes at least interactive data of a user operating a terminal device, that is, data generated when a user interacts with a target terminal when using the target terminal. At the same time, in order to facilitate more accurate intent recognition in the future, the sample data also includes at least user attributes and application attributes, wherein user attributes are relevant information of the user who actually operates the terminal device, such as age, gender, etc., and application attributes are application information of all applications installed on the terminal, such as application categories, etc. That is, the sample data obtained for training is historical behavior data including at least interactive data, user attributes, and application attributes.

[0039] It should be noted that the interaction data is generated by the target user operating the target terminal including the target application. The target user corresponds to the user attribute, and the target application corresponds to the application attribute.

[0040] In actual applications, the interaction data obtained includes at least the user's app switching behavior (app A->app B), application installation and uninstallation behavior, and the user's screen turning on and off behavior, etc. Operations within the application are not obtained, such as frequent click operations during instant messaging.

[0041] Step 102: filter the interaction data to obtain processed interaction data, and extract features from the processed interaction data to obtain user behavior representation.

[0042] In one embodiment, after obtaining the historical behavior data, it will be processed accordingly to ensure better model training. Specifically, after obtaining the historical behavior data, since the interaction data is the actual interaction data of the user, in order to ensure the accuracy of the data, the data needs to be processed accordingly. At this time, it can be filtered, and then features are extracted to obtain the user behavior representation corresponding to the interaction data.

[0043] In practical applications, the interactions between users and terminal devices are diverse, and whether the interaction information is valid or repeated requires accurate judgment to avoid the impact of invalid or repeated interaction data on model training.

[0044] Furthermore, when processing the interaction data to obtain the corresponding user behavior representation, you can refer to Figure 2 , Figure 2 It is a flowchart of the steps of obtaining user behavior representation provided in an embodiment of the present application, wherein the steps include steps 201 to 204.

[0045] Step 201, performing behavior recognition on the interaction data, and obtaining a behavior coding matrix of the interaction data according to the recognition result;

[0046] Step 202, performing behavior position analysis according to the behavior coding matrix to obtain a position coding matrix of the interaction data;

[0047] Step 203, performing matrix addition on the position coding matrix and the behavior coding matrix to obtain a behavior matrix of the interaction data;

[0048] Step 204 , filtering and feature extraction processing are performed on the behavior matrix to obtain user behavior representations corresponding to the interaction data.

[0049] When processing interactive data, the reliability of the data is improved by identifying the interactive behaviors contained therein and then performing screening and filtering. Specifically, the interactive data is subjected to behavioral identification, and the behavioral coding matrix of the interactive data is obtained based on the identification results. The behavioral position analysis is then performed based on the behavioral coding matrix to determine the position coding matrix of the interactive data. The behavioral coding matrix and the position coding matrix are then concatenated to obtain the behavioral matrix of the interactive data. Finally, the behavioral matrix is ​​filtered and feature extracted to obtain the user behavior representation corresponding to the interactive data.

[0050] For example, when performing behavior recognition on interaction data, all behaviors contained therein are recognized. However, due to possible misoperation or frequent use in a short period of time, the recognized behaviors may be inaccurate, such as when a user opens the same application multiple times within a certain period of time, or when a certain application is used for a short time, such as 5 seconds. In this case, the data needs to be filtered to improve the accuracy of the data. After completing the filtering process of the behaviors, a matrix corresponding to the behaviors is constructed to record the behaviors contained in the interaction data, and a behavior coding matrix is ​​obtained.

[0051] Specifically, when performing behavior recognition on interaction data to obtain a behavior coding matrix, it includes: performing behavior recognition on the interaction data, and obtaining behavior information of the interaction data based on the recognition result; screening and deduplication processing on the initial behavior information to obtain processed behavior information; and constructing based on the behavior information based on the time sequence to obtain a behavior coding matrix using the behavior data.

[0052] Among them, when processing, first of all, all behaviors contained in the interactive data are identified, and the identified behaviors include but are not limited to: APP switching, APP startup, APP installation and uninstallation, and screen on and off operations, so as to obtain corresponding behavior information, and then the behavior information is screened and deduplicated, and the erroneous operation behaviors or frequent behaviors in the behavior information are filtered or deduplicated to complete the cleaning of the behavior information, and then according to the data after the cleaning, the matrix is ​​constructed according to the time series information to obtain the corresponding behavior coding matrix.

[0053] Exemplarily, when cleaning the interaction data, the focus is on frequent behaviors and invalid behaviors in the data, where frequent behaviors are behaviors that occur multiple times in a short period of time, such as a user opening an application multiple times in a short period of time; invalid behaviors are behaviors that exist for an extreme amount of time, such as a user using an application for an extreme amount of time (e.g., 5 seconds). By cleaning the data, the cleanliness of the data is improved, thereby making the model training more accurate.

[0054] After the behavior information is cleaned, the corresponding behavior coding matrix will be constructed based on the order of interaction data records and the processed behavior information. The behavior coding matrix records the behaviors contained in the interaction data and can be described in other ways.

[0055] Furthermore, when the behavior coding matrix of the interaction data is obtained, the position information corresponding to the behavior will also be determined, wherein the position information is the position of the behavior in the interaction data, which can be specifically time information. At this time, when the position information is determined and the position coding matrix is ​​obtained, it includes: determining the timestamp of each behavior recorded in the behavior coding matrix according to the time information of the interaction data; and constructing the position coding matrix of the interaction data according to the timestamp.

[0056] Exemplarily, when determining the position coding matrix, the time information of the behaviors recorded in the behavior coding matrix in the interaction data is determined to determine the timestamps of each behavior, and then the corresponding position coding matrix is ​​constructed based on the timestamps.

[0057] After obtaining the behavior coding matrix and the position coding matrix, the two are added together to obtain a behavior matrix for representing user behaviors in the interaction data. Then, by processing the behavior matrix, a user behavior representation containing behavioral characteristics is obtained.

[0058] It should be noted that the behavior encoding matrix, position encoding matrix and behavior matrix can be an embedding representation. By defining representations under different requirements, different matrices are obtained to represent behaviors and positions, etc., to facilitate subsequent processing and model training.

[0059] Furthermore, when processing the behavior matrix to obtain the user behavior representation, you can refer to Figure 3 , Figure 3 It is a flowchart of the steps of obtaining user behavior representation provided in an embodiment of the present application, wherein the steps include steps 301 to 303.

[0060] Step 301, extracting features from the behavior matrix in the frequency domain to obtain feature representation in the frequency domain;

[0061] Step 302, denoising the feature representation using a gradient descent method to obtain a denoised feature representation;

[0062] Step 303: Perform a time domain transformation on the denoised feature representation to obtain a user behavior representation in the time domain.

[0063] Specifically, when processing the behavior matrix, it is subjected to feature extraction and denoising to obtain more accurate feature information, i.e., user behavior representation. When processing specifically, the behavior matrix is ​​first subjected to feature extraction in the frequency domain to obtain a feature representation in the frequency domain, and then denoised using a filter, and denoised using the gradient descent method to obtain a feature representation after denoising, and finally subjected to a time domain transformation to obtain a user behavior representation in the time domain.

[0064] For example, for the processing of features and signals, processing in the frequency domain is more convenient, so when processing, the processing efficiency is improved by processing in the frequency domain, and then conversion is performed to obtain the user behavior representation in the time domain.

[0065] When processing, the obtained behavior matrix is ​​first Fourier transformed, then filtered, and finally inverse Fourier transformed to obtain the user behavior representation. When the behavior matrix is ​​represented by embedding, when processing, Fourier transform is performed on each dimension represented by embedding. Each dimension represented by embedding forms a discrete sequence in the time domain, and filtering operations are performed on the spectrum corresponding to each sequence.

[0066]

[0067] Among them, X l is the frequency domain representation of the sequence, F I ∈R n×d is a behavioral matrix, n is the maximum length of the sequence, and d is the dimension.

[0068] Then, the learnable filter is passed through the complex matrix W n×dTo achieve this, signal adjustment is performed in the frequency domain, the gradient is set for the filter matrix, and self-adjustment can be achieved through the gradient descent method to remove signal noise and adaptively reduce the signal corresponding to the high-frequency application in the sequence.

[0069]

[0070] When processing, multiplication in the frequency domain is equivalent to circular convolution in the time domain, which is equivalent to one-dimensional convolution of two functions with the same period. Therefore, the convolution operation with a field of view of sequence length n can be implemented through the filter, realizing context-based feature extraction.

[0071] Finally, the inverse Fourier transform is used to convert back to the time domain representation.

[0072]

[0073] Finally, when obtaining the user behavior representation, we can establish an attention mechanism for each application in the application and interaction data, pay attention to the application-related interest migration, and use the weighted average method to obtain the user behavior representation.

[0074] Step 103 , extracting features from user attributes, application attributes, and terminal attributes of the target terminal, and combining the obtained feature representation with the user behavior representation to obtain a historical behavior representation.

[0075] In one embodiment, after completing the processing of the interaction data and obtaining the user behavior representation, the input data for model training will be constructed in combination with the user's relevant information, wherein the user's relevant information at least includes the user's attributes, the attributes of the target terminal used, and the attributes of the target application contained in the target terminal. Specifically, feature extraction is performed on the user attributes, the target application, and the target terminal respectively, and then the obtained feature representation is spliced ​​with the user behavior representation corresponding to the interaction data to obtain the historical behavior representation as the input for model training.

[0076] Among them, when extracting features of user attributes, target applications and target terminals respectively, the features of user attributes obtained include gender and age, etc., the terminal attributes of the target terminal include terminal power, WiFi connection status and Bluetooth connection status, etc., and the features of the target application are the categories of the application programs. That is, under normal circumstances, user attributes and target application attributes are fixed data, while target terminal attributes are variable.

[0077] Step 104 , training the intent recognition model according to the historical behavior representation to obtain a trained intent recognition model.

[0078] In one embodiment, after feature extraction is performed to obtain historical behavior representations, the historical behavior representations are used as input for model training to train the intent recognition model, and then a trained intent recognition model is obtained when the training is completed. There is no restriction on the method for judging whether the training is completed, which can be the number of training times or the use of a loss function, etc.

[0079] In actual applications, the interaction data records the interaction data when the user operates the target terminal. Generally speaking, the interaction data is the situation of the user using the application. In real life, there are many target applications installed in the target terminal, but not all applications are used in the same way, that is, the frequency of use will vary greatly. When the interaction data is obtained, popular high-frequency applications will occupy most of the behavior information, while unpopular applications may not be started for a week or a month. Therefore, in order to prevent the trained intent recognition model from being overly biased towards high-frequency applications, the positive sampling probability of the application is adjusted before training. By suppressing the probability of positive sampling of popular applications, the model is prevented from being overly biased towards people's high-frequency applications.

[0080] Specifically, when training the intent recognition model, it includes: analyzing and counting the interaction data, and adjusting the positive sampling probability of each application in the target application according to the set probability adjustment method to obtain the positive sampling probability of the target application; training the intent recognition model according to the positive sampling probability and historical behavior representation, and obtaining a trained intent recognition model when the training is completed.

[0081] Exemplarily, it can be seen from the above description that before training, the positive sampling probability of each target application included in the target terminal is adjusted to prevent overfitting of the model and bias towards high-frequency applications for intent recognition and prediction.

[0082] When processing, the adjustment is performed according to the set adjustment method, wherein after the adjustment, the application APP (ω i ) becomes a positive sample:

[0083]

[0084] in, X is clicked i The number of users, Y is the total number of users who have clicked in the same period, z(w i ) indicates APP(ω i ) usage rate, α is a hyperparameter, generally set in [1e -5 ,1e -3 ].

[0085] Based on the above adjustment method, the higher the usage rate of the app, the lower the corresponding positive sample sampling rate. Then, after the adjustment is completed, the intent recognition model is trained based on the positive sampling probability and historical behavior representation of each application.

[0086] Furthermore, after the training is completed, the intent recognition model obtained through training can be verified and evaluated to determine whether the model is accurate. Specifically, during the verification and evaluation process, the intent recognition output is performed to obtain the user's current intent, and service recommendations are made based on the identified and predicted intent, and then the evaluation is completed based on the user's feedback.

[0087] For example, when evaluating, the intent recognition model will recognize the intent based on the user's behavior data in the previous period, and then recommend and display services based on the recognition results. At this time, the user's operation on the recommended service, such as selection or click, can be obtained, and the model is evaluated through user feedback. For example, when the user gives feedback, if the model's intent recognition and service recommendation are selected by the user every time, it means that the model's intent recognition effect is good, otherwise further processing may be required.

[0088] Furthermore, after completing the training to obtain the trained intent recognition model, when used for intent recognition, it includes: when performing intent recognition, obtaining behavior data within a preset time period based on the current time; performing feature extraction on the behavior data to obtain behavior features corresponding to the behavior data, and inputting the behavior features into the trained intent recognition model to obtain several recognition results and probability values ​​for each recognition result; determining recommended services in the recognition results according to the probability values, and generating corresponding service cards for display based on the recommended services.

[0089] In one embodiment, when performing intent recognition, it is actively based on the user's historical behavior data to perform intent recognition and service recommendation. Therefore, at this time, based on the current time, the behavior data within the preset time period is obtained, and then the behavior data is feature extracted to obtain the behavior features corresponding to the behavior data, and then the trained intent recognition model is used to perform intent recognition processing. When performing intent recognition processing, the result obtained is the predicted probability of different services, that is, the probability value of the service currently required by the user is determined through intent recognition, and then the service is recommended and displayed according to the size of the probability value, such as displaying and recommending the three services with the highest probability value. As for the obtained behavior data, the data one and a half hours before the current time can be obtained as the behavior data.

[0090] It should be noted that for the trained intent recognition model, it will be regularly updated during use based on actual predictions and user feedback. The model will be regularly updated by obtaining the prediction results and feedback data during the prediction process. When updating, the method described in steps 101 to 104 can also be referred to. The difference is that the update is for the existing and used intent recognition model.

[0091] Furthermore, based on the above method, a system for training and deploying a model is also provided, referring to Figure 4 , Figure 4 It is a block diagram of a training deployment system provided in an embodiment of the present application.

[0092] The entire processing process can be divided into four parts, namely data collection, data processing, algorithm model, training evaluation and model deployment. Among them, the data obtained during data collection includes static data and user behavior data (i.e., the interaction data described above), and the static data is divided into user attributes and APP attributes. The user behavior data is the behavior data of operating the terminal, such as switching APPs, installing and uninstalling applications, and turning the terminal screen on and off, etc. Data processing includes spark data processing and python data processing, and the algorithm model is the intention recognition model described in each embodiment.

[0093] When training the model, data collection and processing are performed first. Data collection and processing are performed on the data acquisition side by embedding points. User behavior data such as app switching behavior (app A->app B), application installation and uninstallation behavior, and user screen on and off are collected. At the same time, user attributes (age, gender, etc.) and app attributes (app category) are also obtained. The acquired data is then processed, specifically feature extraction. Finally, the constructed algorithm model (intent recognition model) is trained and evaluated based on the extracted features based on the training and evaluation module to obtain a trained and optimized intent recognition model. Finally, the trained intent recognition model is deployed on the corresponding terminal or device through the model deployment module.

[0094] Then, when processing the data, a small number of popular applications account for most of the clicks based on the distribution of application usage times in each frequency band. Therefore, in order to prevent the model from being biased towards high-frequency applications, popular applications are suppressed when sampling positive samples, so that the more popular the application, the lower its probability of becoming a positive sample.

[0095] Appω i The probability of being a positive sample:

[0096]

[0097] in, represents the usage rate of the app, a is a hyperparameter, generally set in [1e -5 ,1e -3 ].

[0098] Then, perform python data processing, for details, please refer to Figure 5 ,exist Figure 5 The data processing process shown includes first obtaining a user feature set (user attributes), an item feature set (behavior attributes) and a user installation list merge (application attributes), and then using hdfs (Hadoop Distributed File System) to preprocess the obtained data, and determine negative sample data based on the interaction data and the user application installation set in hdfs, and then perform feature extraction to obtain the features of the negative sample data. At the same time, the interaction data can be stored in CSV (Comma-Separated Values), and finally generate a user-item interaction set (user behavior data) based on a streaming processing method, and process the user-item interaction set based on the loss type to generate corresponding positive and negative samples to avoid data overfitting. At the same time, the processed data can be divided into data to obtain training set data and verification set data, wherein the training set data can account for 80% and the verification set data can account for 20%.

[0099] When constructing and training the algorithm model, the constructed intent recognition model can refer to Figure 6 , Figure 6 It is a schematic diagram of the structural block diagram and data processing flow diagram of the intent recognition model provided by the embodiment of the present application. For the entire intent recognition model, it mainly includes two parts, namely: a feature extraction module (composed of feature extractor) and a click probability prediction module (composed of multilayer perception), wherein the feature extraction module is used to extract features from historical behavior data, and the click probability prediction module is used to make predictions based on the obtained features.

[0100] At the same time, based on Figure 6As shown, the input of model training consists of four parts: user profile, context, user behavior session, and target action. User profiles include gender and age, context information includes mobile phone battery level, wifi connection, and Bluetooth connection. Context is the terminal attribute of the target terminal described in the above embodiment, and candidate application attributes include first, second, and third level category labels. When performing identification and prediction, the click probability of each candidate application is inferred one by one, and all candidate applications are sorted from large to small based on click probability, and the top 3 probabilities are taken out for display.

[0101] For the constructed model, in order to improve the accuracy of model recognition, the feature extraction module can set different structures and extraction methods for different objects when performing feature extraction. Specifically, different structures and feature extraction methods are used when extracting special features from user behavior data.

[0102] When extracting features from user attributes, context information, and candidate applications, you can use embedding table mapping to obtain them. By converting static attributes such as user attributes and APP attributes into discrete features, such as continuous features such as power and age, you can discretize them in a bucket-based manner. Each attribute gets the corresponding embedding vector from the preset embedding table based on the discretized value, and then concatenates the embedding vectors of each attribute as the final feature representation.

[0103] When extracting features from user behavior data, the conventional feature extraction method cannot be used because the data contains noise information. It is necessary to use a reasonable method to filter out the noise in the data to improve the reliability of the data. Specifically, the structure of the feature extraction module for extracting features from user behavior data can refer to Figure 7 , Figure 7 It is a block diagram of an extractor for extracting features from user behavior data provided in an embodiment of the present application.

[0104] For the feature extractor of user behavior data, an interest extraction layer and an interest transfer layer are designed.

[0105] In the interest extraction layer, the context-based modeling method is followed, and the Fourier transform + learnable filter + inverse Fourier transform structure is adopted. The input is the item embedding representation (behavior encoding matrix) of each application in the user behavior data. At the same time, in order to simulate the temporal relationship between behaviors in the behavior sequence, the position embedding representation (position encoding matrix) is defined. The position encoding matrix dimension is (n, d), where n is the maximum length of the sequence, and d is the dimension of the embedding representation after splicing. The position encoding value is obtained through model learning.

[0106] Then add the item embedding representation and the position embedding representation, and use F l ∈R n×d Represented as input to the interest extraction layer. Perform Fourier transform on each dimension represented by embedding. Each dimension represented by embedding forms a discrete sequence in the time domain. Perform filtering operations on the spectrum corresponding to each sequence:

[0107]

[0108] Among them, X l is the frequency domain representation of the sequence, F l ∈R n×d is a behavioral matrix, n is the maximum length of the sequence, and d is the dimension.

[0109] Then, the learnable filter is passed through the complex matrix W n×d To achieve this, signal adjustment is performed in the frequency domain, the gradient is set for the filter matrix, and self-adjustment can be achieved through the gradient descent method to remove signal noise and adaptively reduce the signal corresponding to the high-frequency application in the sequence.

[0110]

[0111] When processing, multiplication in the frequency domain is equivalent to circular convolution in the time domain, which is equivalent to one-dimensional convolution of two functions with the same period. Therefore, the convolution operation with a field of view of sequence length n can be implemented through the filter, realizing context-based feature extraction.

[0112] Finally, the inverse Fourier transform is used to convert back to the time domain representation.

[0113]

[0114] In the interest transfer layer, an attention mechanism is established for candidate applications and each application in the user's historical behavior sequence, focusing on the interest transfer related to the candidate applications, and using the weighted average method to obtain the final user behavior representation.

[0115] Finally, in the click probability prediction module, the final feature information is obtained by splicing the user representation, context representation, candidate application representation and user behavior representation, and then the model is trained based on the feature information. Specifically, the obtained feature information is sent to the multilayer perceptron (MLP) for implicit feature fusion. At this time, the corresponding inference result can be obtained based on the input feature data, and then the accuracy of the inference result can be determined based on the historical behavior data. The structure diagram of MLP can be referred to Figure 8 , Figure 8 It is a schematic diagram of the structural block diagram of the multi-layer perceptron provided in an embodiment of the present application.

[0116] Specifically, the MLP consists of two layers of multilayer perceptrons (mlp layer) followed by a classifier (linear). Each layer of multilayer perceptron (mlp layer) is composed of dropout, linear fully connected layer, BatchNorm and activation function (such as tanh) in sequence. The click rate predictor is a binary classification, which directly outputs the probability of the application being clicked at the next moment.

[0117] When evaluating the model, the evaluation is completed through interaction with the user. Specifically, during the evaluation, identification and prediction are performed based on the user's behavior data. At this time, the three applications recommended to the user contain the probability that the user actually uses the application. In order to further verify the improvement of the learnable filter on the indicators of medium and low-frequency applications, on the basis of the overall recall@3 indicator, the recall indicator distribution of applications with different usage frequencies, the recall indicator distribution of different application types, and the recall indicator distribution of type and frequency are statistically analyzed to complete the evaluation of the model.

[0118] For the intent recognition model obtained after training and evaluation, the corresponding model deployment method can be used to deploy it on the corresponding terminal or device, so that the terminals and devices deployed with the intent recognition model can complete the relevant intent recognition operations.

[0119] In practical applications, the constructed intent recognition model can be constructed using the torch deep learning framework model, and then corresponding training and evaluation are performed to obtain an intent recognition model based on the torch deep learning framework. Then, when the intent recognition model is deployed, since it needs to rely on the tflite engine when used on the terminal side, the model format needs to be converted during deployment. Specifically, the model format conversion can be implemented by following the standard format conversion process of torch->onnx->pb->tflite, so that the intent recognition model (tflite format intent recognition model) finally deployed on the terminal can use the tflite native operator library operator, so that when performing intent recognition processing, the speed of intent recognition and reasoning can be accelerated, and the dependence on other libraries can be avoided to cause a sudden increase in the apk size.

[0120] Reference Fig. 9 , Fig. 9 This is a schematic diagram of the model deployment and use provided by the embodiment of the present application. Fig. 9 In the processing process shown, for the intent recognition model built based on the torch architecture, after completing the training and evaluation of the model, when performing the conversion, the torch operator can be customized, and then the format conversion of the model is completed according to the deployed model conversion method to obtain the model file in the tflite format, and then the deployment on the terminal device is completed by loading. When the terminal performs intent recognition, the user behavior data is first obtained, and the user behavior data is converted into a data format type that can be currently recognized and processed through data format conversion, and then subsequent processing, such as feature extraction and matching, is performed to obtain the current intent recognition result, that is, the inference result of the model.

[0121] Furthermore, taking Android as an example, the model inference process on the Android side refers to Fig.10 In order to ensure that the size of the model file corresponding to the converted tflite format intent recognition model is as small as possible, the sorted tflite format intent recognition model can strip the embedding layer and put the embedding layer mapping work in the data processing part before model reasoning. For the embedding table of large-scale ids (such as the embedding of application ids), the end side only saves the embedding of the part of applications that the user has installed, which can greatly reduce the space occupied by the end side.

[0122] After the deployment of the intent recognition model is completed, historical behavior data is obtained during the intent recognition process, wherein the historical behavior data includes user attributes (user profile), context (context), and user historical behavior sequence (user behavior session), and feature extraction is performed on the embedding table side based on the mapping of the embedding layer to obtain user profile embedding (user features), context embedding (context features) and user behavior session embedding (user behavior representation), and then the loaded intent recognition model is used to complete the recognition and reasoning of the intent, and the obtained reasoning result (intent recognition result) is the relevant instructions and services for the operation of the part of the application installed by the user.

[0123] In summary, the present application discloses a method for training an intent recognition model, which obtains the user's actual interaction data, that is, historical behavior data, including interaction data, user attributes and application data, and extracts features after filtering the historical behavior data to obtain user behavior representation, and then trains the intent recognition model in combination with user attributes, terminal attributes and application attributes. The method reduces the noise in the historical behavior data and the overfitting effect of high-frequency data on the model through filtering during the training process, thereby improving the accuracy of model training. At the same time, based on user behavior data as the basis for user intent recognition, there is no need to manually set rules, making intent recognition more accurate and closer to user needs. While actively performing intent recognition and service recommendations, it also effectively improves the accuracy of subsequent intent recognition.

[0124] According to the method described in the above embodiment, this embodiment will be further described from the perspective of a training device for an intention recognition model. The training device for the intention recognition model can be implemented as an independent entity or integrated into an electronic device, such as a terminal. The terminal may include a mobile phone, a tablet computer, etc.

[0125] See also Fig.11 , Fig.11 is a structural diagram of a training device for an intent recognition model provided in an embodiment of the present application, such as Fig.11 As shown, the training device 1100 of the intention recognition model provided in the embodiment of the present application includes:

[0126] The data acquisition module 1101 is used to acquire historical behavior data, wherein the historical behavior data includes interaction data, user attributes and application attributes, and the interaction data is generated by the target user operating the target terminal including the target application, the target user corresponds to the user attribute, and the target application corresponds to the application attribute;

[0127] The first processing module 1102 is used to filter the interaction data to obtain processed interaction data, and extract features from the processed interaction data to obtain user behavior representation;

[0128] The second processing module 1103 is used to extract features from user attributes, application attributes and terminal attributes of the first terminal, and combine the obtained feature representation with the user behavior representation to obtain a historical behavior table;

[0129] The model training module 1104 is used to train the intent recognition model according to the historical behavior representation to obtain a trained intent recognition model.

[0130] In one embodiment, the first processing module 1102 is further configured to:

[0131] Performing behavior recognition on the interaction data, and obtaining a behavior coding matrix of the interaction data based on the recognition results;

[0132] Perform behavior position analysis based on the behavior coding matrix to obtain the position coding matrix of the interaction data;

[0133] Add the position encoding matrix and the behavior encoding matrix to obtain the behavior interaction matrix of the interaction data;

[0134] The behavior matrix is ​​filtered and feature extracted to obtain the user behavior representation corresponding to the interaction data.

[0135] In one embodiment, the first processing module 1102 is further configured to:

[0136] Performing behavior recognition on the interaction data, and obtaining behavior information of the interaction data according to the recognition result;

[0137] Screening and processing the initial behavior information to obtain processed behavior information;

[0138] Based on the time series, it is constructed according to the behavioral information to obtain a behavioral coding matrix using behavioral data.

[0139] In one embodiment, the first processing module 1102 is further configured to:

[0140] Position analysis, to obtain the position encoding matrix of the interaction data, including:

[0141] Determine the timestamp of each behavior recorded in the behavior coding matrix according to the time information of the interaction data;

[0142] According to the timestamp, a position encoding matrix of the interaction data is constructed.

[0143] In one embodiment, the first processing module 1102 is further configured to:

[0144] Perform feature extraction on the behavior matrix in the frequency domain to obtain feature representation in the frequency domain;

[0145] The feature representation is denoised using the gradient descent method to obtain the denoised feature representation;

[0146] The denoised feature representation is transformed into the time domain to obtain the user behavior representation in the time domain.

[0147] In one embodiment, the model training module 1104 is further used to:

[0148] Analyze and collect statistics on the interaction data, and adjust the positive sampling probability of each application in the target application according to the set probability adjustment method to obtain the positive sampling probability of the target application;

[0149] The intent recognition model is trained according to the positive sampling probability and the historical behavior representation, and a trained intent recognition model is obtained when the training is completed.

[0150] In one embodiment, the training device 1100 for the intention recognition model further includes an intention recognition module for:

[0151] When performing intent recognition, behavior data within a preset time period is obtained based on the current time;

[0152] Extract features from the behavior data to obtain behavior features corresponding to the behavior data, and input the behavior features into the trained intent recognition model to obtain several recognition results and the probability value of each recognition result;

[0153] The recommended service is determined in the recognition result according to the probability value, and the corresponding service card is generated according to the recommended service for display.

[0154] During specific implementation, the above modules and / or units can be implemented as independent entities, or can be arbitrarily combined to be implemented as the same or several entities. The specific implementation of the above modules and / or units can refer to the previous method embodiments. The specific beneficial effects that can be achieved can also refer to the beneficial effects in the previous method embodiments, which will not be repeated here.

[0155] Also, see Fig.12 , Fig.12 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, and the electronic device may be a mobile terminal such as a smart phone, a tablet computer, or the like. Fig.12 As shown, the electronic device 1200 includes a processor 1201 and a memory 1202. The processor 1201 is electrically connected to the memory 1202.

[0156] Processor 1201 is the control center of electronic device 1200. It uses various interfaces and lines to connect various parts of the entire electronic device. It executes various functions of electronic device 1200 and processes data by running or loading applications stored in memory 1202 and calling data stored in memory 1202, thereby monitoring electronic device 1200 as a whole.

[0157] In this embodiment, the processor 1201 in the electronic device 1200 will load the instructions corresponding to the processes of one or more applications into the memory 1202 according to the following steps, and the processor 1201 will run the application stored in the memory 1202, thereby implementing any step in the training method of the intent recognition model provided in the above embodiment.

[0158] The electronic device 1200 can implement the steps in any embodiment of the training method of the intent recognition model provided in the embodiments of the present application. Therefore, it can achieve the beneficial effects that can be achieved by any training method of the intent recognition model provided in the embodiments of the present application. Please refer to the previous embodiments for details and will not be repeated here.

[0159] See also Fig.13 , Fig.13 is another structural schematic diagram of an electronic device provided in an embodiment of the present application, such as Fig.13 As shown, Fig.13 The specific structural block diagram of the electronic device provided in the embodiment of the present application is shown, and the electronic device can be used to implement the training method of the intention recognition model provided in the above embodiment. The electronic device 1300 can be a mobile terminal such as a smart phone or a laptop computer.

[0160] The RF circuit 1310 is used to receive and send electromagnetic waves, realize the mutual conversion between electromagnetic waves and electrical signals, and thus communicate with a communication network or other devices. The RF circuit 1310 may include various existing circuit elements for performing these functions, such as antennas, radio frequency transceivers, digital signal processors, encryption / decryption chips, user identity module (SIM) cards, memories, etc. The RF circuit 1310 can communicate with various networks such as the Internet, corporate intranets, wireless networks, or communicate with other devices through wireless networks. The above-mentioned wireless networks may include cellular telephone networks, wireless local area networks, or metropolitan area networks. The above-mentioned wireless networks can use various communication standards, protocols and technologies, including but not limited to Global System for Mobile Communication (GSM), Enhanced Data GSM Environment (EDGE), Wideband Code Division Multiple Access (WCDMA), Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Wireless Fidelity (Wi-Fi) (such as Institute of Electrical and Electronics Engineers standards IEEE 802.11a, IEEE 802.11b, IEEE802.11g and / or IEEE802.11n), Voice over Internet Protocol (VoIP), Worldwide Interoperability for Microwave Access (Wi-Max), other protocols for email, instant messaging and short messages, and any other suitable communication protocols, even those that have not yet been developed.

[0161] The memory 1320 can be used to store software programs and modules, such as program instructions / modules corresponding to the training method of the intent recognition model in the above-mentioned embodiment. The processor 1380 executes various functional applications and the training method of the intent recognition model by running the software programs and modules stored in the memory 1320.

[0162] The memory 1320 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 1320 may further include a memory remotely arranged relative to the processor 1380, and these remote memories may be connected to the electronic device 1300 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0163] The input unit 1330 can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control. Specifically, the input unit 1330 may include a touch-sensitive surface 1331 and other input devices 1332. The touch-sensitive surface 1331, also known as a touch display screen or touch pad, can collect user touch operations on or near it (such as operations performed by the user using any suitable object or accessory such as a finger, stylus, etc. on or near the touch-sensitive surface 1331), and drive the corresponding connection device according to a pre-set program. Optionally, the touch-sensitive surface 1331 may include a touch detection device and a touch controller. Among them, the touch detection device detects the user's touch direction, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into touch point coordinates, and then sends it to the processor 1380, and can receive and execute commands sent by the processor 1380. In addition, the touch-sensitive surface 1331 may be implemented in various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the touch-sensitive surface 1331, the input unit 1330 may further include other input devices 1332. Specifically, the other input devices 1332 may include, but are not limited to, one or more of a physical keyboard, a function key (such as a volume control key, a switch key, etc.), a trackball, a mouse, a joystick, and the like.

[0164] The display unit 1340 can be used to display information input by the user or information provided to the user and various graphical user interfaces of the electronic device 1300, which can be composed of graphics, text, icons, videos and any combination thereof. The display unit 1340 may include a display panel 1341, and optionally, the display panel 1341 may be configured in the form of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), etc. Further, the touch-sensitive surface 1331 may cover the display panel 1341, and when the touch-sensitive surface 1331 detects a touch operation on or near it, it is transmitted to the processor 1380 to determine the type of the touch event, and then the processor 1380 provides corresponding visual output on the display panel 1341 according to the type of the touch event. Although in the figure, the touch-sensitive surface 1331 and the display panel 1341 are implemented as two independent components to implement input and output functions, in some embodiments, the touch-sensitive surface 1331 and the display panel 1341 can be integrated to implement input and output functions.

[0165] The electronic device 1300 may also include at least one sensor 1350, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor, wherein the ambient light sensor may adjust the brightness of the display panel 1341 according to the brightness of the ambient light, and the proximity sensor may generate an interrupt when the flip cover is closed or closed. As a type of motion sensor, the gravity acceleration sensor can detect the magnitude of acceleration in each direction (generally three axes), and can detect the magnitude and direction of gravity when stationary, which can be used for applications that identify the posture of the mobile phone (such as horizontal and vertical screen switching, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, tapping), etc.; as for other sensors such as gyroscopes, barometers, hygrometers, thermometers, infrared sensors, etc. that can also be configured in the electronic device 1300, they will not be repeated here.

[0166] The audio circuit 1360, the speaker 1361, and the microphone 1362 can provide an audio interface between the user and the electronic device 1300. The audio circuit 1360 can transmit the received audio data to the speaker 1361 after converting the received audio data into an electrical signal, which is converted into a sound signal for output by the speaker 1361; on the other hand, the microphone 1362 converts the collected sound signal into an electrical signal, which is received by the audio circuit 1360 and converted into audio data, and then the audio data is processed by the output processor 1380 and sent to, for example, another terminal through the RF circuit 1310, or the audio data is output to the memory 1320 for further processing. The audio circuit 1360 may also include an earplug jack to provide communication between an external headset and the electronic device 1300.

[0167] The electronic device 1300 can help the user receive requests, send information, etc. through the transmission module 1370 (such as a Wi-Fi module), which provides the user with wireless broadband Internet access. Although the transmission module 1370 is shown in the figure, it is understandable that it is not a necessary component of the electronic device 1300 and can be omitted as needed without changing the essence of the invention.

[0168] The processor 1380 is the control center of the electronic device 1300. It uses various interfaces and lines to connect various parts of the entire mobile phone. By running or executing software programs and / or modules stored in the memory 1320, and calling data stored in the memory 1320, it executes various functions of the electronic device 1300 and processes data, thereby monitoring the electronic device as a whole. Optionally, the processor 1380 may include one or more processing cores; in some embodiments, the processor 1380 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, etc., and the modem processor mainly processes wireless communications. It can be understood that the above-mentioned modem processor may not be integrated into the processor 1380.

[0169] The electronic device 1300 also includes a power supply 1390 (such as a battery) for supplying power to various components. In some embodiments, the power supply can be logically connected to the processor 1380 through a power management system, so that the power management system can manage charging, discharging, and power consumption management. The power supply 1390 can also include any components such as one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, and power status indicators.

[0170] Although not shown, the electronic device 1300 also includes a camera (such as a front camera, a rear camera), a Bluetooth module, etc., which will not be described in detail here. Specifically in this embodiment, the display unit of the electronic device is a touch screen display, and the mobile terminal also includes a memory, and one or more programs, wherein one or more programs are stored in the memory, and are configured to be executed by one or more processors to implement any step in the training method of the intent recognition model provided in the above embodiment.

[0171] In specific implementation, the above modules can be implemented as independent entities, or can be arbitrarily combined and implemented as the same or several entities. The specific implementation of the above modules can be found in the previous method embodiments, which will not be repeated here.

[0172] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by controlling related hardware through instructions, and the instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. To this end, an embodiment of the present application provides a storage medium in which a plurality of instructions are stored, and when the instructions are executed by the processor, any step in the training method of the intent recognition model provided in the above embodiments can be implemented.

[0173] The storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0174] Since the instructions stored in the storage medium can execute the steps in any embodiment of the training method for the intent recognition model provided in the embodiments of the present application, the beneficial effects that can be achieved by any training method for the intent recognition model provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0175] The above is a detailed introduction to the training method, device, electronic device and storage medium of an intent recognition model provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and its core idea of ​​the present application. At the same time, for those skilled in the art, according to the ideas of the present application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present application. Moreover, for those of ordinary skill in the art, without departing from the principles of the present application, several improvements and modifications can be made, and these improvements and modifications are also regarded as the scope of protection of the present application.

Claims

1. A training method for an intent recognition model, characterized in that: include: Acquire historical behavior data, wherein the historical behavior data includes interaction data, user attributes, and application attributes, and the interaction data is generated by a target user operating a target terminal including a target application, the target user corresponds to the user attributes, and the target application corresponds to the application attributes; Performing filtering processing on the interaction data to obtain processed interaction data, and performing feature extraction on the processed interaction data to obtain user behavior representation; Extracting features from the user attributes, the application attributes, and the terminal attributes of the target terminal, and combining the obtained feature representation with the user behavior representation to obtain a historical behavior representation; The intent recognition model is trained according to the historical behavior representation to obtain a trained intent recognition model.

2. The method according to claim 1, characterized in that The filtering the interaction data to obtain processed interaction data, and extracting features from the processed interaction data to obtain user behavior representation, includes: Performing behavior recognition on the interaction data, and obtaining a behavior coding matrix of the interaction data according to the recognition result; Performing behavior position analysis according to the behavior coding matrix to obtain a position coding matrix of the interaction data; Performing matrix addition on the position coding matrix and the behavior coding matrix to obtain a behavior interaction matrix of the interaction data; The behavior matrix is ​​filtered and feature extracted to obtain a user behavior representation corresponding to the interaction data.

3. The method according to claim 2, characterized in that The performing behavior recognition on the interaction data and obtaining a behavior coding matrix of the interaction data according to the recognition result includes: Performing behavior recognition on the interaction data, and obtaining behavior information of the interaction data according to the recognition result; Screening the initial behavior information to obtain processed behavior information; The behavior information is constructed based on the time sequence to obtain the behavior coding matrix using the behavior data.

4. The method according to claim 2, characterized in that The performing behavior position analysis according to the behavior coding matrix to obtain the position coding matrix of the interaction data includes: Determining the timestamp of each behavior recorded in the behavior coding matrix according to the time information of the interaction data; A position coding matrix of the interaction data is constructed according to the timestamp.

5. The method according to claim 2, characterized in that The filtering and feature extraction processing of the behavior matrix to obtain the user behavior representation corresponding to the interaction data includes: Performing feature extraction on the behavior matrix in the frequency domain to obtain feature representation in the frequency domain; De-noising the feature representation using a gradient descent method to obtain a de-noised feature representation; The denoised feature representation is transformed into the time domain to obtain the user behavior representation in the time domain.

6. The method according to claim 1, characterized in that The training of the intent recognition model according to the historical behavior representation and obtaining the trained intent recognition model when the training is completed include: Analyzing and counting the interaction data, and adjusting the positive sampling probability of each application in the target application according to the set probability adjustment method, to obtain the positive sampling probability of the target application; The intent recognition model is trained according to the positive sampling probability and the historical behavior representation, and a trained intent recognition model is obtained when the training is completed.

7. The method according to claim 1, characterized in that The method further comprises: When performing intent recognition, behavior data within a preset time period is obtained based on the current time; Extracting features from the behavior data to obtain behavior features corresponding to the behavior data, and inputting the behavior features into a trained intention recognition model to obtain a number of recognition results and a probability value for each recognition result; A recommended service is determined in the recognition result according to the probability value, and a corresponding service card is generated according to the recommended service for display.

8. A training device for an intention recognition model, characterized in that: include: A data acquisition module, used to acquire historical behavior data, wherein the historical behavior data includes interaction data, user attributes and application attributes, and the interaction data is generated by a target user operating a target terminal including a target application, the target user corresponds to the user attributes, and the target application corresponds to the application attributes; A first processing module is used to filter the interaction data to obtain processed interaction data, and extract features from the processed interaction data to obtain user behavior representation; A second processing module is used to extract features from the user attributes, the application attributes and the terminal attributes of the target terminal, and to combine the obtained feature representation with the user behavior representation to obtain a historical behavior table; The model training module is used to train the intent recognition model according to the historical behavior representation to obtain a trained intent recognition model.

9. An electronic device, characterized in that: The electronic device comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, and the processor implements the steps in the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 7 are implemented.

Images