Information processing method based on heterogeneous data fusion and multi-task learning

Abstract

The present application provides an information processing method based on heterogeneous data fusion and multi-task learning, the method comprising: inputting the current data of a business object into a target model; determining the first feature vector of numerical data, the second feature vector of categorical data, and the third feature vector of time series data by a preprocessing module; obtaining the first feature selection mask corresponding to the first feature vector, the second feature selection mask corresponding to the second feature vector, and the third feature selection mask corresponding to the third feature vector by the preprocessing module; obtaining a shared feature vector by a data fusion module; obtaining the task parameters of each task in a plurality of tasks, and determining the target task feature vector of each task by a task feature extraction module corresponding to each task according to the task parameters of each task and the shared feature vector; obtaining the prediction result of the business object under each task by a prediction module corresponding to each task feature extraction module, so that the prediction result is more comprehensive and accurate.

Description

Technical Field

[0001] The present application relates to the technical field of big data analysis and processing, and specifically, to an information processing method based on heterogeneous data fusion and multi-task learning. Background Art

[0002] In various business fields, there is a need to predict information about business objects.

[0003] At present, when predicting information of business objects, neural network models can be used for prediction. However, currently, neural network models are generally used to predict business objects in a single dimension, resulting in certain limitations in the prediction results. Summary of the invention

[0004] The purpose of this application is to provide an information processing method based on heterogeneous data fusion and multi-task learning to address the deficiencies in the above-mentioned prior art, so as to improve the accuracy of information processing based on heterogeneous data fusion and multi-task learning.

[0005] To achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:

[0006] In a first aspect, an embodiment of the present application provides an information processing method based on heterogeneous data fusion and multi-task learning, the method comprising:

[0007] Inputting current data of the business object into the target model, the current data includes: numerical data, categorical data and time series data, the target model includes: a preprocessing module, a data fusion module, a plurality of task feature extraction modules connected to the data fusion module, and a prediction module corresponding to each task feature extraction module, wherein each task feature extraction module is used to extract features under a task type;

[0008] The preprocessing module determines a first eigenvector of numerical data, a second eigenvector of categorical data, and a third eigenvector of time series data;

[0009] The preprocessing module performs mask processing on the first feature vector, the second feature vector, and the third feature vector, respectively, to obtain a first feature selection mask corresponding to the first feature vector, a second feature selection mask corresponding to the second feature vector, and a third feature selection mask corresponding to the third feature vector;

[0010] The data fusion module performs fusion processing according to the first feature vector, the first feature selection mask, the second feature vector, the second feature selection mask, the third feature vector and the third feature selection mask to obtain a shared feature vector;

[0011] Obtaining task parameters of each task in the plurality of tasks, and determining a target task feature vector of each task by a task feature extraction module corresponding to each task according to the task parameters of each task and the shared feature vector;

[0012] The prediction module corresponding to each task feature extraction module performs prediction processing according to the target task feature vector of each task to obtain the prediction result of the business object under each task.

[0013] Optionally, the determining, by the preprocessing module, the first eigenvector of the numerical data, the second eigenvector of the categorical data, and the third eigenvector of the time series data includes:

[0014] The numerical data is directly input into the feature extraction submodule of the preprocessing module to extract features, and obtain a first feature vector;

[0015] Inputting the categorical data into an embedding submodule and a feature extraction submodule in a preprocessing module in sequence to obtain the second feature vector;

[0016] The time series data is sequentially input into the first sub-linear layer, the multi-head attention sub-module, the feedforward neural sub-module, the average pooling sub-module, and the second sub-linear layer in the preprocessing module to obtain the third eigenvector.

[0017] Optionally, the step of sequentially inputting the categorical data into an embedding submodule and a feature extraction submodule in a preprocessing module to obtain the second feature vector comprises:

[0018] The embedding submodule embeds each subcategory type field into sub-embedded data corresponding to each subcategory type according to the embedding matrix corresponding to each subcategory type, and obtains embedded data corresponding to the category type data according to each sub-embedded data;

[0019] The embedded data corresponding to the categorical data is input into a feature extraction submodule, and the feature extraction submodule performs feature extraction on the embedded data according to extraction parameters corresponding to the categorical data to obtain the second feature vector.

[0020] Optionally, the step of sequentially inputting the time series data into a first linear layer, a multi-head attention submodule, a feedforward neural submodule, an average pooling submodule, and a second linear layer in a preprocessing module to obtain the third feature vector includes:

[0021] The first linear layer performs dimension conversion and position encoding on the time series data to obtain first time series data;

[0022] Input the first time series data into the multi-head attention submodule to extract time features, obtain the second time series data output by each attention head, and concatenate and linearly transform each of the second time series data to obtain a multi-head attention feature vector;

[0023] Inputting the multi-head attention feature vector into the feedforward neural submodule to extract type features to obtain a type feature vector;

[0024] Inputting the type feature vector into the average pooling submodule to aggregate and pool the time features in the type feature vector to obtain pooled features;

[0025] The pooled features are input into the second linear layer for projection processing to obtain the third feature vector.

[0026] Optionally, the preprocessing module performs mask processing on the first feature vector, the second feature vector, and the third feature vector respectively to obtain a first feature selection mask corresponding to the first feature vector, a second feature selection mask corresponding to the second feature vector, and a third feature selection mask corresponding to the third feature vector, including:

[0027] Obtaining a first feature selection mask corresponding to the first feature vector according to the first feature vector, a weight parameter of a masked fully connected layer corresponding to the numerical data, and a mask bias item of the masked fully connected layer corresponding to the numerical data;

[0028] Obtaining a second feature selection mask corresponding to the second feature vector according to the second feature vector, a weight parameter of the masked fully connected layer corresponding to the categorical data, and a mask bias item of the masked fully connected layer corresponding to the categorical data;

[0029] A third feature selection mask corresponding to the third feature vector is obtained according to the third feature vector, a weight parameter of the masked fully connected layer corresponding to the time series data, and a mask bias item of the masked fully connected layer corresponding to the time series data.

[0030] Optionally, the data fusion module performs fusion processing according to the first feature vector, the first feature selection mask, the second feature vector, the second feature selection mask, the third feature vector and the third feature selection mask to obtain a shared feature vector, including:

[0031] Determining a first weighted feature according to the first feature vector and the first feature selection mask;

[0032] Determining a second weighted feature according to the second feature vector and the second feature selection mask;

[0033] Determining a third weighted feature according to the third feature vector and the third feature selection mask;

[0034] The first weighted feature, the second weighted feature and the third weighted feature are concatenated to obtain a shared feature vector.

[0035] Optionally, determining a first weighted feature according to the first feature vector and the first feature selection mask includes:

[0036] The first feature selection mask is multiplied element-by-element by the first feature vector to obtain the first weighted feature.

[0037] Optionally, the task feature extraction module corresponding to each task determines the target task feature vector of each task according to the task parameters of each task and the shared feature vector, including:

[0038] Extracting features from the shared feature vector according to task parameters of each task to obtain a task feature vector corresponding to each task, wherein the task parameters include: a weight matrix and a bias item;

[0039] The task feature vectors are cross-task fused to obtain the target task feature vectors of each task.

[0040] Optionally, performing cross-task fusion processing on the task feature vectors to obtain a target task feature vector of each task includes:

[0041] Obtain a cross weight matrix, wherein the parameter value in the i-th row and j-th column in the cross weight matrix is ​​used to represent the weight of the information obtained by the task feature vector of the i-th task from the task feature vector of the j-th task, and i and j are integers greater than 0 respectively;

[0042] For task k, obtain the parameter values ​​of the kth row in the cross weight matrix, where k is an integer greater than 0;

[0043] Calculate the product of each parameter value in the kth row and the target task feature vector of each associated task corresponding to each parameter value;

[0044] The sum of the products of each parameter value in the kth row and each associated task corresponding to each parameter value is taken as the target task feature vector of task k, where if the parameter value is in the mth column, the associated task corresponding to the parameter value is task m, and m is an integer greater than 0.

[0045] In a second aspect, an embodiment of the present application further provides an electronic device, comprising: a processor, a storage medium and a bus, wherein the storage medium stores program instructions executable by the processor, and when an application is running, the processor communicates with the storage medium via the bus, and the processor executes the program instructions to perform the steps of the information processing method based on heterogeneous data fusion and multi-task learning described in the first aspect above.

[0046] In a third aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and the computer program is read and executes the steps of the information processing method based on heterogeneous data fusion and multi-task learning described in the first aspect above.

[0047] The beneficial effects of this application are:

[0048] The present application provides an information processing method based on heterogeneous data fusion and multi-task learning, wherein the preprocessing module in the target model preprocesses the heterogeneous data sources of the business object first, obtains the feature vectors of each heterogeneous data source, and performs mask processing on the obtained feature vectors to obtain the feature selection mask corresponding to the feature vectors of each heterogeneous data source, and then the data fusion module performs fusion processing according to the feature vectors of each heterogeneous data source and the feature selection mask corresponding to the feature vectors of each heterogeneous data source to obtain the shared feature vectors of all heterogeneous data sources, so that the features of each heterogeneous data source can be effectively fused, and the obtained shared feature vector can contain the important features of each heterogeneous data source, and then the task feature extraction module determines the target task vector of each task according to the task parameters of each task and the shared feature vector obtained by fusion, and the prediction module corresponding to each task feature extraction module performs prediction processing according to the target task feature vector of each task, so as to obtain the prediction results of the business object under each task. The shared feature vector can be optimized in a targeted manner according to the different characteristics of the task, and the target task vector of each task can be obtained, and the corresponding prediction module can be selected according to different tasks to predict the target task vector of each task, so that the prediction results of the business object are more comprehensive and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 A schematic diagram of the architecture of a target model provided in an embodiment of the present application;

[0051] Figure 2 A flowchart of an information processing method based on heterogeneous data fusion and multi-task learning provided in an embodiment of the present application;

[0052] Figure 3 A flowchart of a method for determining a feature vector provided in an embodiment of the present application;

[0053] Figure 4 A schematic diagram of the structure of a feature extraction submodule provided in an embodiment of the present application;

[0054] Figure 5 A flowchart of another method for determining a feature vector provided in an embodiment of the present application;

[0055] Figure 6 A flowchart of another method for determining a feature vector provided in an embodiment of the present application;

[0056] Figure 7 A flowchart of a method for determining a feature selection mask provided in an embodiment of the present application;

[0057] Figure 8 A flowchart of a method for determining a shared feature vector provided in an embodiment of the present application;

[0058] Fig. 9 A complete flow chart of obtaining a shared feature vector in an embodiment of the present application;

[0059] Fig.10 A flowchart of a method for determining a target task feature vector provided in an embodiment of the present application;

[0060] Fig.11 A flowchart of another method for determining a target task feature vector provided in an embodiment of the present application;

[0061] Fig.12 A structural block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0062] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart under the guidance of the content of the present application, or remove one or more operations from the flowchart.

[0063] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0064] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the existence of the features declared thereafter, but does not exclude the addition of other features.

[0065] With the growing demand for skilled workers among global companies and the rapid transformation of industries, how to comprehensively analyze and predict the multi-dimensional status of employees has become a key issue in corporate management.

[0066] In the existing technology, employee evaluation methods usually focus on a single dimension, such as skills or work performance, and it is difficult to comprehensively process heterogeneous data such as employee education, ability level, and job mobility, resulting in limited and incomplete analysis results for employees. For example, analysis of traffic data and analysis of animal growth and development usually focus on a single dimension, and it is difficult to comprehensively process multi-source heterogeneous data, resulting in limitations in the analysis results.

[0067] This application proposes an information processing method based on heterogeneous data fusion and multi-task learning. By fusing the multi-dimensional heterogeneous data of a business object, a target model is pre-established to enable prediction of the business object under different tasks.

[0068] If the business object is an employee, the learning improvement needs, ability level assessment and turnover probability of the business object can be predicted respectively by integrating the multi-dimensional heterogeneous data of the business object.

[0069] If the business object is a different animal, the growth requirements, growth level assessment and probability of growth problems of the business object can be analyzed by integrating the multi-dimensional heterogeneous data of the business object.

[0070] Optionally, the information processing method based on heterogeneous data fusion and multi-task learning provided in the embodiment of the present application is applied to an electronic device, which may be a terminal device with computing and display capabilities, such as a mobile phone, a tablet computer, a laptop computer, a PDA, a desktop computer, or a server. Specifically, it can be applied to an application in a terminal device, such as an APP (application) of a mobile phone, an application system on a computer, etc.

[0071] The specific implementation process of information processing based on heterogeneous data fusion and multi-task learning provided in the embodiments of the present application is explained in detail below.

[0072] Figure 1 A schematic diagram of the architecture of a target model provided in an embodiment of the present application, such as Figure 1 As shown, the target model may include: a preprocessing module, a data fusion module, a multi-task feature extraction module connected to the data fusion module, and a prediction module corresponding to each task feature extraction module one by one.

[0073] The preprocessing module may preprocess the multi-dimensional heterogeneous data sources of the business objects so that the dimensions of the heterogeneous data sources are the same.

[0074] The data fusion module can fuse the heterogeneous data of the same dimension output by the preprocessing module to obtain a shared feature vector.

[0075] The multi-task feature extraction module can extract task features from shared feature vectors according to different tasks to obtain target task feature vectors under different tasks.

[0076] The prediction module can predict the target task feature vector output by the task feature extraction module corresponding to the prediction module to obtain the prediction results under each task.

[0077] Figure 2 The present invention provides a flowchart of an information processing method based on heterogeneous data fusion and multi-task learning, wherein the execution subject of the method is the aforementioned electronic device. Figure 2 As shown, the method includes:

[0078] S101. Input the current data of the business object into the target model.

[0079] The current data may include: numerical data, categorical data, and time series data. The business object may include multiple data. If the business object is an employee, the numerical data may include, for example, the working years and age of each business object; the categorical data may include, for example, the study and job title data of each business object; and the time series data may include, for example, the work mobility data and performance change trend data of each business object.

[0080] If the business object is an animal, the numerical data may be, for example, the age data of each business object; the categorical data may be, for example, data such as the type of each business object; and the time series data may be, for example, data such as the growth trend of each business object.

[0081] Optionally, numeric data The dimension is , where n is the number of business objects, The number of numeric features of the business object. The dimension is ,in is the number of business objects, is the number of categorical features. Time series data The dimension is , where n is the number of business objects, is the number of time series features, and t is time. The number of business objects in the current data of processing a batch of business objects, the number of business objects in the numerical data, the categorical data, and the time series data are the same, that is, the value of n in the current data of processing a batch of business objects, the value of n in the categorical data, and the value of n in the time series data are the same.

[0082] Optionally, the target model may include: a preprocessing module, a data fusion module, a multi-task feature extraction module connected to the data fusion module, and a prediction module corresponding to each task feature extraction module. Each task feature extraction module is used to extract features of a task type.

[0083] S102 , determining, by a preprocessing module, a first eigenvector of numerical data, a second eigenvector of categorical data, and a third eigenvector of time series data.

[0084] The dimensions of the first eigenvector, the second eigenvector, and the third eigenvector are all the same. For example, the dimensions of the first eigenvector, the second eigenvector, and the third eigenvector are all The preprocessing module is used to process multi-dimensional heterogeneous data, that is, numerical data, categorical data, and time series data, into data of the same dimension.

[0085] S103, the preprocessing module performs mask processing on the first eigenvector, the second eigenvector and the third eigenvector respectively to obtain a first feature selection mask corresponding to the first eigenvector, a second feature selection mask corresponding to the second eigenvector and a third feature selection mask corresponding to the third eigenvector.

[0086] Specifically, the preprocessing module can perform mask processing on the first feature vector to obtain a first feature selection mask corresponding to the first feature vector. ; The preprocessing module performs mask processing on the second feature vector to obtain the second feature selection mask corresponding to the second feature vector ; The preprocessing module performs mask processing on the third feature vector to obtain the third feature selection mask corresponding to the third feature vector .

[0087] Optionally, by performing masking processing on the first eigenvector, the second eigenvector and the third eigenvector respectively, the obtained first feature selection mask, the second feature selection mask and the third feature selection mask can be used to facilitate the subsequent feature selection of the first eigenvector, the second eigenvector and the third eigenvector, thereby facilitating the accuracy of the features extracted in the subsequent task feature extraction, and subsequently making the prediction results more accurate.

[0088] S104: The data fusion module performs fusion processing according to the first feature vector, the first feature selection mask, the second feature vector, the second feature selection mask, the third feature vector and the third feature selection mask to obtain a shared feature vector.

[0089] The shared feature vector is a common feature vector among the first feature vector, the second feature vector and the third feature vector. The shared feature vector can be used, for example, To express.

[0090] S105 , obtaining task parameters of each task among the multiple tasks, and having the task feature extraction module corresponding to each task determine a target task feature vector of each task according to the task parameters of each task and the shared feature vector.

[0091] Optionally, each task may correspond to a task parameter, and the task parameter may be used for task feature extraction, so that the target task feature vector obtained by the task feature extraction module for each task is different.

[0092] For example, if there are three tasks, such as task 1, task 2 and task 3, task 1 corresponds to a task feature extraction module 1, task 2 corresponds to a task feature extraction module 2 and task 3 corresponds to a task feature extraction module 3. Then, task feature extraction module 1 determines the target task feature vector of task 1 according to the task parameters and shared feature vector of task 1; task feature extraction module 2 determines the target task feature vector of task 2 according to the task parameters and shared feature vector of task 2; and task feature extraction module 3 determines the target task feature vector of task 3 according to the task parameters and shared feature vector of task 3.

[0093] S106. The prediction module corresponding to each task feature extraction module performs prediction processing according to the target task feature vector of each task to obtain the prediction result of the business object under each task.

[0094] Optionally, each task feature extraction module may correspond to a prediction module, and then the prediction module corresponding to each task feature extraction module performs prediction processing on the target task feature vector output by each task feature extraction module to obtain a prediction result for each task.

[0095] The prediction module may be, for example, a binary prediction module, a ternary prediction module, and a regression prediction module, and different prediction modules may output different prediction results. For different tasks, the prediction module corresponding to the task may be used to predict the target task feature vector under the task.

[0096] Exemplarily, the prediction module may include prediction module 1, prediction module 2 and prediction module 3, wherein prediction module 1 corresponds to task feature extraction module 1, prediction module 2 corresponds to task feature extraction module 2, and prediction module 3 corresponds to task feature extraction module 3. Then, prediction module 1 can perform prediction processing according to the target task feature vector of task 1 to obtain the prediction result of the business object under task 1; prediction module 2 can perform prediction processing according to the target task feature vector of task 2 to obtain the prediction result of the business object under task 2; prediction module 3 can perform prediction processing according to the target task feature vector of task 3 to obtain the prediction result of the business object under task 3.

[0097] In this embodiment, the preprocessing module in the target model first preprocesses the heterogeneous data sources of the business object to obtain the feature vectors of each heterogeneous data source, and performs mask processing on the obtained feature vectors to obtain the feature selection mask corresponding to the feature vectors of each heterogeneous data source. Then, the data fusion module performs fusion processing according to the feature vectors of each heterogeneous data source and the feature selection mask corresponding to the feature vectors of each heterogeneous data source to obtain the shared feature vectors of all heterogeneous data sources, which can effectively fuse the features of each heterogeneous data source. The obtained shared feature vector can contain important features of each heterogeneous data source. Then, the task feature extraction module determines the target task vector of each task according to the task parameters of each task and the fused shared feature vector, and the prediction module corresponding to each task feature extraction module performs prediction processing according to the target task feature vector of each task, thereby obtaining the prediction results of the business object under each task. The shared feature vector can be optimized in a targeted manner according to the different characteristics of the task to obtain the target task vector of each task, and the corresponding prediction module is selected according to different tasks to predict the target task vector of each task, so that the prediction results of the business object are more comprehensive and accurate.

[0098] Figure 3 A flow chart of a method for determining a feature vector provided in an embodiment of the present application is shown as follows: Figure 3 As shown, the determination of the first feature vector of the numerical data, the second feature vector of the categorical data, and the third feature vector of the time series data by the preprocessing module in S102 may include:

[0099] S201 , directly inputting numerical data into a feature extraction submodule of a preprocessing module to perform feature extraction, and obtaining a first feature vector.

[0100] Optionally, the numerical data can be directly input into the feature extraction submodule of the preprocessing module, and the feature submodule performs feature extraction on the numerical data according to the extraction parameters corresponding to the numerical data to obtain a first feature vector. The extraction parameters corresponding to the numerical data may include the weight matrix corresponding to the numerical data in the first fully connected layer, the bias item corresponding to the numerical data in the first fully connected layer, the weight matrix corresponding to the numerical data in the second fully connected layer, and the bias item corresponding to the numerical data in the second fully connected layer.

[0101] Specifically, the dimension is Numeric data Input to the feature extraction submodule in the preprocessing module, where the structure diagram of the feature extraction submodule is as follows Figure 4 As shown in Figure 1, it includes two fully connected layers. You can first convert the numerical data Input to the first fully connected layer, after ReLU activation of the first fully connected layer, the generated dimension is Features The specific process is ,in It is the weight matrix corresponding to the numerical data in the first fully connected layer. is the bias term corresponding to the numerical data in the first fully connected layer. Input to the second fully connected layer, after ReLU activation of the second fully connected layer, the calculation process is: ,in It is the weight matrix corresponding to the numerical data in the second fully connected layer. is the bias term corresponding to the numerical data in the second fully connected layer, and the generated dimension is The first eigenvector .

[0102] S202: input the categorical data into the embedding submodule and the feature extraction submodule in the preprocessing module in sequence to obtain a second feature vector.

[0103] Specifically, the categorical data may be first input into the embedding submodule in the preprocessing module to perform embedding processing on the embedded data, and then the embedded data may be input into the feature extraction module to perform feature extraction, thereby obtaining a second feature vector of the categorical data.

[0104] S203. Input the time series data into the first sub-linear layer, the multi-head attention sub-module, the feedforward neural sub-module, the average pooling sub-module, and the second sub-linear layer in the preprocessing module in sequence to obtain the third eigenvector.

[0105] Specifically, the time series data can be input into the first sub-linear layer, the multi-head attention sub-module, the feedforward neural sub-module, the average pooling sub-module, and the second sub-linear layer in the preprocessing in sequence, and the third eigenvector of the time series data can be obtained using a preset method.

[0106] In this embodiment, the preprocessing processes for different heterogeneous data sources are different, but the dimensions of the feature vectors of the heterogeneous data sources obtained through preprocessing are the same, thereby facilitating subsequent data fusion and prediction processing.

[0107] Figure 5 A flow chart of another method for determining a feature vector provided in an embodiment of the present application is shown as follows: Figure 5 As shown, in the above S202, the category data is sequentially input into the embedding submodule and the feature extraction submodule in the preprocessing module to obtain the second feature vector, which may include:

[0108] S301, the embedding submodule embeds each subcategory type field into sub-embedded data corresponding to each subcategory type according to the embedding matrix corresponding to each subcategory type, and obtains embedded data corresponding to the category type data according to each sub-embedded data.

[0109] Optionally, the categorical data includes at least one subcategorical type, for example, for a dimension Categorical data , The number of category features is also the number of subcategories. Each subcategory contains the subcategory fields that can be used. To represent, where j is the serial number of the sub-category type. For example, if =2, the number of sub-categories is 2. It is a subcategory type field in subcategory type 1, such as the subcategory field of the age subcategory of each business object. It is a subcategory type field in subcategory type 2, such as the subcategory field of the age subcategory of each business object.

[0110] Optionally, each sub-category type corresponds to an embedding matrix, specifically the embedding matrix is , where j is the serial number of the subclass type, is the number of sub-categories, is the embedding dimension corresponding to each sub-category type. Then, according to the embedding matrix corresponding to each sub-category type , each sub-category field The embedding is the sub-embedded data corresponding to each sub-category type. Specifically, each sub-embedded data is . The combination of each sub-embedded data is used as the embedded data corresponding to the category data, then the embedded data corresponding to the category data is .in, The sub-embedding data corresponding to each sub-category type.

[0111] S302: Input the embedded data corresponding to the categorical data into a feature extraction submodule, and the feature extraction module extracts features from the embedded data according to extraction parameters corresponding to the categorical data to obtain a second feature vector.

[0112] Among them, the extraction parameters corresponding to the categorical data may include the weight matrix corresponding to the categorical data in the fully connected layer and the bias item corresponding to the categorical data in the fully connected layer.

[0113] Optionally, only one fully connected layer is needed for feature extraction of data type data. Specifically, feature extraction can be performed according to the following formula: .in, is the weight matrix corresponding to the categorical data in the fully connected layer, The bias term corresponding to the categorical data in the fully connected layer. And the second eigenvector of the categorical data obtained The dimension is .

[0114] In this embodiment, the category data is first input into the embedding submodule for numerical conversion. Different embedding matrices are used for conversion for different subcategory types. This can make the sub-embedded data corresponding to each subcategory type more accurate, so that the embedded data corresponding to the category data obtained based on each sub-embedded data is more comprehensive and accurate.

[0115] Figure 6 A flowchart of another method for determining a feature vector provided in an embodiment of the present application is shown as follows: Figure 6 As shown, in the above S203, the time series data is sequentially input into the first sub-linear layer, the multi-head attention sub-module, the feedforward neural sub-module, the average pooling sub-module, and the second sub-linear layer in the preprocessing module to obtain the third feature vector, which may include:

[0116] S401. Perform dimension conversion and position encoding on time series data by a first linear layer to obtain first time series data.

[0117] Optionally, the input time series data can be First, it is embedded into a high-dimensional space through the first linear layer. The dimension from Projection to the internal dimensions of the Transformer Specifically, the projection formula is ,in is the weight matrix of the first linear layer, is the bias term of the first linear layer. The dimension of the embedded time series data becomes , then the dimension of the first time series data is .

[0118] Optionally, after the time series data is dimensionally transformed, an encoder may be used to positionally encode the transformed time series data. Specifically, position encoding may be performed using the following formulas (1) and (2).

[0119] Formula (I)

[0120] Formula (II)

[0121] Where i is the position obtained by encoding. Thus, the first time series data with position information is obtained , its dimension remains .

[0122] S402: Input the first time series data into the multi-head attention submodule for time feature extraction to obtain the second time series data output by each attention head, and concatenate and linearly transform each time series data to obtain a multi-head attention feature vector.

[0123] Specifically, we first use a learnable weight matrix The first time series data to be input They are projected into query Q, key K and value V respectively. The specific formula is: ,in , is the dimension of a single attention head. The dimensions of Q, K, and V generated by each attention head are .

[0124] The dot product between the query Q and the key K is used to calculate the attention weight between each time step, which is specifically calculated by the following formula (III).

[0125] Formula (III)

[0126] Optionally, after computing the attention weights, The function normalizes the calculated weights to obtain the dependency between each time step. Then, the output of each attention head is generated by weighted summation with the value matrix V, that is, the second time series data. The dimension of each attention head output is .

[0127] Optionally, the multi-head attention mechanism computes multiple attention heads in parallel to capture multiple time-step patterns. Specifically, you can set h Attention heads, the output of each attention head is , then concatenate the outputs of each attention head to obtain a concatenated feature vector, the dimension of which is And through a linear transformation, the output of each attention head is integrated to obtain the multi-head attention feature vector ,in , the dimension of the multi-head attention feature vector is .

[0128] S403. Input the multi-head attention feature vector into the feedforward neural submodule to extract type features and obtain a type feature vector.

[0129] Optionally, after the multi-head self-attention mechanism, the obtained multi-head attention feature vector can be input into the feedforward neural submodule for further processing. The feedforward neural submodule consists of two fully connected layers. The calculation formula of the first fully connected layer is ,in, is the multi-head attention feature vector, is the weight matrix of the first fully connected layer in the feedforward neural submodule, and , is the bias term of the first fully connected layer in the feedforward neural submodule, and the calculation formula of the second fully connected layer is ,in is the weight matrix of the second fully connected layer in the feedforward neural submodule, and , is the bias term of the second fully connected layer in the feedforward neural submodule. Then, the feature vector output by the second fully connected layer is used as the feature vector output by the feedforward neural submodule to obtain the type feature vector , and the dimension of the type feature vector remains .

[0130] S404: Input the type feature vector into the average pooling submodule to aggregate and pool the time features in the type feature vector to obtain pooled features.

[0131] Specifically, the average pooling submodule type feature vector is used The time dimension t in is aggregated, and the calculation process is , after the pooling operation, the pooled features obtained The feature dimension is .

[0132] S405: Input the pooled features into the second linear layer for projection processing to obtain a third feature vector.

[0133] Specifically, through the second linear layer Projected into a 64-dimensional feature space, the formula is ,in is the weight matrix of the second linear layer, is the bias term of the second linear layer. The third eigenvector obtained is The output dimension is .

[0134] Figure 7 A flowchart of a method for determining a feature selection mask provided in an embodiment of the present application is shown as follows: Figure 7As shown, in the above S103, the preprocessing module performs mask processing on the first feature vector, the second feature vector and the third feature vector respectively to obtain a first feature selection mask corresponding to the first feature vector, a second feature selection mask corresponding to the second feature vector and a third feature selection mask corresponding to the third feature vector, which may include:

[0135] S501. Obtain a first feature selection mask corresponding to the first feature vector according to the first feature vector, a mask weight parameter of the masked fully connected layer corresponding to the numerical data, and a mask bias item of the masked fully connected layer corresponding to the numerical data.

[0136] Specifically, the sparse attention mechanism generates a first feature selection mask for the first feature vector The specific calculation formula is: in is the mask weight parameter of the masked fully connected layer corresponding to the numerical data, with a dimension of , It is the mask bias item of the masked fully connected layer corresponding to the numerical data.

[0137] S502: Obtain a second feature selection mask corresponding to the second feature vector according to the second feature vector, a weight parameter of the masked fully connected layer corresponding to the categorical data, and a mask bias item of the masked fully connected layer corresponding to the categorical data.

[0138] Specifically, the sparse attention mechanism generates a second feature selection mask for the second feature vector The specific calculation formula is: ,in is the weight parameter of the masked fully connected layer corresponding to the categorical data, It is the mask bias item of the masked fully connected layer corresponding to the categorical data.

[0139] S503. Obtain a third feature selection mask corresponding to the third feature vector according to the third feature vector, a weight parameter of the masked fully connected layer corresponding to the time series data, and a mask bias item of the masked fully connected layer corresponding to the time series data.

[0140] Specifically, the sparse attention mechanism generates a second feature selection mask for the second feature vector The specific calculation formula is: ,in is the weight parameter of the masked fully connected layer corresponding to the time series data, It is the mask bias term of the masked fully connected layer corresponding to the time series data.

[0141] Figure 8 A flow chart of a method for determining a shared feature vector provided in an embodiment of the present application is shown as follows: Figure 8 As shown, in the above S104, the data fusion module performs fusion processing according to the first feature vector, the first feature selection mask, the second feature vector, the second feature selection mask, the third feature vector and the third feature selection mask to obtain a shared feature vector, which may include:

[0142] S601. Select a mask according to a first feature vector and a first characteristic to determine a first weighted feature.

[0143] Optionally, the first characteristic selection mask After generation, the first feature selection mask is combined with the extracted first feature vector Multiply element by element to perform feature weighting operation. Specifically, the first weighted feature is ,in Represents element-wise multiplication.

[0144] Optionally, element-by-element multiplication ensures that the most important features of the numerical data are retained, while unimportant features are suppressed. The first weighted feature of the final output is The dimension is .

[0145] S602: Determine a second weighted feature according to the second feature vector and the second feature selection mask.

[0146] Optionally, the second feature selection mask After generation, the second feature selection mask is combined with the extracted second feature vector Multiply element by element to perform feature weighting operation. Specifically, the second weighted feature is ,in Represents element-wise multiplication.

[0147] Optionally, element-by-element multiplication ensures that the most important features of the categorical data are retained, while unimportant features are suppressed. The second weighted feature of the final output is The dimension is .

[0148] S603: Determine a third weighted feature according to the third feature vector and the third feature selection mask.

[0149] Optionally, the third feature selection mask After generation, the third feature selection mask is combined with the extracted third feature vector Multiply each element and perform feature weighting operation. Specifically, the third weighted feature is ,in Represents element-wise multiplication.

[0150] Optionally, element-by-element multiplication ensures that the most important time steps in the time series data for the current task are retained, while unimportant time step information is suppressed. The third weighted feature of the final output is The dimension is .

[0151] S604: Concatenate the first weighted feature, the second weighted feature, and the third weighted feature to obtain a shared feature vector.

[0152] Optionally, since the dimensions of the first weighted feature, the second weighted feature and the third weighted feature are the same, the first weighted feature, the second weighted feature and the third weighted feature can be concatenated to obtain a shared feature vector , shared feature vector The size is .

[0153] In this embodiment, by performing mask operations on each feature vector, important features of each heterogeneous data source can be retained and unimportant features can be suppressed, thereby facilitating subsequent feature extraction of shared feature vectors in different tasks, making the extracted task feature vectors more precise, thereby facilitating accurate task prediction.

[0154] Fig. 9 This is a complete flow chart for obtaining a shared feature vector in an embodiment of the present application, such as Fig. 9 As shown, for numerical data, the numerical data is first input into the two fully connected layers in the feature extraction module in the preprocessing module to obtain the first feature vector, and then the first feature vector is masked using a mask mechanism to obtain the first weighted feature.

[0155] For categorical data, the categorical data is first input into an embedding layer for embedding processing, and then the embedded data is input into a fully connected layer for feature extraction to obtain a second feature vector, and then the second feature vector is masked using a mask mechanism to obtain a second weighted feature.

[0156] For time series data, a multi-head attention feature vector is obtained through embedding representation and multi-head attention mechanism, and then the multi-head attention feature vector is input into the average of the feedforward neural sub-module, the pooling sub-module and the second linear layer to obtain the third feature vector. Then, the third feature vector is masked using the mask mechanism to obtain the third weighted feature.

[0157] The first weighted feature, the second weighted feature and the third weighted feature are fused using a data fusion module to obtain a shared feature vector.

[0158] Optionally, the above S601, selecting a mask according to the first feature vector and the first characteristic to determine the first weighted feature, may include:

[0159] Optionally, the first feature selection mask is multiplied element-by-element by the first feature vector to obtain a first weighted feature.

[0160] Fig.10 A flow chart of a method for determining a target task feature vector provided in an embodiment of the present application is shown as follows: Fig.10 As shown, in S105, the task feature extraction module corresponding to each task determines the target task feature vector of each task according to the task parameters and the shared feature vector of each task, which may include:

[0161] S701 . Extract features from a shared feature vector according to task parameters of each task to obtain a task feature vector corresponding to each task.

[0162] The dimensions of the task feature vectors corresponding to each task are the same. The task parameters may include a weight matrix and a bias term.

[0163] Optionally, the task feature extraction module may include two fully connected layers for task feature extraction.

[0164] Exemplarily, for task 1, it may include the weight matrix and bias item corresponding to task 1, specifically, it may include the weight matrix corresponding to task 1 in the first fully connected layer, the bias item corresponding to task 1 in the first fully connected layer, the weight matrix corresponding to task 1 in the second fully connected layer, and the bias item corresponding to task 1 in the second fully connected layer. For task 2, it may include the weight matrix and bias item corresponding to task 2, specifically, it may include the weight matrix corresponding to task 2 in the first fully connected layer, the bias item corresponding to task 2 in the first fully connected layer, the weight matrix corresponding to task 2 in the second fully connected layer, and the bias item corresponding to task 2 in the second fully connected layer. For task 3, it may include the weight matrix and bias item corresponding to task 3, specifically, it may include the weight matrix corresponding to task 3 in the first fully connected layer, the bias item corresponding to task 3 in the first fully connected layer, the weight matrix corresponding to task 3 in the second fully connected layer, and the bias item corresponding to task 3 in the second fully connected layer.

[0165] Optionally, the specific calculation process of the task feature vector corresponding to task 1 is as follows: first, the shared feature vector Input to the first fully connected layer for feature extraction. The output dimension of the first fully connected layer is 128, and the activation function uses the ReLU activation function. The specific calculation formula of the first fully connected layer is: ,in is the weight matrix corresponding to task 1 in the first fully connected layer, is the bias term corresponding to the first fully connected layer of task 1. Input to the second fully connected layer for further feature extraction. The specific calculation formula is: ,in is the weight matrix corresponding to task 1 in the second fully connected layer, The bias item corresponding to task 1 in the second fully connected layer is used to generate the task feature vector of task 1 through the second fully connected layer. The dimension is .

[0166] Optionally, the specific calculation process of the task feature vector corresponding to task 2 is as follows: first, the shared feature vector Input to the first fully connected layer for feature extraction. The output dimension of the first fully connected layer is 128, and the activation function uses the ReLU activation function. The specific calculation formula of the first fully connected layer is: ,in is the weight matrix corresponding to the first fully connected layer of task 2 ， is the bias term corresponding to the first fully connected layer of Task 2. Input to the second fully connected layer for further feature extraction. The specific calculation formula is: ,in is the weight matrix corresponding to task 2 in the second fully connected layer, The bias item corresponding to task 2 in the second fully connected layer is used to generate the task feature vector of task 2 through the second fully connected layer. The dimension is .

[0167] Optionally, the specific calculation process of the task feature vector corresponding to task 3 is as follows: first, the shared feature vector Input to the first fully connected layer for feature extraction. The output dimension of the first fully connected layer is 128, and the activation function uses the ReLU activation function. The specific calculation formula of the first fully connected layer is: ， in is the weight matrix corresponding to the first fully connected layer of Task 3, is the bias term corresponding to the first fully connected layer of Task 3. Input to the second fully connected layer for further feature extraction. The specific calculation formula is: ,in is the weight matrix corresponding to Task 3 in the second fully connected layer, The bias item corresponding to task 3 in the second fully connected layer is used to generate the task feature vector of task 3 through the second fully connected layer. The dimension is .

[0168] S702: Perform cross-task fusion processing on the feature vectors of each task to obtain the target task characteristic vector of each task.

[0169] Optionally, the task feature vectors corresponding to each task are the features of the tasks themselves extracted by the feature extraction module. Therefore, a cross-weight matrix can be used to perform cross-task fusion processing on the task feature vectors to achieve cross-task information fusion, thereby obtaining the target task feature vectors of each task.

[0170] Among them, the dimension of the target task feature vector of each task obtained by fusion processing still remains .

[0171] Fig.11 A flowchart of another method for determining a target task feature vector provided in an embodiment of the present application is shown as follows: Fig.11 As shown, the above S702, performing cross-task fusion processing on the feature vectors of each task to obtain the target task characteristic vector of each task, may include:

[0172] S801. Obtain a cross weight matrix.

[0173] Optionally, the parameter value in the i-th row and j-th column in the cross weight matrix is ​​used to represent the weight of the information obtained by the task feature vector of the i-th task from the task feature vector of the j-th task, and i and j are integers greater than 0 respectively.

[0174] Among them, the cross weight matrix is ​​as shown in the following formula (IV).

[0175] Formula (IV)

[0176] Among them, the parameter value represents the information weight that the task feature vector of task i obtains from the task feature vector of task j. is the information weight obtained from the task feature vector of task 1, is the information weight obtained by the task feature vector of task 1 from the task feature vector of task 2, is the information weight obtained by the task feature vector of task 1 from the task feature vector of task 3, is the information weight obtained by the task feature vector of task 2 from the task feature vector of task 1, is the information weight obtained by the task feature vector of task 3 from the task feature vector of task 1, and so on.

[0177] S802: For task k, obtain the parameter values ​​of the k-th row in the cross weight matrix.

[0178] Here, k is an integer greater than 0.

[0179] For example, for Task 1, get the parameter values ​​of row 1. , as well as ; For Task 2, get the parameter values ​​of the second row, specifically , as well as ; For Task 3, get the parameter values ​​of row 3, specifically , as well as .

[0180] S803 , calculating the product of each parameter value in the kth row and the target task feature vector of each associated task corresponding to each parameter value.

[0181] Among them, if the parameter value is located in the m column in the cross weight matrix, the associated task corresponding to the parameter value is task m, and m is an integer greater than 0.

[0182] For example, for task 1, the parameter values ​​in the first row are obtained. , as well as , where the parameter value The corresponding associated task is task 1, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 1 is ; Parameter value The corresponding associated task is task 2, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 2 is ; Parameter value The corresponding associated task is task 3, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 3 is .

[0183] For example, for task 2, the parameter values ​​in the first row are obtained. , as well as , where the parameter value The corresponding associated task is task 1, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 1 is ; Parameter value The corresponding associated task is task 2, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 2 is ; Parameter value The corresponding associated task is task 3, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 3 is .

[0184] For example, for task 3, the parameter values ​​in the first row are obtained. , as well as , where the parameter value The corresponding associated task is task 1, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 1 is ; Parameter value The corresponding associated task is task 2, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 2 is ; Parameter value The corresponding associated task is task 3, then the parameter value With parameter value The product of the target task feature vector of the corresponding associated task 3 is .

[0185] S803 . The sum of the products of each parameter value in the kth row and each associated task corresponding to each parameter value is taken as the target task feature vector of task k.

[0186] Specifically, the target task feature vector of Task 1 is ; The target task feature vector of task 2 is ; The target task feature vector of task 3 is .

[0187] In this embodiment, the task feature vectors of each task are cross-fused through a cross-weight matrix, so that the target task features of each task obtained can include the influence of task features of other tasks, and the target task features of each task obtained can be more comprehensive and accurate.

[0188] Optionally, after obtaining the target task vector of each task, the prediction module corresponding to each task feature extraction module may be used to perform prediction processing on the target task feature vector of each task, specifically:

[0189] Optionally, when the business object is an employee, if Task 1 is a need to improve academic qualifications, a binary classification prediction module can be used. When the business object is an animal, if Task 1 is a need to increase nutrition, a binary classification prediction module can also be used.

[0190] Specifically, first enter After the first fully connected layer in the binary classification prediction module, the output dimension is 32, and the activation function uses ReLU. ,in, is the weight matrix of the first fully connected layer in the binary classification prediction model, is the bias term of the first fully connected layer in the binary classification prediction model. The input is sent to the second fully connected layer, and the output dimension is a feature vector of 16, and the activation function is ReLU. Specifically, ,in, is the weight matrix of the second fully connected layer in the binary classification prediction model, It is the bias term of the second fully connected layer in the binary classification prediction model. The dimension is 16, the activation function is Sigmoid, and it is used for binary classification probability prediction. Specifically, in, is the prediction weight matrix of the binary prediction module, It is the prediction bias term of the binary classification prediction module. is a A vector of , classified as 0 or 1 by judging whether each value is greater than 0.5.

[0191] Optionally, when the business object is an employee, if Task 2 is a capability level assessment, a three-category prediction module may be used. When the business object is an animal, if Task 2 is a growth level assessment, a three-category prediction module may also be used.

[0192] Specifically, the target task feature vector of task 2 can be After the first fully connected layer in the three-class prediction module, the output dimension is 32, and the activation function uses ReLU. The calculation process is: ,in, is the weight matrix of the first fully connected layer in the three-classification prediction model, is the bias term of the first fully connected layer in the three-class prediction model. Input to the second fully connected layer in the three-class prediction module, the output dimension is 16, the activation function is ReLU, and the calculation process is , is the weight matrix of the second fully connected layer in the three-class prediction model, is the bias term of the second fully connected layer in the three-class prediction model. Input to the classification layer, output dimension 3, activation function is Softmax, used for the probability distribution of three categories. in, is the prediction weight matrix of the three-class prediction module, is the prediction bias item of the three-class prediction module, and the output The dimension is , is n one-hot vectors.

[0193] Optionally, when the business object is an employee, if Task 3 is the prediction of the probability of attrition, a regression prediction module may be used; when the business object is an animal, if Task 3 is the probability of gaining weight, a regression prediction module may also be used.

[0194] Specifically, the target task feature vector of task 3 is Input to the first fully connected layer in the regression prediction module, output feature vector with dimension 32, and activation function is ReLU. The calculation process is: in, is the weight matrix of the first fully connected layer of the regression prediction module, is the bias term of the first fully connected layer of the regression prediction module. Input to the second fully connected layer, output feature vector of dimension 16, activation function is ReLU, calculation process is ,in, is the weight matrix of the second fully connected layer in the regression prediction module, is the bias term of the second fully connected layer in the regression prediction module. Input to the final regression layer, the output dimension is 1, and the activation function is a linear activation function for regression prediction. = ,in, is the prediction weight matrix of the prediction layer in the regression prediction module, It is the prediction bias item of the prediction layer in the regression prediction module.

[0195] Optionally, the loss function for the target model is calculated as follows:

[0196] Optionally, Task 1 is a binary classification task, and the loss can be calculated using binary cross entropy loss, as shown in the following formula (V).

[0197] Formula (V)

[0198] in, yes The true labels of samples, is the predicted value of the model, is the total sample size.

[0199] Optionally, Task 2 is a three-classification task, and the loss is calculated using multivariate cross entropy loss, as specifically shown in the following formula (VI).

[0200] Formula (VI)

[0201] in, It is The sample in The true label of the class, The model predicts The samples belong to The probability of the class.

[0202] Optionally, Task 3 is a regression task, and the loss is calculated using mean square error loss, as specifically shown in the following formula (VII).

[0203] Formula (VII)

[0204] in, It is The true value of the samples, is the predicted value of the model.

[0205] The following table 1 is for sample size The result of one run, in which the binary classification task Represented as 0-1 classification, three-classification task results Represented by one-hot vector, the predicted probability A floating point number.

[0206] Table 1

[0207]

[0208] Optionally, in order to jointly optimize the models of these three tasks, the loss function of each task can be weighted and summed according to different weights. Assume that the weights are , , , then the total loss function is:

[0209]

[0210] In practical applications, the weight The contribution of each task to the overall training can be balanced through cross-validation or adjustment based on the importance of the task. The target model will ensure that the three tasks learn the optimal features at the same time by jointly optimizing the loss of the three tasks.

[0211] Fig.12 This is a structural block diagram of an electronic device 900 provided in an embodiment of the present application. The electronic device may be, for example, an information processor based on heterogeneous data fusion and multi-task learning as described in the above embodiments. Fig.12 As shown, the electronic device may include: a processor 901 and a memory 902 .

[0212] Optionally, a bus 903 may also be included, wherein the memory 902 is used to store machine-readable instructions executable by the processor 901. When the electronic device 900 is running, the processor 901 communicates with the memory 902 through the bus 903. When the machine-readable instructions are executed by the processor 901, the method steps in the above method embodiment are performed.

[0213] An embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method steps in the above-mentioned information processing method embodiment based on heterogeneous data fusion and multi-task learning are executed.

[0214] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the method embodiment, and will not be repeated in this application. In the several embodiments provided in this application, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0215] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), disk or optical disk and other media that can store program code.

[0216] The above are only specific implementation methods of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be covered by the protection scope of the present application.

Claims

1. An information processing method based on heterogeneous data fusion and multi-task learning, characterized in that: The method comprises: Input the current data of the business object into the target model, the current data includes: numerical data, categorical data and time series data, the target model includes: a preprocessing module, a data fusion module, a plurality of task feature extraction modules connected to the data fusion module, and a prediction module corresponding to each task feature extraction module, wherein each task feature extraction module is used to extract features under a task type, the business object is an employee, the numerical data includes the employee's working years and age, the categorical data includes the employee's education and title, and the time series data includes the employee's work mobility data and the employee's performance change trend data; The preprocessing module determines a first eigenvector of the numerical data, a second eigenvector of the categorical data, and a third eigenvector of the time series data, wherein the dimensions of the first eigenvector, the second eigenvector, and the third eigenvector are the same; The preprocessing module performs mask processing on the first feature vector, the second feature vector and the third feature vector respectively to obtain a first feature selection mask corresponding to the first feature vector, a second feature selection mask corresponding to the second feature vector and a third feature selection mask corresponding to the third feature vector; The data fusion module performs fusion processing according to the first feature vector, the first feature selection mask, the second feature vector, the second feature selection mask, the third feature vector and the third feature selection mask to obtain a shared feature vector; Obtaining task parameters of each task in a plurality of tasks, performing feature extraction on the shared feature vector according to the task parameters of each task, and obtaining a task feature vector corresponding to each task, wherein the task parameters include: a weight matrix and a bias term; performing cross-task fusion processing on each of the task feature vectors, and obtaining a target task feature vector of each task, wherein the plurality of tasks include: academic qualification improvement needs, ability level assessment, and loss probability prediction; The prediction module corresponding to each task feature extraction module performs prediction processing according to the target task feature vector of each task to obtain the prediction result of the business object under each task.

2. The information processing method based on heterogeneous data fusion and multi-task learning according to claim 1 is characterized in that: The determining of the first feature vector of the numerical data, the second feature vector of the categorical data, and the third feature vector of the time series data by the preprocessing module includes: The numerical data is directly input into the feature extraction submodule of the preprocessing module to extract features, and obtain a first feature vector; Inputting the categorical data into an embedding submodule and a feature extraction submodule in a preprocessing module in sequence to obtain the second feature vector; The time series data is sequentially input into the first sub-linear layer, the multi-head attention sub-module, the feedforward neural sub-module, the average pooling sub-module, and the second sub-linear layer in the preprocessing module to obtain the third eigenvector.

3. The information processing method based on heterogeneous data fusion and multi-task learning according to claim 2 is characterized in that: The step of sequentially inputting the categorical data into an embedding submodule and a feature extraction submodule in a preprocessing module to obtain the second feature vector includes: The embedding submodule embeds each subcategory type field into sub-embedded data corresponding to each subcategory type according to the embedding matrix corresponding to each subcategory type, and obtains embedded data corresponding to the category type data according to each sub-embedded data; The embedded data corresponding to the categorical data is input into a feature extraction submodule, and the feature extraction submodule performs feature extraction on the embedded data according to extraction parameters corresponding to the categorical data to obtain the second feature vector.

4. The information processing method based on heterogeneous data fusion and multi-task learning according to claim 2 is characterized in that: The step of sequentially inputting the time series data into the first linear layer, the multi-head attention submodule, the feedforward neural submodule, the average pooling submodule, and the second linear layer in the preprocessing module to obtain the third feature vector includes: The first linear layer performs dimension conversion and position encoding on the time series data to obtain first time series data; Input the first time series data into the multi-head attention submodule to extract time features, obtain the second time series data output by each attention head, and concatenate and linearly transform each of the second time series data to obtain a multi-head attention feature vector; Inputting the multi-head attention feature vector into the feedforward neural submodule to extract type features to obtain a type feature vector; Inputting the type feature vector into the average pooling submodule to aggregate and pool the time features in the type feature vector to obtain pooled features; The pooled features are input into the second linear layer for projection processing to obtain the third feature vector.

5. The information processing method based on heterogeneous data fusion and multi-task learning according to claim 1 is characterized in that: The preprocessing module performs mask processing on the first feature vector, the second feature vector, and the third feature vector respectively to obtain a first feature selection mask corresponding to the first feature vector, a second feature selection mask corresponding to the second feature vector, and a third feature selection mask corresponding to the third feature vector, including: Obtaining a first feature selection mask corresponding to the first feature vector according to the first feature vector, a weight parameter of a masked fully connected layer corresponding to the numerical data, and a mask bias item of the masked fully connected layer corresponding to the numerical data; Obtaining a second feature selection mask corresponding to the second feature vector according to the second feature vector, a weight parameter of the masked fully connected layer corresponding to the categorical data, and a mask bias item of the masked fully connected layer corresponding to the categorical data; A third feature selection mask corresponding to the third feature vector is obtained according to the third feature vector, a weight parameter of the masked fully connected layer corresponding to the time series data, and a mask bias item of the masked fully connected layer corresponding to the time series data.

6. The information processing method based on heterogeneous data fusion and multi-task learning according to claim 1 is characterized in that: The data fusion module performs fusion processing according to the first feature vector, the first feature selection mask, the second feature vector, the second feature selection mask, the third feature vector and the third feature selection mask to obtain a shared feature vector, including: Determining a first weighted feature according to the first feature vector and the first feature selection mask; Determining a second weighted feature according to the second feature vector and the second feature selection mask; Determining a third weighted feature according to the third feature vector and the third feature selection mask; The first weighted feature, the second weighted feature and the third weighted feature are concatenated to obtain a shared feature vector.

7. The information processing method based on heterogeneous data fusion and multi-task learning according to claim 6 is characterized in that: The determining a first weighted feature according to the first feature vector and the first feature selection mask includes: The first feature selection mask is multiplied element-by-element by the first feature vector to obtain the first weighted feature.

8. The information processing method based on heterogeneous data fusion and multi-task learning according to claim 1 is characterized in that: The cross-task fusion processing of each task feature vector to obtain the target task feature vector of each task includes: Obtain a cross weight matrix, wherein the parameter value in the i-th row and j-th column in the cross weight matrix is ​​used to represent the weight of the information obtained by the task feature vector of the i-th task from the task feature vector of the j-th task, and i and j are integers greater than 0 respectively; For task k, obtain the parameter values ​​of the kth row in the cross weight matrix, where k is an integer greater than 0; Calculate the product of each parameter value in the kth row and the target task feature vector of each associated task corresponding to each parameter value; The sum of the products of each parameter value in the kth row and each associated task corresponding to each parameter value is taken as the target task feature vector of task k, where if the parameter value is in the mth column, the associated task corresponding to the parameter value is task m, and m is an integer greater than 0.

9. An electronic device, characterized in that: It includes a memory and a processor, the memory stores a computer program executable by the processor, and the processor implements the steps of the information processing method based on heterogeneous data fusion and multi-task learning as described in any one of claims 1 to 8 when executing the computer program.

Images