CN-121981337-A - Training method, generating method, device, equipment and medium of demand prediction model

CN121981337ACN 121981337 ACN121981337 ACN 121981337ACN-121981337-A

Abstract

The invention provides a training method of a demand prediction model, a push content generation method, a push content generation device, electronic equipment and a computer readable storage medium, and relates to the technical field of machine learning. The training method comprises the steps of obtaining a plurality of client demands and corresponding multi-source behavior data and multi-source product information thereof, carrying out feature extraction based on paragraph semantic equilibrium normalization PSBN on the multi-source behavior data and the multi-source product information to obtain behavior feature vectors and product feature vectors corresponding to the client demands, constructing a feature matrix of the self, training a deep neural network DNN model based on the feature matrix of the self to obtain an initial demand prediction model of the self, and training a target demand prediction model of the self through a central server and other local nodes based on federation learning, the feature matrix of the self and the initial demand prediction model. To at least solve the problem of insufficient model prediction accuracy in the related art. The method is suitable for demand prediction and content pushing scenes.

Inventors

YANG ZHONGYUE
XU RUI
Jing Xiaopi
Jia Yaoming
ZHANG YAN
MA JUN
LI SHAOHUI
ZHENG JIE
FU YUNJIAN

Assignees

中国联合网络通信集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260320

Claims (10)

1. The training method of the demand prediction model is applied to any one of a plurality of local nodes, and is characterized by comprising the following steps: acquiring a plurality of customer demands and corresponding multi-source behavior data and multi-source product information thereof; Performing feature extraction based on paragraph semantic equilibrium normalization PSBN on the multi-source behavior data and the multi-source product information to obtain behavior feature vectors and product feature vectors corresponding to a plurality of client demands, and constructing a feature matrix of the multi-source behavior data and the multi-source product information, wherein the feature matrix comprises the behavior feature vectors and the product feature vectors corresponding to the plurality of client demands and association relations of the behavior feature vectors and the product feature vectors corresponding to the plurality of client demands; training a deep neural network DNN model based on the characteristic matrix of the model to obtain an initial demand prediction model of the model; Based on federal learning, own feature matrix and initial demand prediction model, the own target demand prediction model is trained jointly with other local nodes through a central server.
2. The method of claim 1, wherein the customer requirements include a demand product, The obtaining the multi-source behavior data and the multi-source product information corresponding to the plurality of client demands specifically comprises the following steps: Acquiring product information of a required product in different data sources, and historical communication records and historical browsing records of clients corresponding to a plurality of client requirements in different data sources; respectively calculating the similarity between the historical communication records of the client in different data sources, the similarity between the historical communication records of the client in different data sources and the similarity between the product information of the required product in different data sources; and combining the product information with the similarity larger than a preset threshold, the history communication record and the history browsing record to obtain multi-source behavior data and multi-source product information corresponding to each client requirement.
3. The method for training a demand prediction model according to claim 1, wherein the feature extraction based on paragraph semantic equilibrium normalization PSBN is performed on the multi-source behavior data and the multi-source product information to obtain behavior feature vectors and product feature vectors corresponding to a plurality of customer demands, and the method specifically comprises: inputting the multi-source behavior data and the multi-source product information into a pre-training language model, and respectively extracting multi-source behavior semantic vectors and multi-source product semantic vectors, wherein the pre-training language model comprises a bi-directional encoder representation BERT model based on a transducer; Respectively calculating the mean value and standard deviation of the multi-source behavior semantic vector and the multi-source product semantic vector; based on the mean value and standard deviation of the multi-source behavior semantic vector and the multi-source product semantic vector, the multi-source behavior semantic vector and the multi-source product semantic vector are normalized respectively, and behavior feature vectors and product feature vectors corresponding to a plurality of customer demands are obtained.
4. The training method of a demand prediction model according to claim 3, wherein the normalizing the multi-source behavior semantic vector and the multi-source product semantic vector based on the mean and the standard deviation of the multi-source behavior semantic vector and the multi-source product semantic vector respectively obtains a behavior feature vector and a product feature vector corresponding to a plurality of customer demands, specifically comprising: Based on the mean value, standard deviation and the following formulas of the multi-source behavior semantic vector and the multi-source product semantic vector, respectively normalizing the multi-source behavior semantic vector and the multi-source product semantic vector to obtain behavior feature vectors and product feature vectors corresponding to a plurality of customer demands: , Wherein, the Representing behavior feature vectors or product feature vectors corresponding to a plurality of customer demands, Representing multi-source behavioral semantic vectors or multi-source product semantic vectors, Representing the mean of the multi-source behavioral semantic vector or the multi-source product semantic vector, Representing standard deviation of multi-source behavior semantic vectors or multi-source product semantic vectors, Representing preset parameters.
5. The method for training a demand prediction model according to claim 1, wherein the training of the target demand prediction model based on federal learning, a feature matrix of the target demand prediction model and an initial demand prediction model is performed jointly with other local nodes via a central server, specifically includes: S1, assigning i as1, and determining an initial demand prediction model of the device as a demand prediction local model of the device in the ith round, wherein i is used for representing the frequency of using a feature matrix in the process of obtaining the demand prediction local model of the current round; S2, evaluating the model gradient of the self in the ith round, and sending the model gradient of the self in the ith round to a central server, so that the central server aggregates the model gradients of a plurality of local nodes in the ith round to obtain a global model gradient of the plurality of local nodes in the ith round; S3, receiving a global model gradient of the demand of the ith round sent by the central server, training a self demand prediction local model of the ith round based on a characteristic matrix of the self global model gradient, the global model gradient of the ith round sent by the central server, obtaining a self demand prediction local model of the ith round in the (i+1) th round, and assigning i as i+1; And repeatedly executing S2-S3 until the demand prediction local model of the self in the (i+1) th round converges, and determining the demand prediction local model of the self in the (i+1) th round as a target demand prediction model of the self.
6. A push content generation method applied to a client of any one of a plurality of local nodes, comprising: acquiring multisource behavior data of a current client and a target demand prediction model of a local node, wherein the target demand prediction model is trained by a training method of the demand prediction model according to any one of claims 1-5; inputting the multisource behavior data of the current customer into a target demand prediction model, and predicting the customer demand of the current customer and the corresponding multisource product information thereof; Based on the client requirements of the current client and the corresponding multi-source product information, corresponding push content is generated.
7. The training device of the demand prediction model is applied to any one of a plurality of local nodes and is characterized by comprising a first acquisition module, a feature extraction module, a first training module and a second training module, A first acquisition module for acquiring a plurality of customer demands and corresponding multi-source behavior data and multi-source product information thereof, The feature extraction module is used for carrying out feature extraction based on paragraph semantic equilibrium normalization PSBN on the multi-source behavior data and the multi-source product information to obtain behavior feature vectors and product feature vectors corresponding to a plurality of client demands and constructing a feature matrix of the feature extraction module, wherein the feature matrix comprises the behavior feature vectors and the product feature vectors corresponding to the client demands and the association relations thereof, A first training module for training a deep neural network DNN model based on the characteristic matrix of the first training module to obtain an initial demand prediction model of the first training module, The second training module is used for training the target demand prediction model of the second training module with other local nodes through the central server based on federal learning, the characteristic matrix of the second training module and the initial demand prediction model.
8. A push content generating device is applied to a client of any one of a plurality of local nodes and is characterized by comprising a second acquisition module, a prediction module and a generation module, A second obtaining module, configured to obtain multi-source behavior data of a current client and a target demand prediction model of a local node to which the current client belongs, where the target demand prediction model is obtained by training the training device of the demand prediction model according to claim 7, A prediction module for inputting the multisource behavior data of the current customer into a target demand prediction model to predict the customer demand of the current customer and the multisource product information corresponding to the customer demand, And the generation module is used for generating corresponding push content based on the client requirements of the current client and the corresponding multi-source product information.
9. An electronic device comprising a memory and a processor, the memory having stored therein a computer program, the processor being arranged to run the computer program to implement a method of training a demand prediction model according to any one of claims 1 to 5 or a method of generating push content according to claim 6.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements a method for training a demand prediction model according to any one of claims 1 to 5 or a method for generating push content according to claim 6.

Description

Training method, generating method, device, equipment and medium of demand prediction model Technical Field The present invention relates to the field of machine learning technologies, and in particular, to a training method of a demand prediction model, a method and apparatus for generating push content, an electronic device, and a computer readable storage medium. Background In the cross-regional and cross-department product marketing collaboration of clustered enterprises, the sharing and collaboration of the related knowledge of the product marketing is of great importance. However, data isolation between the branch companies and departments causes difficulty in information sharing, data privacy and security are ensured when sensitive data are processed, and in addition, updating and propagation speed of related knowledge of product promotion are slow, so that service response efficiency is affected. Thus, the above problems limit the quick response and collaboration efficiency of collaboration services between the local branches and departments. Federal learning is used as a distributed machine learning method, which allows multiple participants to train a model together without sharing original data, so that centralized storage and transmission of data are avoided, and the risk of data leakage is reduced. However, current federal learning does not model individuality for the business characteristics of the product promotion collaboration business, and does not fully consider the impact of data heterogeneity of raw data of various local companies and departments on the model performance. In addition, the related art generally adopts a simple feature extraction method, semantic information in original data cannot be fully utilized, and a model may excessively depend on a single data source or neglect key association, so that the prediction accuracy of the model is reduced. Disclosure of Invention The technical problem to be solved by the invention is to provide a training method of a demand prediction model, a push content generating method, a device, electronic equipment and a computer readable storage medium, aiming at the defects of the prior art, and the method can realize the training of the demand prediction model and the push of the content efficiently and accurately. The invention provides a training method of a demand prediction model, which is applied to any one of a plurality of local nodes and comprises the steps of obtaining a plurality of client demands and corresponding multisource behavior data and multisource product information thereof, carrying out feature extraction based on paragraph semantic equilibrium normalization PSBN on the multisource behavior data and the multisource product information to obtain behavior feature vectors and product feature vectors corresponding to the client demands, constructing a feature matrix of the self, wherein the feature matrix comprises the client demands, the behavior feature vectors and the product feature vectors corresponding to the client demands and association relations of the client demands, training a deep neural network DNN model based on the feature matrix of the self to obtain an initial demand prediction model of the self, and training a target demand prediction model of the self through a central server and other local nodes based on federal learning, the feature matrix of the self and the initial demand prediction model. The method comprises the steps of obtaining product information of a plurality of products in different data sources, historical communication records and historical browsing records of the clients in different data sources, calculating similarity among the historical communication records of the clients in different data sources, similarity among the historical browsing records of the clients in different data sources, similarity among the product information of the products in different data sources, and combining the product information, the historical communication records and the historical browsing records, wherein the similarity is larger than a preset threshold, of the product information, the historical communication records and the historical browsing records of the products in different data sources, and obtaining the multi-source behavior data and the multi-source product information corresponding to each client demand. Preferably, feature extraction based on paragraph semantic equilibrium normalization PSBN is performed on multi-source behavior data and multi-source product information to obtain behavior feature vectors and product feature vectors corresponding to a plurality of customer demands, and the method specifically comprises the steps of inputting the multi-source behavior data and the multi-source product information into a pre-training language model to respectively extract the multi-source behavior semantic vectors and the multi-source product semantic vectors, wherein the pre-training langua