CN115630224A

CN115630224A - Information push model construction method and device, computer equipment and storage medium

Info

Publication number: CN115630224A
Application number: CN202211212355.8A
Authority: CN
Inventors: 陈亮; 于宙鑫; 周克涌; 林昊; 余俭; 郑子彬; 张鹏
Original assignee: Merchants Union Consumer Finance Co Ltd; Sun Yat Sen University
Current assignee: Merchants Union Consumer Finance Co Ltd; Sun Yat Sen University
Priority date: 2022-09-29
Filing date: 2022-09-29
Publication date: 2023-01-20

Abstract

The application relates to an information push model construction method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring historical characteristic data; obtaining characterization embedded data according to the historical characteristic data, wherein the characterization embedded data comprises object characterization data, push content characterization data, push mode characterization data and push time characterization data; taking the object representation data and the push content representation data as training data to perform neural network training to obtain a push content sub-model; taking the object representation data and the push mode representation data as training data to perform neural network training to obtain a push mode sub-model; taking the object representation data and the push time representation data as training data to perform neural network training to obtain a push time sub-model; and parallelly splicing to obtain an information pushing model based on a pushing content submodel, a pushing mode submodel and a pushing time submodel. By adopting the method, the efficiency of generating online push information is improved.

Description

Information push model construction method and device, computer equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method and an apparatus for constructing an information push model, a computer device, and a storage medium.

Background

With the development of computer technology, methods for generating network push information by analyzing data through an intelligent algorithm are more and more widely applied, and more attention is paid to how to efficiently and accurately generate online push information.

In the conventional technology, online push information is generated simply according to a single decision condition, so that highly targeted push information cannot be accurately and efficiently generated, and the push cost is high.

Disclosure of Invention

In view of the foregoing, it is desirable to provide an information push model construction method, apparatus, computer device and computer readable storage medium, which can improve the efficiency of generating push information on line.

An information push model construction method is characterized by comprising the following steps:

acquiring historical characteristic data;

obtaining characterization embedded data according to the historical characteristic data, wherein the characterization embedded data comprises object characterization data, push content characterization data, push mode characterization data and push time characterization data;

taking the object representation data and the push content representation data as training data to perform neural network training to obtain a push content sub-model;

taking the object representation data and the push mode representation data as training data to carry out neural network training to obtain a push mode sub-model;

taking the object representation data and the push time representation data as training data to perform neural network training to obtain a push time sub-model;

and the information pushing model is obtained by parallelly splicing the pushing content submodel, the pushing mode submodel and the pushing time submodel, and is used for obtaining the corresponding information pushing content category, the information pushing mode category and the information pushing time according to the object representation data corresponding to the object to be pushed.

In one embodiment, after obtaining the historical feature data, the method further includes:

acquiring initial historical characteristic data;

performing relevance analysis on each dimension characteristic data in the initial historical characteristic data to obtain relevant dimension characteristic data;

analyzing the importance of the related dimension characteristic data to determine the characteristic importance of the related dimension characteristic data;

and sequencing the related dimension feature data according to the feature importance to generate historical feature data.

In one embodiment, deriving characterization embedding data from historical characterization data includes:

acquiring target object characteristic data in the historical characteristic data;

acquiring historical activity data corresponding to the characteristic data of the target object;

normalizing the browsing and staying time of the objects in the historical activity data to obtain a time weight parameter;

mapping each historical activity in the historical activity data into a corresponding historical activity vector, wherein the dimensions of the historical activity vectors corresponding to the historical activities are the same;

and fusing the historical activity vectors corresponding to the historical activities according to the time weight parameters to obtain characterization embedded data, wherein the characterization embedded data is used for characterizing the corresponding relation of the push content, the push mode and the push time of each historical activity corresponding to the characteristic data of the target object.

In one embodiment, the obtaining of the characterization embedding data by fusing the historical activity vectors corresponding to the time weight parameter and the historical activities includes:

fusing the historical activity vectors corresponding to the historical activities according to the time weight parameters to obtain initial embedded characterization data;

and processing the initial embedded characterization data by a multi-layer perceptron to obtain characterization embedded data, wherein the dimensionality of the characterization embedded data is lower than that of the initial embedded characterization data.

In one embodiment, the method for obtaining the push content submodel by performing neural network training on the object characterization data and the push content characterization data as training data includes:

projecting the object representation data through a multilayer perceptron to generate an object representation vector;

projecting the push content representation data through a multilayer perceptron to generate a push content representation vector, wherein the dimension of the push content representation vector is the same as that of the object representation vector;

obtaining a push weight based on a dot product of the push content characterization vector and the object characterization vector;

and training the object representation vector, the push content representation vector and the push weight as training sample data of the multilayer perceptron to obtain a push content sub-model.

In one embodiment, the method for obtaining the push mode submodel by training the neural network with the object characterization data and the push mode characterization data as training data includes:

projecting object representation data through a multilayer perceptron to generate an object representation vector;

respectively generating an application push characterization vector, a short message push characterization vector and a telephone push characterization vector according to application push characterization data, short message push characterization data and telephone push characterization data in the push mode characterization data;

obtaining an application push weight based on a dot product of the object representation vector and the application push representation vector;

obtaining a short message pushing weight based on the dot product of the object representation vector and the short message pushing representation vector;

obtaining a phone push weight based on a dot product of the object representation vector and the phone push representation vector;

and obtaining a pushing mode sub-model by taking the object representation vector, the application pushing weight, the short message pushing weight and the telephone pushing weight as model training sample data.

In one embodiment, the obtaining of the push time sub-model by performing neural network training using the object characterization data and the push time characterization data as training data includes:

dividing the pushing time characterization data to obtain a preset number of sub-time period characterization data;

projecting the sub-time period representation data through a multilayer perceptron to generate sub-time period representation vectors;

obtaining the object idle time probability of the sub-time period corresponding to the sub-time period characterization data according to the object characterization vector and the sub-time period characterization vector;

and training by taking the object idle time probability and the object representation vector of each sub-time period as input sample data of the multilayer perceptron to obtain a push time sub-model.

An information push model construction apparatus, the apparatus comprising:

the data acquisition module is used for acquiring historical characteristic data;

the representation embedded data generation module is used for obtaining representation embedded data according to the historical characteristic data, and the representation embedded data comprises object representation data, push content representation data, push mode representation data and push time representation data;

the model generation module is used for carrying out neural network training by taking the object representation data and the pushed content representation data as training data to obtain a pushed content sub-model; taking the object representation data and the push mode representation data as training data to carry out neural network training to obtain a push mode sub-model; taking the object representation data and the push time representation data as training data to perform neural network training to obtain a push time sub-model; and the information pushing model is obtained by parallelly splicing the pushing content submodel, the pushing mode submodel and the pushing time submodel, and is used for obtaining the corresponding information pushing content category, the information pushing mode category and the information pushing time according to the object representation data corresponding to the object to be pushed.

A computer device comprising a memory and a processor, the memory storing a computer program, the processor when executing the computer program implementing the steps of:

acquiring historical characteristic data;

taking the object representation data and the push mode representation data as training data to perform neural network training to obtain a push mode sub-model;

A computer-readable storage medium storing a computer program which, when executed by a processor, performs the steps of:

acquiring historical characteristic data;

taking the object representation data and the pushed content representation data as training data to carry out neural network training to obtain a pushed content sub-model;

The information push model construction method, the information push model construction device, the computer equipment and the storage medium are characterized in that historical characteristic data are obtained, characterization embedded data are obtained according to the historical characteristic data, the characterization embedded data comprise object characterization data, push content characterization data, push mode characterization data and push time characterization data, the object characterization data and the push content characterization data are used as training data to conduct neural network training to obtain a push content submodel, the object characterization data and the push mode characterization data are used as training data to conduct neural network training to obtain a push mode submodel, the object characterization data and the push time characterization data are used as training data to conduct neural network training to obtain a push time submodel, the information push model is obtained based on the push content submodel, the push mode submodel and the push time submodel in a parallel splicing mode, and is used for obtaining corresponding information push content types, information push mode types and information push time according to the object characterization data corresponding to the object to-be-pushed. Therefore, each submodel is obtained by performing neural network training on the object representation data, the push mode representation data and the push time representation data respectively, and then all the submodels are spliced in parallel to obtain the whole information push model.

Drawings

FIG. 1 is a diagram of an application environment of a method for constructing an information push model in one embodiment;

FIG. 2 is a flowchart illustrating a method for constructing an information push model according to an embodiment;

FIG. 3 is a flow diagram illustrating historical feature data generation in one embodiment;

FIG. 4 is a schematic flow chart diagram that illustrates the generation of the characterizing embedded data in one embodiment;

FIG. 5 is a schematic flow chart diagram illustrating the generation of characterization embedding data in one embodiment;

FIG. 6 is a flow diagram illustrating the generation of a push content submodel in one embodiment;

FIG. 7 is a flow diagram illustrating the generation of a push mode sub-model in one embodiment;

FIG. 8 is a flow diagram illustrating the generation of a push time submodel in one embodiment;

FIG. 9 is a block diagram showing an arrangement of an information push model building apparatus according to an embodiment;

FIG. 10 is a diagram showing an internal structure of a computer device according to an embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of and not restrictive on the broad application.

The information push model construction method provided by the embodiment of the application can be applied to an application environment shown in fig. 1. The computer device 102 obtains historical feature data; obtaining characterization embedded data according to the historical characteristic data, wherein the characterization embedded data comprises object characterization data, pushed content characterization data, pushed mode characterization data and pushed time characterization data; taking the object representation data and the push content representation data as training data to perform neural network training to obtain a push content sub-model; taking the object representation data and the push mode representation data as training data to perform neural network training to obtain a push mode sub-model; taking the object representation data and the push time representation data as training data to perform neural network training to obtain a push time sub-model; and the information pushing model is obtained by parallelly splicing the pushing content submodel, the pushing mode submodel and the pushing time submodel, and is used for obtaining the corresponding information pushing content category, the information pushing mode category and the information pushing time according to the object representation data corresponding to the object to be pushed. The computer device 102 may specifically include, but is not limited to, various personal computers, laptops, servers, smartphones, tablets, smart cameras, portable wearable devices, and the like.

In one embodiment, as shown in fig. 2, an information push model building method is provided, which is described by taking the method as an example applied to the computer device 102 in fig. 1, and includes the following steps:

in step S202, historical feature data is acquired.

The historical characteristic data is historical user behavior characteristic data and is data generated by previous user participation activities, the historical characteristic data comprises user side historical data and activity side historical data, the user side historical data is characteristic data of the user, such as the age, sex, occupation and the like of the user, and the activity side historical data is characteristic data of the activities, such as the holding time of the activities, the holding content of the activities, the activity pushing information mode, the pushing information time and the like.

Specifically, the computer device screens out the characteristic to be adopted of the historical activity data in a data pool according to the business logic of a business department and expert knowledge, screens out user side characteristic data and activity side characteristic data, and combines the user side characteristic data and the activity side characteristic data into historical characteristic data.

And S204, obtaining characterization embedded data according to the historical characteristic data, wherein the characterization embedded data comprises object characterization data, push content characterization data, push mode characterization data and push time characterization data.

The characterization embedded data is used for characterizing data of various dimension characteristics of historical characteristic data, the object characterization data is used for characterizing data of various dimension characteristics of user side characteristic data, the push content characterization data is used for characterizing data of various types of push contents of activity side characteristic data, the push mode characterization data is used for characterizing data of information push mode types in the activity side characteristic data, and the push time characterization data is used for characterizing data of different push time period types in the activity side characteristic data.

Specifically, the computer device performs embedding representation processing on user-side feature data in the history feature data acquired in the previous step to acquire user history interaction information and feature statistical information, and in consideration of the fact that each user may participate in a plurality of different history activities, sets corresponding activity weights according to the history interaction information corresponding to the current user, performs weighting fusion to obtain object representation parameters of the history activity information corresponding to the current user in the user-side feature data, splices the object representation parameters and the feature statistical information, generates lower-dimension embedding representation data, namely object representation data, as input of a neural network model, and performs the same processing on the activity-side feature data to obtain lower-dimension push content representation data, push mode representation data and push time representation data, thereby obtaining representation embedding data.

And S206, performing neural network training by using the object representation data and the pushed content representation data as training data to obtain a pushed content sub-model.

The dimensions and contents of the object representation data and the pushed content representation data are different, that is, the data dimensions of the object representation data and the pushed content representation data are different, and the neural network may be a model such as a multilayer perceptron, a generalized neural network, a convolutional neural network, and the like, which is not particularly limited herein.

Specifically, the computer device projects object representation data and pushed content representation data into the same low-dimensional space by using a multilayer perceptron to obtain representations in a new space, then obtains a probability score between the object representation data and recommended content representation data of a current user according to the representations of the object representation data and the pushed content representation data in the new space, takes the probability score as standard output data of a model, and trains by taking binary cross entropy as a loss function of the multilayer perceptron model to obtain a pushed content submodel.

And S208, performing neural network training by using the object characterization data and the push mode characterization data as training data to obtain a push mode submodel.

The dimensions and contents of the object representation data and the push mode representation data are different, that is, the data dimensions of the object representation data and the push mode representation data are different, the neural network can be a model such as a multilayer perceptron, a generalized neural network, a convolutional neural network and the like, and the push mode representation data is not particularly limited herein and comprises various push modes of different types such as application push, short message push, telephone push and the like.

Specifically, the computer device projects object representation data and different push mode type data in the push mode representation data to the same low-dimensional space by using a multilayer perceptron to obtain a representation in a new space, and obtains probability scores between the object representation data of a current user and the different push mode type data according to the representations of the object representation data and the different push mode type data in the new space, the probability scores are used as standard output data of a model, and the user is likely to accept multiple push modes, so a Pair-wise model is adopted for optimization, and finally a push mode submodel is obtained through training.

Step S210, the object representation data and the push time representation data are used as training data to conduct neural network training to obtain a push time sub-model.

The dimensionality and content of the object characterization data and the pushed time characterization data are different, namely the data dimensionalities of the object characterization data and the pushed time characterization data are different, the neural network can be a model such as a multilayer perceptron, a generalized neural network and a convolutional neural network, specific limitation is not made, the pushed time characterization data is obtained by labeling the idle time of a user from 8.

Specifically, the computer device inputs object representation data and push time representation data into a multilayer perceptron to project the object representation data and the push time representation data into the same low-dimensional space, and assigns corresponding labels to data in different time periods in the push time representation data, for example, the used labels include various labels such as transaction generation after push, user access service after push, no response after push, complaint after push, and the like, so as to obtain probability scores corresponding to the labels in each time period, the probability scores are used as standard output of a neural network, and multi-class cross entropy is used as a loss function to optimize in the training process, so as to obtain a push time submodel.

And S212, parallelly splicing to obtain an information pushing model based on the pushing content submodel, the pushing mode submodel and the pushing time submodel, wherein the information pushing model is used for obtaining the corresponding information pushing content category, the information pushing mode category and the information pushing time according to the object representation data corresponding to the object to be pushed.

In this embodiment, by obtaining historical feature data, token embedded data is obtained according to the historical feature data, where the token embedded data includes object token data, push content token data, push mode token data, and push time token data, the object token data and the push content token data are used as training data to perform neural network training to obtain a push content submodel, the object token data and the push mode token data are used as training data to perform neural network training to obtain a push mode submodel, the object token data and the push time token data are used as training data to perform neural network training to obtain a push time submodel, an information push model is obtained by parallelly splicing the push content submodel, the push mode submodel, and the push time submodel, and the information push model is used to obtain corresponding information push content categories, information push mode categories, and information push time according to the object token data corresponding to an object to be pushed. Therefore, each submodel is obtained by performing neural network training on the object representation data, the push mode representation data and the push time representation data respectively, and then all the submodels are spliced in parallel to obtain the whole information push model.

In one embodiment, as shown in fig. 3, after acquiring the historical feature data, the method further includes:

step S302, initial historical feature data is obtained.

Wherein the initial historical feature data is feature data generated by all users participating in historical activities.

Step S304, carrying out correlation analysis on each dimension characteristic data in the initial historical characteristic data to obtain related dimension characteristic data.

Specifically, the computer device screens out the features to be adopted (the number of the features is from tens to hundreds, etc.) according to experience, then analyzes the features with strong correlation in all the features to be adopted by using a Spearman correlation coefficient, takes the features with the correlation reaching a preset threshold as the relevant dimension feature data, and specifically calculates the correlation among the features according to the formula shown in the following formula 1:

represents the two characteristic variables (x) _i ,y _i ) The values in pairs are respectively from small to smallRank in order to big (or big to small), R _i Represents x _i Order of (2), Q _i Represents y _i In the order of (a) to (b), R _i -Q _i Is x _i 、y _i The difference in rank of (2).

Step S306, carrying out importance analysis on the related dimension characteristic data to determine the characteristic importance of the related dimension characteristic data.

Specifically, the computer device performs simple logistic regression and preliminary modeling of the tree model according to the relevant dimension feature data determined in the previous step, and calculates the feature importance of each feature in the relevant dimension feature data.

And step S308, sequencing the related dimension characteristic data according to the characteristic importance to generate historical characteristic data.

In this embodiment, the computer device obtains the initial historical feature data, performs relevance analysis on each dimension feature to obtain relevant dimension feature data with high relevance, then obtains the feature importance of each dimension feature in the relevant dimension feature data by using simple logistic regression and tree model initial modeling analysis, and performs ranking according to the feature importance to obtain the historical feature data, so that invalid redundant data is effectively removed, and the reliability of the historical feature data is enhanced.

In one embodiment, as shown in FIG. 4, deriving characterization embedding data from historical feature data includes:

in step S402, target object feature data in the history feature data is acquired.

The target object feature data is user data in historical activities, and comprises basic information such as user ages, professions, places of work and the like.

And S404, acquiring historical activity data corresponding to the characteristic data of the target object.

Specifically, the computer device obtains all historical activity data associated with the target object feature data in the database.

Step S406, normalizing the object browsing retention time in the historical activity data to obtain a time weight parameter.

Specifically, the computer device determines historical activity data corresponding to each target object characteristic data according to the previous steps, obtains the browsing stay time of the target user in each historical activity data corresponding to the target object characteristic data, and normalizes the browsing stay time to obtain the time weight.

Step S408, mapping each historical activity in the historical activity data to a corresponding historical activity vector, where the dimensions of the historical activity vectors corresponding to each historical activity are the same.

Specifically, the computer device analyzes each dimension of each historical activity in the historical activity data, maps the historical activity data into a corresponding historical activity vector according to the characteristics of the historical activity according to a preset rule, and ensures that the dimensions of the historical activity vectors corresponding to the historical activities are the same.

And step S410, fusing the historical activity vectors corresponding to the historical activities according to the time weight parameters to obtain characterization embedded data, wherein the characterization embedded data is used for characterizing the corresponding relation of the push content, the push mode and the push time of each historical activity corresponding to the characteristic data of the target object.

Specifically, the computer device performs embedding representation mapping on each historical activity vector, and then fuses the historical activity vector with the time weight parameters corresponding to each activity to obtain representation embedding data.

In the embodiment, the target object characteristic data in the historical characteristic data is obtained, the historical activity data corresponding to the target object characteristic data is obtained, the browsing retention time of the object in the historical activity data is normalized to obtain the time weight parameter, each historical activity in the historical activity data is mapped into the corresponding historical activity vector, the dimensions of the historical activity vector corresponding to each historical activity are the same, finally, the representation embedded data is obtained by fusing the time weight parameter and the historical activity vector corresponding to each historical activity, the representation embedded data is used for representing the corresponding relation of the push content, the push mode and the push time of each historical activity corresponding to the target object characteristic data, appropriate time weight can be accurately distributed according to the browsing time of the current user for each historical activity, the weight importance of the historical activity with high user relevance is improved, the generated representation embedded data reflects the relevance degree of different activities, and the accuracy of the generated representation embedded data is improved.

In an embodiment, as shown in fig. 5, the obtaining of the characterization embedding data by fusing the historical activity vectors corresponding to the historical activities according to the time weighting parameter includes:

and step S502, fusing according to the time weight parameter and the historical activity vector corresponding to each historical activity to obtain initial embedded characterization data.

And step S504, processing the initial embedded characterization data through a multilayer perceptron to obtain characterization embedded data, wherein the dimensionality of the characterization embedded data is lower than that of the initial embedded characterization data.

Specifically, the computer device calculates the characterization embedded data User in the manner shown in the following formula 2 to formula 4 _emb ：

User _{history_emb} ＝RELU(a ₁ *EMB(sale ₁ )+a ₂ *EMB(sale ₂ )+…+a _n *EMB(sale _n ))，

Equation 2

Wherein RELU is the activation function, a _n For time weighting, the browsing dwell time of the user browsing the corresponding activity can be directly defined or automatically learned by an attention mechanism, and the EMB represents an Embedding function which can uniquely project discrete historical activities into a high-dimensional dense space.

Then use the User _{history_emb} The method is spliced with the basic statistical characteristic User _ feature of the User, and the specific formula is as follows:

User_allemb＝Concat(User _{history_emb} user _ feature), equation 3

The basic statistical characteristic User _ feature of the User is a basic characteristic and a client behavior characteristic of the client.

And then, processing the data by a multilayer perceptron according to User _ allomb to obtain low-dimensional characterization embedded data User _emb Specifically, the following formula 4 shows:

User _emb = MLP (User _ ensemble), equation 4

Where MLP denotes a multi-layer perceptron.

In the embodiment, initial embedded characterization data are obtained by fusing the time weight parameters and the historical activity vectors corresponding to the historical activities, the initial embedded characterization data are processed by the multilayer perceptron to obtain characterization embedded data, the dimensionality of the characterization embedded data is lower than that of the initial embedded characterization data, the redundant initial embedded characterization data are simplified by the multilayer perceptron in such a way to generate the characterization embedded data with lower dimensionality, and the computing resources of the data are effectively saved.

In an embodiment, as shown in fig. 6, performing neural network training on object characterization data and push content characterization data as training data to obtain a push content sub-model, including:

step S602, projecting the object representation data through a multilayer perceptron to generate an object representation vector.

Specifically, the computer device projects object representation data capable of representing various dimensional features of the user-side feature data as input of a multilayer perceptron to generate corresponding object representation vectors, wherein the multilayer perceptron can also be replaced by other neural network models.

And step S604, projecting the push content representation data through a multilayer perceptron to generate a push content representation vector, wherein the dimension of the push content representation vector is the same as that of the object representation vector.

Specifically, the computer device projects push content characterization data capable of characterizing feature data of the active side as input of a multilayer perceptron, and generates a corresponding push content characterization vector, wherein the multilayer perceptron can also be replaced by other neural network models, and the dimension of the push content characterization vector is the same as that of the object characterization vector.

Step S606, a push weight is obtained based on a dot product of the push content characterization vector and the object characterization vector.

The push weight is used for representing the probability of successful push of the content of the current push information.

Specifically, the computer device calculates a dot product of the push content characterization vector and the object characterization vector according to a formula shown in the following formula 5, i.e. a push weight score:

score = Sigmoid (u (x) · v (y). T), formula 5

Wherein, T represents the transposition of the vector, u (x) represents the object representation vector, v (y) represents the push content representation vector, u (x) · v (y) represents the dot product of the push content representation vector and the object representation vector, and Sigmoid (u (x) · v (y) · T) represents the integration of the dot products into the interval (0, 1).

Step S608, training the object representation vector, the push content representation vector and the push weight as training sample data of the multilayer perceptron to obtain a push content sub-model.

Specifically, the computer device performs label identification on the push weight of each group of object representation vectors and push content representation vectors, the label identification includes that the push success rate is very high, the push success rate is relatively high, the push success rate is general, the push success rate is relatively low, and the push success rate is extremely low, the push weight and the corresponding label identification form training sample data, the training sample data is used as a training sample of the multilayer perceptron to train to obtain a push content sub-model, and Binary Cross Entropy (BCE) is used as a loss function to participate in training, as shown in the following formula 6:

L ₁ = BCE (score, label), equation 6

Wherein label is a label identifier, score is a push weight, and L ₁ To lose error.

In the embodiment, object representation data are projected through a multilayer perceptron to generate an object representation vector, push content representation data are projected through the multilayer perceptron to generate a push content representation vector, the dimension of the push content representation vector is the same as that of the object representation vector, push weights are obtained based on dot products of the push content representation vector and the object representation vector, the push content representation vector and the push weights are used as training sample data of the multilayer perceptron to train to obtain a push content sub-model, the push content representation data and the object representation data are projected to be vectors of the same dimension space, the dot products of the object representation vector and the push content representation data are used for calculating the probability of successful push, the probability is used as sample data of model training, and the precision of model training is effectively improved.

In an embodiment, as shown in fig. 7, performing neural network training on object characterization data and push mode characterization data as training data to obtain a push mode submodel includes:

step S702, projecting the object representation data through a multilayer perceptron to generate an object representation vector.

Step S704, respectively generating an application push characterization vector, a short message push characterization vector, and a telephone push characterization vector according to the application push characterization data, the short message push characterization data, and the telephone push characterization data in the push mode characterization data.

Specifically, the computer device projects push mode characterization data capable of characterizing feature data of the active side as input of a multilayer perceptron, and generates corresponding push content characterization vectors, specifically, generates application push characterization vectors, short message push characterization vectors, and telephone push characterization vectors from application push characterization data, short message push characterization data, and telephone push characterization data in the push mode characterization data, respectively, wherein the multilayer perceptron can also be replaced by other neural network models, and the vector dimensions of the application push characterization vectors, the short message push characterization vectors, and the telephone push characterization vectors are the same as the object characterization vectors.

Step S706, obtaining the application push weight based on the dot product of the object representation vector and the application push representation vector.

The application push weight is used for representing the successful push probability when the application software is adopted to push the information.

Step S708, a short message push weight is obtained based on a dot product of the object representation vector and the short message push representation vector.

The short message pushing weight is used for representing the successful pushing probability when the short message pushing information is adopted.

Step S710, a phone push weight is obtained based on a dot product of the object token vector and the phone push token vector.

Wherein the telephone push weight is used for representing the successful push probability when the telephone push information is adopted.

And step S712, obtaining a pushing mode sub-model by taking the object representation vector, the application pushing weight, the short message pushing weight and the telephone pushing weight as model training sample data.

Specifically, the computer device calculates dot products of application push weights, short message push weights, telephone push weights and current object characterization vectors respectively to obtain push success probabilities of the current object characterization vectors corresponding to various push modes, then performs label identification on respective recommended weights of different push modes to combine sample training data to train the multilayer perceptron to obtain a push mode submodel, and since the object characterization vectors correspond to various push modes, a Pair-wise model is adopted for optimization, as shown in the following formula 7:

wherein N is _S Set of all activities, L, representing that object S produces an association ₂ To lose error, σ denotes an activation function, which may be a sigmoid function, s _i Score for positive sample, s _j The score of the negative sample is obtained by training to maximize the difference between the scores of the positive and negative samples.

In the embodiment, object representation data are projected through a multilayer perceptron to generate object representation vectors, an application push representation vector, a short message push representation vector and a telephone push representation vector are respectively generated according to application push representation data, short message push representation data and telephone push representation data in push mode representation data, an application push weight is obtained based on a dot product of the object representation vector and the application push representation vector, a short message push weight is obtained based on a dot product of the object representation vector and the short message push representation vector, a telephone push weight is obtained based on a dot product of the object representation vector and the telephone push representation vector, the object representation vector, the application push weight, the short message push weight and the telephone push weight are used as model training sample data to obtain a push mode submodel, and a submodule takes a dot product between various recommendation modes and the object representation vector as a few value of push success, so that the reliability of the model is improved.

In one embodiment, as shown in fig. 8, performing neural network training on object characterization data and push time characterization data as training data to obtain a push time submodel, includes:

and S802, projecting the object representation data through a multilayer perceptron to generate an object representation vector.

Step S804, dividing the push time characterization data to obtain a preset number of sub-time period characterization data.

Specifically, the computer device performs segmentation labeling on the idle Time of the object corresponding to the object representation data in the step S802 from 8 to 20 every two hours, and uses One-Hot coding, and simultaneously splices whether the idle Time is a working day, a resting day, a holiday, an activity day, and the like as a combined feature (Torch _ feature) of a Time plane, and learns the sub-period representation data Time of the user Time by using the following formula 8 _emb ：

Time _emb = MLP (Torch _ feature), formula 8

Wherein the MLP is a multi-layer perceptron.

And step S806, projecting the sub-time period characterization data through a multilayer perceptron to generate sub-time period characterization vectors.

The sub-period token vector has the same dimension as the object token vector in step S802.

And step S808, obtaining the object idle time probability of the sub-time period corresponding to the sub-time period characterization data according to the object characterization vector and the sub-time period characterization vector.

Specifically, the computer device obtains the representation of six label dimensions of the time dimension by using a multilayer perceptron, and obtains the object idle time probability of the sub-time period corresponding to the sub-time period representation data.

And step S810, training the object idle time probability of each sub-time period and the object characterization vector as input sample data of the multilayer perceptron to obtain a push time sub-model.

Specifically, the computer device activates by using the softmax function, using the multi-class cross entropy as a loss function, as shown in the following equations 9 and 10:

p＝Softmax(MLP(Concat(User _emb ,Time _emb ) ))) of formula 9

Wherein p = [ p ] ₀ ,…,p _C-1 ]Is a probability distribution of each element p _i Representing the probability of the sample belonging to the ith class; y = [ y ₀ ,…,y _C-1 ]Is One-Hot representation of the sample label, y when the sample belongs to the ith category _i =1, otherwise y _i =0; c is a sample label.

And taking the object idle time probability of each sub-time period and the object characterization vector as input sample data of the multilayer perceptron.

In this embodiment, object characterization data is projected through a multilayer perceptron to generate object characterization vectors, a preset number of sub-period characterization data are obtained by dividing the object characterization data based on the push time characterization data, the sub-period characterization data is projected through the multilayer perceptron to generate sub-period characterization vectors, the object idle time probability of the sub-period corresponding to the sub-period characterization data is obtained according to the object characterization vectors and the sub-period characterization vectors, the object idle time probability of each sub-period and the object characterization vectors are used as input sample data of the multilayer perceptron to be trained to obtain a push time submodel, the push time submodel trained in different periods can provide an accurate successful push probability for each period, and accuracy of the push time submodel is improved.

The application also provides an application scenario, wherein the application scenario applies the information push model construction method, the method is applied to a scenario of pushing marketing activity information, and specifically, the application of the information push model construction method to the application scenario is as follows:

the computer equipment collects behavior characteristics from historical data of a user side and a marketing side, screens out mode entering and exiting characteristics according to means such as expert knowledge and characteristic importance calculation of a business department, and sends the mode entering characteristics to shallow Embedding characteristics of a neural network learning user and marketing contents.

The method comprises the steps of screening out characteristics to be adopted from a non-sensitive data pool of an internet financial institution, carrying out Spearman correlation coefficient analysis on all the characteristics, analyzing the correlation between specific characteristics and marketing touch and reach labels, eliminating invalid redundant characteristics, carrying out preliminary modeling by using simple logistic regression and a tree model, calculating the characteristic importance of each characteristic, sorting according to the corresponding characteristic importance, selecting the optimal tens of dimensions from the characteristics to carry out subsequent modeling, accelerating model training and reasoning speed.

After the in-mold feature is selected, an Embedding Embedding characterization is performed on the user and the marketing campaign. The user's Embedding representation mainly includes two sources, which are historical interaction information and feature statistical information thereof, respectively.

Its formulaic description is:

User _{history_emb} ＝RELU(a ₁ *EMB(sale ₁ )+a ₂ *EMB(sale_2)+…+a _n *EMB(sale _n ))，

equation 11

Where the RELU is an activation function,a _n the Attention weight can be directly defined by the user dwell time or obtained by automatic learning of an Attention mechanism, and only the sum of all Attention weights is 1, and the EMB () represents an Embedding function which can uniquely project discrete historical activity IDs into a high-dimensional dense space.

Then, the embedded representation obtained by the historical interaction is spliced with the basic statistical characteristics of the user, and the specific formula is as follows:

User_allemb＝Concat(User _{history_emb} user _ feature) equation 12

And then, processing by a multilayer perceptron to obtain a lower-dimensional embedded representation of the user:

User _emb = MLP (User _ ensemble) formula 13

Wherein the MLP is a multi-layer perceptron.

The marketing campaign side also has the same principle, and at this time, the embedded representations at the two ends are obtained for three downstream tasks to use. Since these tasks have strong correlation, three tasks are learned simultaneously using a multi-task learning technique.

Task one: since the user and the activity are different in nature, the multilayer perceptron is used for projection, and the representations (x and y) of the user and the activity are projected into the same low-dimensional space to obtain the representations, u (x) and v (y), in a new space. After both characterizations, the formula is used:

score = Sigmoid (u (x) · v (y). T), formula 14

T represents the transpose of the vector, and the final prediction score is the dot product of the two vectors, which is then integrated into the (0, 1) interval using Sigmoid.

Task 1 was then optimized with a Binary Cross Entropy (BCE) penalty.

L ₁ = BCE (score, label), equation 15

And a second task: and finely scoring the marketing reaching mode, and recommending the reaching mode for the marketing activity.

In the task, the marketing touch and reach modes (Torch _ way) are mainly divided into APP marketing, short message marketing, telephone marketing and the like, and the number of selectable modes is small, so that different touch and reach modes are scored on the basis of a Pair-wise model.

Also, we perform Embedding for different reach patterns:

Torchway _emb = EMB (Torch _ way), equation 16

The Embellding of the reach modes obtains preference scores of the user for different reach modes through a model structure similar to the task one.

Since the user is likely to accept multiple reach patterns, a Pair-wise model is employed to optimize

Wherein s is _i Score for positive sample, s _j The score of the negative sample is obtained by training to maximize the difference between the scores of the positive sample and the negative sample.

And a third task: the idle time probability of the user is learned, the user is prevented from being disturbed, and meanwhile the touch probability of the user is improved.

The idle time of a user often varies with objective factors. Task three sets of idle time of the user from 8:00-20:00, labeling by sections every two hours, using One-Hot coding, simultaneously splicing whether the time is a workday, a holiday, an activity promotion day and the like as the combined characteristics (Torch _ feature) of the time layer, and learning the embedded expression of the user time by using the following formula:

Time _emb = MLP (Torch _ feature), formula 18

Wherein the MLP is a multi-layer perceptron.

And after the embedded representation of the time is obtained, the embedded representation of the time is also sent to a model structure similar to the task one to be processed, and the preference degrees of the user on different touch times are obtained.

A multi-classification loss function is used in the training process, and the used labels comprise various labels for generating transactions after marketing, accessing services by users after marketing, no response after marketing, complaints after marketing and the like.

At the moment, vector dot products are not used as final scores, MLP is used for obtaining representation of six label dimensions, softmax is used for activation, and multi-class cross entropy loss is used for optimization.

The formula is as follows:

p＝Softmax(MLP(Concat(User _emb ,Time _emb ) ))) of formula 19

Wherein p = [ p ] ₀ ,…,p _C-1 ]Is a probability distribution of each element p _i Representing the probability of the sample belonging to the ith class; y = [ y) ₀ ,…,y _C-1 ]Is One-Hot representation of the sample label, y when the sample belongs to category i _i =1, otherwise y _i =0; c is a sample label.

In the optimization process, the Loss of different tasks is optimized in a unified way without adopting a weighted summation mode, fixed variables are adopted, and the tasks one, two and three are optimized separately in steps.

And finally, generating a personalized marketing strategy for each client by using the push model, reducing the disturbance to the client, and reaching the client in the most appropriate time in an optimal mode.

The information push model construction method includes the steps of obtaining historical characteristic data, obtaining characterization embedded data according to the historical characteristic data, wherein the characterization embedded data comprise object characterization data, push content characterization data, push mode characterization data and push time characterization data, conducting neural network training by using the object characterization data and the push content characterization data as training data to obtain a push content sub-model, conducting neural network training by using the object characterization data and the push mode characterization data as training data to obtain a push mode sub-model, conducting neural network training by using the object characterization data and the push time characterization data as training data to obtain a push time sub-model, and obtaining an information push model by parallelly splicing the push content sub-model, the push mode sub-model and the push time sub-model, wherein the information push model is used for obtaining corresponding information push content types, information push mode types and information push time according to object characterization data corresponding to an object to be pushed. Therefore, each submodel is obtained by respectively training the neural network with the pushed content characterization data, the pushed mode characterization data and the pushed time characterization data based on the object characterization data, and then all submodels are parallelly spliced to obtain the whole information pushing model.

It should be understood that, although the steps in the flowcharts related to the embodiments as described above are sequentially displayed as indicated by arrows, the steps are not necessarily performed sequentially as indicated by the arrows. The steps are not limited to being performed in the exact order illustrated and, unless explicitly stated herein, may be performed in other orders. Moreover, at least a part of the steps in the flowcharts related to the embodiments described above may include multiple steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, and the execution order of the steps or stages is not necessarily sequential, but may be rotated or alternated with other steps or at least a part of the steps or stages in other steps.

In one embodiment, as shown in fig. 9, an information push model building apparatus is provided, which may be a part of a computer device by using a software module or a hardware module, or a combination of the two modules, and specifically includes: a data acquisition module 902, a characterization embedding data generation module 904, and a module generation module 906, wherein:

a data obtaining module 902, configured to obtain historical feature data;

a representation embedded data generating module 904, configured to obtain representation embedded data according to the historical feature data, where the representation embedded data includes object representation data, push content representation data, push mode representation data, and push time representation data;

the model generation module 906 is configured to perform neural network training on the object representation data and the pushed content representation data as training data to obtain a pushed content sub-model; taking the object representation data and the push mode representation data as training data to carry out neural network training to obtain a push mode sub-model; taking the object representation data and the push time representation data as training data to perform neural network training to obtain a push time sub-model; and the information pushing model is used for obtaining the corresponding information pushing content category, the information pushing mode category and the information pushing time according to the object representation data corresponding to the object to be pushed.

The information push model construction device obtains the characteristic embedded data according to the historical characteristic data by obtaining the historical characteristic data, the characteristic embedded data comprises object characteristic data, push content characteristic data, push mode characteristic data and push time characteristic data, the object characteristic data and the push content characteristic data are used as training data to conduct neural network training to obtain a push content sub-model, the object characteristic data and the push mode characteristic data are used as training data to conduct neural network training to obtain a push mode sub-model, the object characteristic data and the push time characteristic data are used as training data to conduct neural network training to obtain a push time sub-model, and the information push model is obtained by parallelly splicing the push content sub-model, the push mode sub-model and the push time sub-model and is used for obtaining corresponding information push content types, information push mode types and information push time according to the object characteristic data corresponding to the object to be pushed. Therefore, each submodel is obtained by respectively training the neural network with the pushed content characterization data, the pushed mode characterization data and the pushed time characterization data based on the object characterization data, and then all submodels are parallelly spliced to obtain the whole information pushing model.

In one embodiment, the data acquisition module 902 is further configured to acquire initial historical feature data; performing relevance analysis on each dimension characteristic data in the initial historical characteristic data to obtain relevant dimension characteristic data; analyzing the importance of the related dimension characteristic data to determine the characteristic importance of the related dimension characteristic data; and sequencing the related dimension feature data according to the feature importance to generate historical feature data.

In one embodiment, the characterization embedded data generation module 904 is further configured to obtain target object feature data in the historical feature data; acquiring historical activity data corresponding to the characteristic data of the target object; normalizing the browsing and staying time of the objects in the historical activity data to obtain a time weight parameter; mapping each historical activity in the historical activity data into a corresponding historical activity vector, wherein the dimensions of the historical activity vectors corresponding to the historical activities are the same; and fusing the historical activity vectors corresponding to the historical activities according to the time weight parameters to obtain characterization embedded data, wherein the characterization embedded data is used for characterizing the corresponding relation of the push content, the push mode and the push time of each historical activity corresponding to the characteristic data of the target object.

In one embodiment, the characterization embedded data generating module 904 is further configured to fuse historical activity vectors corresponding to the historical activities according to the time weight parameter to obtain initial embedded characterization data; and processing the initial embedded characterization data by a multilayer perceptron to obtain characterization embedded data, wherein the dimensionality of the characterization embedded data is lower than that of the initial embedded characterization data.

In one embodiment, the model generation module 906 is further configured to project the object representation data through a multi-layer perceptron to generate an object representation vector; projecting the push content representation data through a multilayer perceptron to generate a push content representation vector, wherein the dimension of the push content representation vector is the same as that of the object representation vector; obtaining a push weight based on a dot product of a push content characterization vector and the object characterization vector; and training the object representation vector, the push content representation vector and the push weight as training sample data of the multilayer perceptron to obtain a push content sub-model.

In one embodiment, the model generation module 906 is further configured to project the object representation data through a multi-layer perceptron to generate an object representation vector; respectively generating an application push characterization vector, a short message push characterization vector and a telephone push characterization vector according to application push characterization data, short message push characterization data and telephone push characterization data in the push mode characterization data; obtaining an application push weight based on a dot product of the object representation vector and the application push representation vector; obtaining a short message pushing weight based on the dot product of the object representation vector and the short message pushing representation vector; obtaining a phone push weight based on a dot product of the object representation vector and the phone push representation vector; and obtaining a pushing mode submodel by taking the object representation vector, the application pushing weight, the short message pushing weight and the telephone pushing weight as model training sample data.

In one embodiment, the model generation module 906 is further configured to project the object representation data through a multi-layered perceptron to generate an object representation vector; dividing the presentation data based on the push time to obtain a preset number of sub-time period presentation data; projecting the sub-time period representation data through a multilayer perceptron to generate sub-time period representation vectors; obtaining the object idle time probability of the sub-time period corresponding to the sub-time period characterization data according to the object characterization vector and the sub-time period characterization vector; and training by taking the object idle time probability and the object representation vector of each sub-time period as input sample data of the multilayer perceptron to obtain a push time sub-model.

For specific limitations of the information push model building apparatus, reference may be made to the above limitations of the information push model building method, which is not described herein again. All or part of each module in the information push model building device can be realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in fig. 10. The computer device comprises a processor, a memory, a communication interface, a display screen and an input device which are connected through a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program to run on the non-volatile storage medium. The communication interface of the computer device is used for communicating with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. The computer program is executed by a processor to implement an information push model building method. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.

Those skilled in the art will appreciate that the architecture shown in fig. 10 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

In an embodiment, a computer device is further provided, which includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method embodiments when executing the computer program.

In an embodiment, a computer-readable storage medium is provided, in which a computer program is stored which, when being executed by a processor, carries out the steps of the above-mentioned method embodiments.

In one embodiment, a computer program product or computer program is provided that includes computer instructions stored in a computer-readable storage medium. The computer instructions are read by a processor of a computer device from a computer-readable storage medium, and the computer instructions are executed by the processor to cause the computer device to perform the steps in the above-mentioned method embodiments.

It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) referred to in the present application are information and data authorized by the user or sufficiently authorized by each party.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above may be implemented by hardware instructions of a computer program, which may be stored in a non-volatile computer-readable storage medium, and when executed, may include the processes of the embodiments of the methods described above. Any reference to memory, storage, database or other medium used in the embodiments provided herein can include at least one of non-volatile and volatile memory. Non-volatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, or the like. Volatile Memory can include Random Access Memory (RAM) or external cache Memory. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), among others.

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. An information push model construction method is characterized by comprising the following steps:

acquiring historical characteristic data;

obtaining characterization embedded data according to the historical characteristic data, wherein the characterization embedded data comprises object characterization data, pushed content characterization data, pushed mode characterization data and pushed time characterization data;

taking the object representation data and the push content representation data as training data to carry out neural network training to obtain a push content submodel;

and parallelly splicing to obtain an information pushing model based on the pushing content submodel, the pushing mode submodel and the pushing time submodel, wherein the information pushing model is used for obtaining the corresponding information pushing content category, the information pushing mode category and the information pushing time according to the object representation data corresponding to the object to be pushed.

2. The method of claim 1, wherein after obtaining the historical characterization data, further comprising:

acquiring initial historical characteristic data;

performing importance analysis on the related dimension characteristic data to determine the characteristic importance of the related dimension characteristic data;

and sequencing the related dimension characteristic data according to the characteristic importance to generate the historical characteristic data.

3. The method of claim 1, wherein deriving characterization embedding data from the historical characterization data comprises:

acquiring historical activity data corresponding to the target object characteristic data;

normalizing the browsing retention time of the objects in the historical activity data to obtain a time weight parameter;

and fusing the historical activity vectors corresponding to the historical activities according to the time weight parameters to obtain characterization embedded data, wherein the characterization embedded data are used for characterizing the corresponding relation of push contents, push modes and push time of the historical activities corresponding to the characteristic data of the target object.

4. The method according to claim 3, wherein the obtaining of the characterization embedding data by fusing the historical activity vectors corresponding to the historical activities according to the time weighting parameter comprises:

fusing according to the time weight parameters and the historical activity vectors corresponding to the historical activities to obtain initial embedded characterization data;

and processing the initial embedded characterization data by a multilayer perceptron to obtain characterization embedded data, wherein the dimensionality of the characterization embedded data is lower than that of the initial embedded characterization data.

5. The method of claim 1, wherein performing neural network training on the object characterizing data and the pushed content characterizing data as training data to obtain a pushed content sub-model comprises:

6. The method according to claim 1, wherein the obtaining of the sub-model of the push mode by performing neural network training on the object characterization data and the push mode characterization data as training data comprises:

obtaining a phone push weight based on a dot product of the object characterization vector and the phone push characterization vector;

7. The method of claim 1, wherein performing neural network training on the object characterization data and the push time characterization data as training data to obtain a push time submodel comprises:

dividing the push time characterization data to obtain a preset number of sub-time period characterization data;

obtaining the object idle time probability of the sub-time period corresponding to the sub-time period representation data according to the object representation vector and the sub-time period representation vector;

and training the object idle time probability of each sub-time period and the object characterization vector as input sample data of the multilayer perceptron to obtain a push time sub-model.

8. An information push model construction apparatus, the apparatus comprising:

the characterization embedded data generation module is used for obtaining characterization embedded data according to the historical characteristic data, wherein the characterization embedded data comprises object characterization data, push content characterization data, push mode characterization data and push time characterization data;

the model generation module is used for carrying out neural network training on the object representation data and the push content representation data as training data to obtain a push content sub-model; taking the object representation data and the push mode representation data as training data to carry out neural network training to obtain a push mode sub-model; taking the object representation data and the push time representation data as training data to carry out neural network training to obtain a push time submodel; and parallelly splicing the push content submodel, the push mode submodel and the push time submodel to obtain an information push model, wherein the information push model is used for obtaining the corresponding information push content category, the information push mode category and the information push time according to the object representation data corresponding to the object to be pushed.

9. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 7.

10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 7.