CN116029760A

CN116029760A - Message pushing method, device, computer equipment and storage medium

Info

Publication number: CN116029760A
Application number: CN202211655868.6A
Authority: CN
Inventors: 李洁; 张琛; 万化; 李云波; 李健
Original assignee: Shanghai Pudong Development Bank Co Ltd
Current assignee: Shanghai Pudong Development Bank Co Ltd
Priority date: 2022-12-22
Filing date: 2022-12-22
Publication date: 2023-04-28

Abstract

The application relates to a message pushing method, a message pushing device, computer equipment and a storage medium. The method comprises the following steps: acquiring an initial sample set, wherein the initial sample set comprises positive samples with up-to-standard activity participation, and negative samples with up-to-standard activity participation; converting the negative samples meeting the label conversion conditions into positive samples to obtain an intermediate sample set; inputting the middle sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels and individual factor labels for basic information containing the common factors and the individual factors in the samples, and obtaining a final sample set after labeling; and inputting the final sample set into a multi-label multi-classification model, predicting the target activity type corresponding to each sample by the multi-label multi-classification model, and pushing the activity message corresponding to the target activity type to each sample. The method can solve the problems that the individual characteristics are not obvious, training is sparse and the effective expansion cannot be achieved due to too few positive samples, and the like.

Description

Message pushing method, device, computer equipment and storage medium

Technical Field

The present invention relates to the field of active message pushing technologies, and in particular, to a message pushing method, device, computer equipment, and storage medium.

Background

In digital marketing, a batch of approximate or most likely transformed expanded populations can be found based on existing seed populations, thereby expanding the scope of the delivered or accurate marketing populations. In the prior art, activities or advertisement delivery is performed based on feature approximation and preference models, marking is performed based on marketing results to help ring guests, and business personnel select approximate crowds for expansion by analyzing the label distribution of response crowds and using experience when the ring guests; the preference model builds a classification model through trend score matching (PSM) or Decision Tree (DT) or comparative original algorithm (GBDT) and the like, and classifies non-seed users to identify potential approximate population.

The existing model method is characterized in that through business features and a method for reclassifying labels based on result data, due to the fact that common features among different activities are multiple, labels of the same seed user are repeatedly hit in different activities, meanwhile, when labels are marked for the seed user again, the label weight of the seed user is increased, the hit rate of the seeds is increased, and under a complex multi-activity pushing environment, the types of activity features cannot be distinguished due to the fact that the features among the activities are excessively shared, and therefore the accuracy of activity pushing is low.

Disclosure of Invention

In view of the foregoing, it is desirable to provide an activity pushing method, apparatus, computer device, computer readable storage medium, and computer program product that can improve the accuracy of the approximation of the customer premises.

In a first aspect, the present application provides a message pushing method. The method comprises the following steps:

acquiring an initial sample set, wherein the initial sample set comprises a plurality of samples, each sample comprises activity data of different activity types participated by a user, a first label marking the activity type of the activity data, a second label marking whether the activity participation degree corresponding to the sample meets the standard, and a plurality of third labels marking basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample;

converting the negative samples meeting the label conversion conditions into positive samples to obtain an intermediate sample set;

inputting the middle sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels for basic information containing the common factors in the samples, labeling individual factor labels for basic information containing the individual factors in the samples, and obtaining a final sample set after labeling; the commonality factor is the commonality characteristic among different activities; the personality factors are personality characteristics for distinguishing different activity types;

And inputting the final sample set into a multi-label multi-classification model, and predicting target activity types corresponding to all samples by the multi-label multi-classification model according to the first label, the second label, the third label, the commonality factor label and the personality factor label, and pushing activity messages corresponding to the target activity types to all samples.

In one embodiment, converting the negative samples that meet the label conversion condition into positive samples, obtaining an intermediate sample set, includes:

taking positive samples in the initial sample set as the initial positive sample set, and carrying out cluster analysis on label distribution of the initial positive sample set to obtain a trained classifier;

inputting each negative sample in the initial sample set to a classifier, and calculating the probability value of each negative sample marked as a positive sample through the classifier;

taking the negative sample with the probability value larger than the preset value as a target negative sample which meets the label conversion condition, and correcting a second label of the target negative sample to represent the activity participation degree of the target negative sample to reach the standard so as to convert the target negative sample into a positive sample;

adding a target negative sample into an initial positive sample set, clustering and analyzing label distribution of the initial positive sample set to obtain a new classifier, and completing a positive sample expansion process;

And iterating the positive sample expansion process for a plurality of times until the number of positive samples of the final initial positive sample set reaches the requirement, and stopping iterating to obtain an intermediate sample set.

In one embodiment, inputting each negative sample into a classifier, calculating a probability value for each negative sample labeled as a positive sample by the classifier, comprises:

inputting each negative sample into a classifier, predicting a first probability value of each negative sample marked by the classifier, predicting a second probability value of each positive sample marked by the classifier, and calculating a probability average value of the positive samples marked by the classifier;

the ratio of the square of the probability mean and the first probability value is determined as the probability value that the negative sample is marked as a positive sample.

In one embodiment, the final sample set is input into a multi-label multi-classification model, the multi-label multi-classification model predicts a target activity type corresponding to each sample according to a first label, a second label, a third label, a commonality factor label and a personality factor label, and pushes an activity message corresponding to the target activity type to each sample, including:

performing dumb variable conversion processing on the first label, the second label, the third label, the commonality factor label and the personality factor label of each sample to obtain a plurality of single-label classification sub-models;

And respectively predicting the activity types of each sample by adopting a plurality of single-label two-classification sub-models, classifying the activity types predicted by the plurality of single-label two-classification sub-models, and taking the activity type with the largest classification result as the target activity type corresponding to the sample.

In one embodiment, the method further comprises:

acquiring an activity message to be pushed;

and selecting a target sample with the same target activity type as the type of the activity message to be pushed in the initial sample set, and preferentially pushing the activity message to be pushed to the target sample.

In one embodiment, the method further comprises:

acquiring activity data of a user to be pushed, if the activity data of the user to be pushed belongs to an initial sample set, determining a target activity type corresponding to the user to be pushed in the initial sample set, and pushing an activity message corresponding to the target activity type to the user to be pushed.

In a second aspect, the present application further provides a message pushing device. The device comprises:

the system comprises an acquisition module, a storage module and a storage module, wherein the acquisition module is used for acquiring an initial sample set, the initial sample set comprises a plurality of samples, each sample comprises activity data of different activity types participated by a user, a first label for marking the activity type of the activity data, a second label for marking whether the activity participation degree corresponding to the sample meets the standard, and a plurality of third labels for marking basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample;

The sample expansion module is used for converting the negative samples meeting the label conversion conditions into positive samples to obtain an intermediate sample set;

the causal identification module is used for inputting the middle sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels for basic information containing the common factors in the samples, labeling individual factor labels for basic information containing the individual factors in the samples, and obtaining a final sample set after labeling; the commonality factor is the commonality characteristic among different activities; the personality factors are personality characteristics for distinguishing different activity types;

the multi-label multi-classification model predicts the target activity types corresponding to the samples according to the first label, the second label, the third label, the commonality factor label and the personality factor label, and pushes the activity messages corresponding to the target activity types to the samples.

In a third aspect, the present application also provides a computer device. The computer device comprises a memory storing a computer program and a processor which when executing the computer program performs the steps of:

In a fourth aspect, the present application also provides a computer-readable storage medium. The computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:

In a fifth aspect, the present application also provides a computer program product. The computer program product comprises a computer program which, when executed by a processor, implements the steps of:

The message pushing method, the device, the computer equipment and the storage medium convert the negative samples meeting the label conversion conditions into positive samples, so that a sufficient number of positive samples are obtained, and the problems that the individual characteristics are not obvious, the training is sparse and the effective expansion is not realized due to the fact that the number of the positive samples is too small are solved; the middle sample set after the positive sample expansion is used as the input of a causal model, the common factors and the personality factors are identified by combining the causal model through the enhancement and expansion processing of the positive sample, and the identified common factors are weakened, so that the personality factors among different activity types can be enhanced, the common factors and the personality factors among different types of activities can be effectively distinguished, the target activity types corresponding to the samples are predicted by adopting a multi-label multi-classification model, the problem that the personality factors are not obvious is solved, and the accuracy of the guests of different activity groups is improved.

Drawings

FIG. 1 is an application environment diagram of a message pushing method in one embodiment;

FIG. 2 is a flow diagram of a message pushing method in one embodiment;

FIG. 3 is a flow diagram of obtaining an intermediate sample set in one embodiment;

FIG. 4 is a schematic diagram of a training process of a semi-supervised framework model in another embodiment;

FIG. 5 is a causal pictorial intent in one embodiment;

FIG. 6 is a flowchart of a multi-label multi-classification model predicting a target activity type corresponding to each sample in one embodiment;

FIG. 7 is a block diagram of a message pushing device in one embodiment;

fig. 8 is an internal structural diagram of a computer device in one embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.

The message pushing method provided by the embodiment of the application can be applied to an application environment shown in fig. 1. The terminal 102 obtains an initial sample set, wherein the initial sample set comprises a plurality of samples, each sample comprises activity data of users participating in different activity types, a first label marking the activity type of the activity data, a second label marking whether the activity participation degree corresponding to the sample meets the standard, and a plurality of third labels marking basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample; converting the negative samples meeting the label conversion conditions into positive samples to obtain an intermediate sample set; inputting the middle sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels for basic information containing the common factors in the samples, labeling individual factor labels for basic information containing the individual factors in the samples, and obtaining a final sample set after labeling; the commonality factor is the commonality characteristic among different activities; the personality factors are personality characteristics for distinguishing different activity types; and inputting the final sample set into a multi-label multi-classification model, and predicting target activity types corresponding to all samples by the multi-label multi-classification model according to the first label, the second label, the third label, the commonality factor label and the personality factor label, and pushing activity messages corresponding to the target activity types to all samples. The terminal 102 communicates with the server 104 via a network. The terminal 102 may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, internet of things devices, portable wearable devices, and the internet of things devices may be smart televisions, etc. The portable wearable device may be a headset device or the like. The server 104 may be implemented as a stand-alone server or as a server cluster of multiple servers.

In one embodiment, as shown in fig. 2, a message pushing method is provided, and the method is applied to the terminal in fig. 1 for illustration, and includes the following steps:

step 202, an initial sample set is obtained, wherein the initial sample set comprises a plurality of samples, each sample comprises activity data of users participating in different activity types, a first label marking the activity type of the activity data, a second label marking whether the activity participation degree corresponding to the sample meets the standard, and a plurality of third labels marking basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample.

Wherein, one sample corresponds to one user, and the activity data comprises basic information of different activity types of the user participation history, activity types, client numbers, clicking or not, standard or not and other labels. The basic information may be information for distinguishing users and distinguishing different activities, for example, the basic information may be activity level, activity preference, and activity sensitivity, wherein the activity level may be determined by a ratio of the number of times the user participates in an activity to the total number of times the activity; the activity preferences may determine the user preferences for a certain type or types of activities by means of mathematical statistics.

Activity engagement refers to the response to a user when engaged in a certain type of activity. For example, the activity participation degree may be information such as a time period for which the user views the activity information, the number of times, whether to click on the activity link, whether to download the view, and whether to register the member. Whether the activity engagement meets the criterion may be determined from a plurality of dimensions, for example, whether the activity engagement meets the criterion may be determined by whether the user clicks on an activity link.

Several large categories of campaign data for the last year are extracted from the marketing system. Including activity type, click and reach mark. Basic information is extracted from the customer relationship system, etc. And processing the activity related data by combining information such as user multi-channel login and the like. And (3) carrying out data exploration and processing, deleting null values and abnormal values, unifying time of the season related data, and discretizing continuous variables. And identifying data distribution, and carrying out standardization and normalization processing. The statistical basic characteristics adopt the labeled data of the client relation system, the liveness is normalized according to the channel response times, the data such as the asset class and the like adopts labels, and the characteristic processing treatment is carried out on part of activity sensitive factors by adopting a gain model.

Optionally, the terminal acquires activity data of different users participating in different historical activity types, wherein the activity data comprises basic information of the users, one user corresponds to one sample, the activity types of a plurality of activity data in each sample are marked by a first label, whether the activity participation degree corresponding to each sample meets the standard or not is marked by a second label, and each basic information of the activity data in each sample is marked by a third label; and taking a sample with the activity participation degree which does not reach the standard as a negative sample, taking the sample with the activity participation degree which reaches the standard as a positive sample, and forming an initial sample set by the positive sample and the negative sample.

Step 204, converting the negative samples meeting the label conversion condition into positive samples, and obtaining an intermediate sample set.

The negative samples meeting the label conversion conditions are similar crowds of the positive samples and are marked as target negative samples. The intermediate sample set includes an original positive sample in the initial sample set, a target negative sample, and a negative sample remaining after the original negative sample amount in the initial sample has been removed from the target negative sample.

In practical application, a large number of users cannot reach the standard due to the fact that the users are not touched, the activity participation degree cannot reach the standard, and the real potential users cannot be identified due to the fact that the sample size meeting the standard of the activity participation degree is small. Generally, in order to ensure accuracy, different active groups are further split, and effective seeds (i.e., positive samples) are further diluted, so that effective expansion is difficult to achieve. Therefore, in order to solve the above-mentioned problem, in this embodiment, the semi-supervised model is adopted to identify reliable negative samples from the initial sample set, the reliable negative samples are approximate population of positive samples, and training is performed on the positive samples and the reliable negative samples for supervised learning until the positive sample amount in the initial sample set reaches the preset number. According to the embodiment, a real standard-reaching user is used as a positive sample, the positive sample is mixed into a reliable negative sample through an enhancement method for modeling, and enhancement processing and expansion are performed on the positive sample, so that the problems that individual characteristics are not obvious, training sparsity is caused by too few positive samples, and effective expansion is not achieved are solved.

Optionally, the terminal identifies a reliable negative sample, i.e. a target negative sample, from the initial sample set through the semi-supervised model, and trains on the reliable negative sample and the original positive sample until the positive sample amount in the initial sample set reaches a preset number, and an intermediate sample set including the original positive sample in the initial sample set, the target negative sample and the negative sample remaining after the original negative sample amount in the initial sample set is removed from the target negative sample is obtained.

Step 206, inputting the middle sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels for basic information containing the common factors in the samples, labeling individual factor labels for basic information containing the individual factors in the samples, and obtaining a labeled final sample set; the commonality factor is the commonality characteristic among different activities; the personality factors are personality characteristics that distinguish different activity types.

Wherein the activity data includes a plurality of basic information, and the plurality of basic information can be divided into a commonality factor and a personality factor. If the basic information of the user includes a plurality of common factors and a plurality of personality factors, the samples containing the common factors are labeled with common factor labels, for example, different common factor labels may be used in X1, X2, X3, and the like. Each sample of the final sample set includes a first label, a second label, a third label, a commonality factor label, and a personality factor label. The more the commonality factor, the more the commonality features between activities of different activity types, the more difficult it is to distinguish the user's preference for different activity types.

Because the common features among different activities are many, the labels of the same seed user are repeatedly hit in different activities, and meanwhile, when the labels are marked for the seed user again, the label weight of the seed user is increased, the hit rate of the seeds is increased, and in a complex multi-activity pushing environment, the types of activity features cannot be distinguished due to excessive sharing of the features among the activities, so that the accuracy of activity pushing is low. Therefore, in order to solve the above-mentioned problem, in this embodiment, historical activity data is used as an object, a causal model is built based on the existing model result data and service assumptions, causal relationship identification and effect analysis are applied, and robustness verification relationship is used to determine commonality factors and personality factors between different types of activities. The intermediate sample set after positive sample expansion by the semi-supervised model is used as the input of a causal model, the enhancement and expansion processing of the positive sample by the semi-supervised model is combined with the causal model to identify common factors and personality factors, and the identified common factors are weakened, so that the personality factors among different activity types can be enhanced. Therefore, aiming at complex and frequent activities, the common factors and the individual factors among different types of activities can be effectively distinguished by applying the causal model, the problem that the individual factors are not obvious is solved, and the accuracy of the ring guests of different active groups is improved.

Optionally, the terminal takes historical activity data as an object, establishes a causal model based on the result data of the existing model and service hypothesis, uses causal relation identification and effect analysis, uses robustness verification relation to determine commonality factors and personality factors among different types of activities, and obtains a trained causal model; the terminal inputs the intermediate sample set into a pre-trained causal model, identifies common factors and individual factors among samples through the causal model, marks common factor labels for basic information containing the common factors in the samples, marks individual factor labels for basic information containing the individual factors in the samples, and obtains a marked final sample set.

Step 208, inputting the final sample set into a multi-label multi-classification model, and predicting target activity types corresponding to the samples by the multi-label multi-classification model according to the first label, the second label, the third label, the commonality factor label and the personality factor label, and pushing activity messages corresponding to the target activity types to the samples.

The target activity type is the most preferred activity type of the user or the most interesting activity type of the user. After the above steps 204 and 206, on the one hand, the activity-sensitive spread data is obtained, on the other hand, the activity influencing factors are marked, and on the other hand, the sensitivity factors of each item are analyzed in step 208 to obtain the sensitivity characteristics of each item of activity, so as to make a reference for ordering when the item-dividing activity is put in.

It should be noted that: when the multi-label multi-classification model analyzes each individual factor, the common factors are manually reduced in a unified minimum weight mode, so that the weight of non-confounding factors is increased, the sensitivity of the individual factors is improved, and the labels of the same seed user are prevented from being hit repeatedly in different activities.

In the embodiment, the characteristic processing is carried out in an off-line mode and in a real-time mode, basic information is stored locally, and related common characteristics are generated by taking off-line as a main observation data related to the cause and effect. Real-time features (mainly behavior, path class data) are calculated from the marketing system by its own engine and stored in the Redis.

For example, one sample in the final sample set comprises activity data of a user a participating in an activity A, an activity B, an activity C and an activity D, a first label, a second label, a third label, a commonality factor label and a personality factor label, the final sample set is input into a multi-label multi-classification model, the multi-label multi-classification model outputs coding values corresponding to the labels, a threshold value is set for each label and is used as a basis for judging whether the labels belong to the labels, the multi-label multi-classification model outputs preference of the user to the activity, and the maximum value of the preference is taken as a target activity type corresponding to the sample.

Optionally, the terminal inputs the final sample set into a multi-label multi-classification model, and the multi-label multi-classification model predicts the preference degree of each sample for different activity types according to the first label, the second label, the third label, the commonality factor label and the personality factor label, takes the activity type corresponding to the maximum value of the preference degree as the target activity type corresponding to the sample, and pushes the activity message corresponding to the target activity type to each sample.

In one embodiment, after determining the most preferred target activity type for each sample, when pushing the activity message, the samples with the same target activity type as the pushing activity type are preferentially pushed. The method comprises the following specific steps:

step 1, obtaining an activity message to be pushed.

The to-be-pushed activity message may be a historical push activity message, or may be an activity message similar to the type of the historical push activity.

And 2, selecting a target sample with the same target activity type as the type of the activity message to be pushed from the initial sample set, and preferentially pushing the activity message to be pushed to the target sample.

Wherein, through the steps, the preference degree of each user in the user pool for different activity types and the most preferred activity type of each user, namely the target activity type, can be determined. When pushing the to-be-pushed active message, selecting a target sample with the same target activity type as the to-be-pushed active message from the initial sample set, and preferentially pushing the to-be-pushed active message to the target sample.

In one embodiment, after determining the most preferred target activity type of each sample, when analyzing the activity data of the user in the user pool, judging whether the user to be pushed belongs to the initial sample set, namely whether the user to be pushed belongs to the user pool, if the user to be pushed does not belong to the initial sample set, namely does not belong to the user pool, after the activity data of the user to be pushed reaches the preset requirement, then taking the user to be pushed into the user pool, and determining the target activity type corresponding to the user to be pushed according to the mode of the embodiment. If the user to be pushed belongs to the initial sample set, namely belongs to the user pool, determining a target activity type corresponding to the user to be pushed in the initial sample set, and pushing an activity message corresponding to the target activity type to the user to be pushed.

In this embodiment, when pushing an active message to be pushed, a target sample with the same type of target activity as that of the active message to be pushed is selected in the initial sample set, and the active message to be pushed is preferentially pushed to the target sample, so that the accuracy of the delivery of the active message can be improved.

In the message pushing method, negative samples meeting the label conversion conditions are converted into positive samples, so that a sufficient number of positive samples are obtained, and the problems that individual characteristics are not obvious, training is sparse and effective expansion cannot be achieved due to the fact that the number of the positive samples is too small are solved; the middle sample set after the positive sample expansion is used as the input of a causal model, the common factors and the personality factors are identified by combining the causal model through the enhancement and expansion processing of the positive sample, and the identified common factors are weakened, so that the personality factors among different activity types can be enhanced, the common factors and the personality factors among different types of activities can be effectively distinguished, the target activity types corresponding to the samples are predicted by adopting a multi-label multi-classification model, the problem that the personality factors are not obvious is solved, and the accuracy of the guests of different activity groups is improved.

In one embodiment, as shown in fig. 3, converting the negative samples satisfying the label conversion condition into positive samples, obtaining an intermediate sample set, includes:

step 302, taking positive samples in the initial sample set as the initial positive sample set, and clustering and analyzing label distribution of the initial positive sample set to obtain a trained classifier.

Optionally, the terminal uses positive samples in the initial sample set as an initial positive sample set, uses negative samples as an initial negative sample set, performs cluster analysis on label distribution of the initial positive sample set by adopting an expected maximum model, and trains a classifier, wherein the classifier can adopt a gradient lifting model.

Step 304, each negative sample in the initial sample set is input to a classifier, and a probability value of each negative sample marked as a positive sample is calculated through the classifier.

Optionally, the terminal inputs each negative sample in the initial negative sample set to a classifier, the classifier classifies the negative samples, and calculates a probability value that each negative sample is labeled as a positive sample.

In some embodiments, step 304 specifically includes the steps of:

step 1, inputting each negative sample into a classifier, predicting a first probability value of each negative sample marked by the classifier, and predicting a second probability value of each positive sample marked by the classifier, and calculating a probability mean value of the positive samples marked by the classifier.

Alternatively, the terminal will classify the positive and negative samples with a classifier, train using the positive sample values as targets, predict the first probability value P (y _x = 1|x); predicting a second probability value P (y) of positive samples in the initial sample set being labeled using a classifier _x = 1|y = 1|x) and based on the second probability value, calculating the probability mean P (y _x ＝1|y＝1)。

And 2, determining the ratio of the square of the probability mean value and the first probability value as the probability value of the negative sample marked as the positive sample.

Alternatively, the probability value that a negative sample is labeled as a positive sample is expressed by a mathematical formula:

and 306, taking the negative sample with the probability value larger than the preset value as a target negative sample which meets the label conversion condition, and correcting a second label of the target negative sample to represent the activity participation degree of the target negative sample to reach the standard so as to convert the target negative sample into a positive sample.

The terminal, as shown in fig. 4, arranges probability values of negative samples marked as positive samples in order from large to small, selects negative samples with probability values greater than a preset value as target negative samples meeting the label conversion condition, and adds the target negative samples to the initial positive sample set.

Step 308, adding the target negative sample into the initial positive sample set, and clustering and analyzing the label distribution of the initial positive sample set to obtain a new classifier, thereby completing a positive sample expansion process.

Optionally, the terminal rejects the target negative sample from the initial negative sample set, brings the target negative sample into the initial positive sample set, continues training and classifying according to the method of step 302 to obtain a new classifier, and simultaneously completes a positive sample expansion process through one iteration process.

And step 310, iterating the positive sample expansion process for a plurality of times until the number of positive samples of the final initial positive sample set reaches the requirement, and stopping iterating to obtain an intermediate sample set.

In the embodiment, the classifier obtained by training the positive sample set identifies the negative samples meeting the label conversion condition in the negative sample set, namely the approximate population, and the approximate population is brought into the positive sample set, so that the number of the positive samples is expanded, and the problems that the individual characteristics are not obvious, the training is sparse and the effective expansion is not realized due to the fact that the number of the positive samples is too small are solved.

In one embodiment, a causal model is used to distinguish between activity commonality factors and personality factors of activity types in a limited sample. Through multiple activity sedimentation and model duplication, the understanding and demonstration of the relation between partial guest groups and response are already sedimented, so that the method is firstly utilized

The a priori knowledge models causal relationships. By adopting a causal inference graph model method, causal learning cannot exhaust confounding factors, a certain business experience input and a variable relation precipitation of historical analysis demonstration are needed, and a causal graph is constructed based on the predefined commonality factors, tool variables and result variables. The commonality factor of the present embodiment is a factor that acts on both the intervention (activity) and the result (response), for example, characteristic variables such as guest classification and activity level. The tool variable is an active push and the result variable is whether to respond.

The causal graph established is shown in fig. 5. The statistical significance of the above effect estimates was checked by permutation tests. Wherein W represents a commonality factor, Z is a tool variable, Y is a result variable (up to standard), X is an effect modification variable, and V is an intervention variable (activity).

Based on the initial design of the causal graph, a check of back door criteria, front door criteria and IV tool variables is performed. Dividing the initial sample set into two groups, wherein one group is an intervention group affected by intervention, and the other group is a control group not affected by the same intervention; respectively carrying out difference (subtraction) for two times before and after the intervention to obtain two groups of difference values, wherein the difference values represent the relative relation between the intervention group and the control group before and after the intervention; and carrying out second difference on the two groups of differences, thereby eliminating the difference between the intervention group and the control group, and finally obtaining the average intervention effect. The intervention effect of activity or not is estimated preferentially here. The robustness check is carried out by a method of adding the confounding factors, and statistical information and behavior statistical information are added during detection, such as the login frequency of a mobile phone bank, whether wind evaluation is carried out, and the like. And (5) circulating the process, and adjusting the parameter relation until the inspection requirement is met.

The identified confounding factors are marked X1, X2, xn.. The common factor type identified by the causal model comprises the categories of liveness, activity sensitivity, product preference and the like, and is used as an activity common influence factor, and when the multi-label multi-classification model distinguishes activities, the factor characteristics are subjected to weight reduction treatment to enhance other factors.

In one embodiment, as shown in fig. 6, the final sample set is input into a multi-label multi-classification model, the multi-label multi-classification model predicts a target activity type corresponding to each sample according to a first label, a second label, a third label, a commonality factor label and a personality factor label, and pushes an activity message corresponding to the target activity type to each sample, including:

step 602, performing dumb variable conversion processing on the first label, the second label, the third label, the commonality factor label and the personality factor label of each sample to obtain a plurality of single-label molecular models.

In which there is a relationship in which one type of characteristic attribute is parallel to each other in a sample, meaning cannot be given simply in numerical values or characters, and this can be solved by constructing dummy variables. For example, a feature containing three factors may convert it into a vector of three columns, each with only 0-1, such a vector being a dummy variable. The dummy variable conversion process is to convert the multi-classification variable into binary variable.

And combining the two-class data in the K classes by the single-label two-class sub-model, training a model by using the combined data, thereby generating K (K-1)/2 classifiers, fusing the results of the classifiers, and outputting a final predicted result value by using a majority voting mode by using the predicted result of the classifier. The single-label two-classification sub-model adopts an ML-Knn classification algorithm to predict.

Optionally, the terminal converts the first label, the second label, the third label, the commonality factor label and the personality factor label of each sample into binary variables, combines the two-by-two data in multiple categories, and then trains multiple single-label two-category sub-models by using the multiple combined data.

Step 604, predicting the activity types of each sample by using a plurality of single-label two-classification sub-models, classifying the activity types predicted by the plurality of single-label two-classification sub-models, and taking the activity type with the highest classification result as the target activity type corresponding to the sample.

The method comprises the steps that a sample comprises activity data of users participating in different activities, activity types of the sample are respectively predicted by adopting a plurality of single-label two-classification sub-models, the activity types predicted by the plurality of single-label two-classification sub-models are classified by using a majority voting mode, the number of predicted results corresponding to each activity type can be obtained, and the reference of the sequencing when the classified activities are put in is determined according to the sequencing of the number of predicted results corresponding to each activity type. For example, a sample a user participates in activity data of activity a, activity B, activity C and activity D, predicts activity types through a plurality of single-tag two-class sub-models, classifies the activity types predicted by the plurality of single-tag two-class sub-models, and ranks the activity types according to the classification result from high to low, so as to obtain ranks of activity C, activity a, activity D and activity B, namely, a target activity type of sample a is activity C, and an activity put rank of sample a is activity C, activity a, activity D and activity B.

In this embodiment, the first label, the second label, the third label, the commonality factor label and the personality factor label of each sample are subjected to dumb variable conversion processing to obtain a plurality of single-label two-classification sub-models, the activity types of each sample are respectively predicted by adopting the plurality of single-label two-classification sub-models, the target activity types corresponding to the samples are determined according to the prediction results, and the activity message corresponding to the target activity type is preferentially recommended when the activity type quantitative recommendation is distinguished, so that the precision of the ring passenger is improved.

In one embodiment, the present embodiment provides the most detailed steps of a message pushing method, which specifically includes the following steps:

step 1, an initial sample set is obtained, wherein the initial sample set comprises a plurality of samples, each sample comprises activity data of users participating in different activity types, a first label marking the activity type of the activity data, a second label marking whether the activity participation degree corresponding to the sample meets the standard, and a plurality of third labels marking basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample.

And 2, taking positive samples in the initial sample set as the initial positive sample set, and carrying out cluster analysis on label distribution of the initial positive sample set to obtain a trained classifier.

And 3, inputting each negative sample into a classifier, predicting a first probability value of each negative sample marked by the classifier, and predicting a second probability value of each positive sample marked by the classifier, and calculating a probability mean value of the positive samples marked by the classifier.

And 4, determining the ratio of the square of the probability mean value and the first probability value as the probability value of the negative sample marked as the positive sample.

And 5, taking the negative sample with the probability value larger than the preset value as a target negative sample which meets the label conversion condition, and correcting a second label of the target negative sample to represent the activity participation degree of the target negative sample to reach the standard so as to convert the target negative sample into a positive sample.

And 6, adding the target negative sample into the initial positive sample set, and clustering and analyzing the label distribution of the initial positive sample set to obtain a new classifier, thereby completing a positive sample expansion process.

And 7, iterating the positive sample expansion process for a plurality of times until the number of positive samples of the final initial positive sample set reaches the requirement, and stopping iterating to obtain an intermediate sample set.

Step 8, inputting the middle sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels for basic information containing the common factors in the samples, labeling individual factor labels for basic information containing the individual factors in the samples, and obtaining a final sample set after labeling; the commonality factor is the commonality characteristic among different activities; the personality factors are personality characteristics that distinguish different activity types.

And 9, performing dumb variable conversion processing on the first label, the second label, the third label, the commonality factor label and the personality factor label of each sample to obtain a plurality of single-label molecular models.

And 10, respectively predicting the activity types of each sample by adopting a plurality of single-label two-classification sub-models, classifying the activity types predicted by the plurality of single-label two-classification sub-models, and taking the activity type with the largest classification result as the target activity type corresponding to the sample.

In this embodiment, a semi-supervised model framework is applied to realize effective spreading under the condition that the positive sample size is relatively small (response cannot be reached), and the return on investment is not reduced; the causal model identifies activity commonality factors, on one hand, activity expansion is not affected, and on the other hand, a characteristic basis is provided for the recommendation of the subsequent classification priority; the multi-classification model is put on and put off, and the specific scenes of the multi-activity recommended ring guests are sorted and distinguished, so that the marketing effect is maximized. Based on a business analysis framework, the influence of interference factors is reduced by adopting a joint modeling multi-layer model, and business decisions can be effectively guided.

It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.

Based on the same inventive concept, the embodiment of the application also provides a message pushing device for realizing the above related message pushing method. The implementation of the solution provided by the device is similar to the implementation described in the above method, so the specific limitation in one or more embodiments of the message pushing device provided below may refer to the limitation of the message pushing method hereinabove, and will not be repeated herein.

In one embodiment, as shown in fig. 7, there is provided a message pushing apparatus, including: an acquisition module 100, a sample expansion module 200, a cause and effect identification module 300, and a multi-classification module 400, wherein:

the obtaining module 100 is configured to obtain an initial sample set, where the initial sample set includes a plurality of samples, each sample includes activity data of a user participating in different activity types, a first tag for labeling an activity type to which the activity data belongs, a second tag for labeling whether an activity participation degree corresponding to the sample meets standards, and a plurality of third tags for labeling basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample.

The sample expansion module 200 is configured to convert the negative samples that meet the label conversion condition into positive samples, and obtain an intermediate sample set.

The causal identification module 300 is configured to input the intermediate sample set into a pre-trained causal model, identify commonality factors and personality factors between samples through the causal model, label the fundamental information containing the commonality factors in the samples with commonality factor labels, label the fundamental information containing the personality factors in the samples with personality factor labels, and obtain a final sample set after labeling; the commonality factor is the commonality characteristic among different activities; the personality factors are personality characteristics that distinguish different activity types.

The multi-classification module 400 is configured to input the final sample set into a multi-label multi-classification model, and the multi-label multi-classification model predicts a target activity type corresponding to each sample according to the first label, the second label, the third label, the commonality factor label and the personality factor label, and pushes an activity message corresponding to the target activity type to each sample.

In one embodiment, the sample expansion module 200 is further configured to use the positive samples in the initial sample set as the initial positive sample set, and perform cluster analysis on the label distribution of the initial positive sample set to obtain a trained classifier;

In one embodiment, the sample expansion module 200 is further configured to input each negative sample to a classifier, predict a first probability value of each negative sample being labeled by the classifier, and predict a second probability value of each positive sample being labeled, and calculate a probability average of the positive samples being labeled;

In one embodiment, the multi-classification module 400 is further configured to perform a dummy variable conversion process on the first label, the second label, the third label, the commonality factor label, and the personality factor label of each sample, so as to obtain a plurality of single-label classification sub-models;

In one embodiment, the multi-classification module 400 is further configured to obtain an activity message to be pushed;

In one embodiment, the multi-classification module 400 is further configured to obtain activity data of a user to be pushed, determine a target activity type corresponding to the user to be pushed in the initial sample set if the activity data of the user to be pushed belongs to the initial sample set, and push an activity message corresponding to the target activity type to the user to be pushed.

The various modules in the message pushing device described above may be implemented in whole or in part by software, hardware, and combinations thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a terminal, and the internal structure thereof may be as shown in fig. 8. The computer device includes a processor, a memory, an input/output interface, a communication interface, a display unit, and an input means. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input/output interface. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile 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 operation of the operating system and computer programs in the non-volatile storage media. The input/output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless mode 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 a message pushing method. The display unit of the computer device is used for forming a visual picture, and can be a display screen, a projection device or a virtual reality imaging device. The display screen 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, can also be a key, a track ball or a touch pad arranged on the shell of the computer equipment, and can also be an external keyboard, a touch pad or a mouse and the like.

It will be appreciated by those skilled in the art that the structure shown in fig. 8 is merely a block diagram of some of the structures associated with the present application and is not limiting of the computer device to which the present application may be applied, and that a particular computer device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is provided comprising a memory and a processor, the memory having stored therein a computer program, the processor when executing the computer program performing the steps of:

In one embodiment, the processor when executing the computer program further performs the steps of:

acquiring an activity message to be pushed;

In one embodiment, a computer readable storage medium is provided having a computer program stored thereon, which when executed by a processor, performs the steps of:

In one embodiment, the computer program when executed by the processor further performs the steps of: taking positive samples in the initial sample set as the initial positive sample set, and carrying out cluster analysis on label distribution of the initial positive sample set to obtain a trained classifier;

In one embodiment, the computer program when executed by the processor further performs the steps of:

acquiring an activity message to be pushed;

In one embodiment, a computer program product is provided comprising a computer program which, when executed by a processor, performs the steps of:

acquiring an activity message to be pushed;

It should be noted that, the user information (including, but not limited to, user equipment information, user personal information, etc.) and the data (including, but not limited to, data for analysis, stored data, presented data, etc.) referred to in the present application are information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data are required to comply with the related laws and regulations and standards of the related countries and regions.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magnetic random access Memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (Phase Change Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like. The databases referred to in the various embodiments provided herein may include at least one of relational databases and non-relational databases. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processors referred to in the embodiments provided herein may be general purpose processors, central processing units, graphics processors, digital signal processors, programmable logic units, quantum computing-based data processing logic units, etc., without being limited thereto.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The above examples only represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the present application. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application shall be subject to the appended claims.

Claims

1. A message pushing method, the method comprising:

acquiring an initial sample set, wherein the initial sample set comprises a plurality of samples, each sample comprises activity data of users participating in different activity types, a first label for marking the activity type of the activity data, a second label for marking whether the activity participation degree corresponding to the sample meets the standard, and a plurality of third labels for marking basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample;

inputting the intermediate sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels for basic information containing the common factors in the samples, labeling individual factor labels for basic information containing the individual factors in the samples, and obtaining a final sample set after labeling; the commonality factors are commonality characteristics among different activities; the personality factors are personality characteristics for distinguishing different activity types;

2. The method of claim 1, wherein converting negative samples that meet a label conversion condition to positive samples, obtaining an intermediate sample set, comprises:

taking positive samples in an initial sample set as the initial positive sample set, and carrying out cluster analysis on label distribution of the initial positive sample set to obtain a trained classifier;

Inputting each negative sample in the initial sample set into the classifier, and calculating a probability value of each negative sample marked as a positive sample through the classifier;

adding a target negative sample into the initial positive sample set, and carrying out cluster analysis on label distribution of the initial positive sample set to obtain a new classifier, so as to complete a positive sample expansion process;

3. The method of claim 2, wherein the inputting each negative sample into the classifier, calculating by the classifier a probability value for each negative sample labeled as a positive sample, comprises:

inputting each negative sample into the classifier, predicting a first probability value of each negative sample marked by the classifier, predicting a second probability value of each positive sample marked by the classifier, and calculating a probability mean value of the positive samples marked by the classifier;

And determining the ratio of the square of the probability mean value and the first probability value as a probability value that the negative sample is marked as a positive sample.

4. The method according to claim 1, wherein the inputting the final sample set into a multi-tag multi-class model, the multi-tag multi-class model predicting a target activity type corresponding to each sample according to the first tag, the second tag, the third tag, the commonality factor tag, and the personality factor tag, pushing an activity message corresponding to the target activity type to each sample, includes:

5. The method according to claim 1, wherein the method further comprises:

acquiring an activity message to be pushed;

and selecting a target sample with the same target activity type as the type of the activity message to be pushed from the initial sample set, and preferentially pushing the activity message to be pushed to the target sample.

6. The method according to claim 1, wherein the method further comprises:

7. A message pushing device, the device comprising:

the system comprises an acquisition module, a storage module and a storage module, wherein the acquisition module is used for acquiring an initial sample set, the initial sample set comprises a plurality of samples, each sample comprises activity data of different activity types participated by a user, a first label for marking the activity type of the activity data, a second label for marking whether the activity participation degree corresponding to the sample meets the standard or not, and a plurality of third labels for marking basic information in the activity data; taking a sample with the up-to-standard activity participation degree as a positive sample, and taking a sample with the up-to-standard activity participation degree as a negative sample;

the causal identification module is used for inputting the intermediate sample set into a pre-trained causal model, identifying common factors and individual factors among samples through the causal model, labeling common factor labels for basic information containing the common factors in the samples, labeling individual factor labels for basic information containing the individual factors in the samples, and obtaining a final sample set after labeling; the commonality factors are commonality characteristics among different activities; the personality factors are personality characteristics for distinguishing different activity types;

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 6 when the computer program is executed.

9. 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 the steps of the method of any of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 6.