CA3143808A1 - Event promoting method, device, computer apparatus, and storage medium - Google Patents

Abstract

The present invention relates to a method, a device, a computer apparatus, and a storage medium for event promotions. The described method includes: acquiring user information of a target user to be promoted with events, and determining user types of the target user based on the user information, wherein user types include active user type, inactive user type, and new user type; acquiring event promoting lists corresponding to a target user based promoting algorithm of user types; and selecting at least one event from the event promoting lists to be promoted to the target user. The present invention improves event promoting accuracy to effectively prevent internet resource wastes.

Description

EVENT PROMOTING METHOD, DEVICE, COMPUTER APPARATUS, AND STORAGE
MEDIUM
Technical Field [0001] The present invention relates to the field of intelligent promoting technologies, in particular, to a method, a device, a computer apparatus, and a storage medium for event promoting.
Background

[0002] In recent years, with the rapid development of internet technologies, more online data volume is accessible to users. However, with limited time to view different websites or Apps, sometimes the huge number of displayed events may become redundant, requiring more time to distinguish by users. To solve the problem, the event promotion method is currently adopted to select events that may interest users from a massive number of events to achieve accurate recommendations.

[0003] The major event promoting method currently is to promote based on collaborative filtering algorithm or conditional strategies.

[0004] The collaborative filtering algorithm generally can be classified into two types. One type is a user-based collaborative filtering algorithm targeting on a statistical calculation of user preference to events, wherein a user activity matrix is constructed with preference activity to all promotion events of each user.
One piece of data is a preference activity vector, using the similarity calculation method to find user groups with similar user activities. The preferred promotion events are internally promoted to users amongst the user group and the described method permits crossing-field event promotions.
However, in practical scenarios, obtaining similar users is complicated, leading to low recommendation accuracy. The other is an event-based collaborative filtering algorithm that is similar to the previous method while targeting events.
By calculating event similarities, events that have not been viewed are promoted based on user activity data.
The described method allows to obtain event similarities within a certain time interval and permits higher accuracy and stability. However, the basic principles and practical scenarios narrow down the range of events to be promoted. Besides low diversity of promoting contents, the updates, processing, and other operations may significantly influence matrix calculation. The problems of synchronizing service data and background databases should be solved. The aforementioned methods are not able to solve the problem of cold start, and cannot provide proper recommendation lists to no user history users.

[0005] The conditional promoting algorithm is to set conditional strategies targeting certain user groups by service operational personnel based on operation targets and platform user activities. The described Date recue/ date received 2021-12-23 method is dependent on the experience of service personnel and invested company resources, leading to low recommendation accuracy with high subjectivity.

[0006] In terms of servers, low activity recommendation accuracy and event diversity would not only cause unnecessary resource wastes of websites and App page displays, but also increase unnecessary interactions between users and servers, causing additional pressure over the server and internet resource wastes.
Summary

[0007] Herein, targeting on aforementioned technical problems, a more diversified and accurate event promoting method, device, computer apparatus, and storage medium are provided, to effectively prevent potential intemet resource wastes.

[0008] An event promoting method, comprises:

[0009] acquiring user information of a target user to be promoted with events, and determining user types of the target user based on the user information, wherein user types include active user type, inactive user type, and new user type;

[0010] acquiring event promoting lists corresponding to the target user based promoting algorithm of user types, wherein the promoting algorithm for the active user type is a promoting model based on gradient boosting decision tree and logistic regression; the promoting algorithm for the inactive user type is a user similarity based promoting algorithm; and the promoting algorithm for the new user type is an event popularity-based promoting algorithm; and

[0011] selecting at least one event from the event promoting lists to be promoted to the target user.

[0012] In one of the embodiments, the described determination of user types of the target user based on the user information, includes:

[0013] determining the user type of the described target user as the active user type when the user information is determined to include user property data and user activity data;

[0014] determining the user type of the described target user as the inactive user type when the user information is determined to include user property data but no user activity data; and

[0015] determining the user type of the described target user as the new user type when the user information is determined to not include user property data or user activity data.

[0016] In one of the embodiments, the described acquisition of event promoting lists corresponding to the target user based promoting algorithm of user types, includes:

[0017] acquiring candidate event set corresponding to the described target user when the user type of the target user is determined as the active user type or new user type, selecting at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the Date recue/ date received 2021-12-23 target user, and obtaining an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions; and

[0018] acquiring user property data of each active user from the currently active user set when the user type of the described target user is the inactive target user types, calculating the similarity between the described target user and each active user based the described target user property data and the user property data of each active user to obtain an event promoting list for an active user with the highest similarity with the described target user, and assigning the described event promoting list as the event promoting list for the described target user, wherein the described active users are classified as the active user type.

[0019] In one of the embodiments, the described selection of at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user and attainment of an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions, include:

[0020] when the user type of the described target user is the active user type, calculating the preference of each event type by the described target user, and assigning preference of individual event types as a section of the described user property data of the described target user;

[0021] pre-processing the user property data for the described target user and event property data of individual events in the candidate event set;

[0022] inputting the processed user property data of the described target user and the event property data of individual events in the candidate event set into a trained promoting model based on gradient boosting decision tree and logistic regression for prediction, selecting individual candidate events with a predicted value greater than a pre-set threshold, and sorting the individual candidate events with a predicted value greater than a pre-set threshold to obtain an event promoting list corresponding to the target user; wherein

[0023] preferably, the described activity property data includes event Chinese name, and the described pre-processing of the user property data for the described target user and event property data of individual events in the candidate event set includes:

[0024] segmenting event Chinese names of the described candidate events in the described candidate sets to obtain segment results for Chinese names of each described candidate event, and converting the Chinese name segment results of each described candidate event into values via word2vec model, so as for acquiring value-based vectors corresponding to Chinese names of each described candidate event; and

[0025] performing missing-refill and vectorization of all data sections besides the described event Chinese names in the described user property data of the target user and the described candidate event property data in the described candidate set.

[0026] In one of the embodiments, before obtaining an event promoting list for an active user with the highest similarity with the described target user, the described method includes:

Date recue/ date received 2021-12-23

[0027] acquiring a candidate event set corresponding to the active user with the highest similarity with the described target user; and

[0028] selecting at least one candidate events satisfying pre-set conditions from the described candidate event set via trained promoting model based on gradient boosting decision tree and logistic regression for prediction, to obtain an event promoting list for an active user with the highest similarity with the described target user based on each described candidate event satisfying pre-set conditions.

[0029] In one of the embodiments, the described selection of at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user and attainment of an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions, include

[0030] acquiring activity popularity data of each candidate event in the described candidate event set within a pre-set time interval when the target user is classified as the new user type;

[0031] calculating current popularity predicted values from the event popularity data of each candidate event in the pre-set time interval via Newton's law of cooling, so as for selecting candidate events satisfying pre-set conditions based on the described current popularity score; and

[0032] sorting candidate events satisfying pre-set conditions based on the described current popularity score, to obtain an event promoting list corresponding to the described target user.

[0033] In one of embodiment, the described promoting model based on gradient boosting decision tree and logistic regression is trained with training data sets including multiple training data combinations and tag information corresponding to each training data combination, wherein each described training combination includes pre-processed user property data of sample users and event property data of sample events, and the described sample users are classified as the active user type;

[0034] the described user property data includes preference scores for each event type by the described sample users, wherein the described preference scores are obtained by calculation based on user activity data of the described sample users;

[0035] the described pre-processed event property data includes value-based vectors, and the value-based vectors for the described sample event Chinese names are obtained by segmenting the event Chinese names of the described sample events and converting from the described event Chinese name segmenting results via word2vec model;

[0036] preferably, after acquiring an event promoting list for the described target user from the promoting algorithm corresponding to the described user type, the method further includes: selecting each valid candidate event from the described event promoting list, so as for obtaining a valid event promoting list based on each valid candidate event; and Date recue/ date received 2021-12-23

[0037] the described selection of at least one event from the event promoting lists to be promoted to the target user includes selecting at least one event from the described valid event promoting lists to be promoted to the target user.

[0038] An event promoting device, comprises:

[0039] a primary acquisition module, configured to acquire user information of a target user to be promoted with events, and determine the user types of the target user based on the user information, wherein user types include active user type, inactive user type, and new user type;

[0040] a secondary acquisition module, configured to acquire event promoting lists corresponding to the target user based promoting algorithm of user types, wherein the promoting algorithm for the active user type is a promoting model based on gradient boosting decision tree and logistic regression, the promoting algorithm for the inactive user type is a user similarity based promoting algorithm, and the promoting algorithm for the new user type is an event popularity-based promoting algorithm; and

[0041] an event promoting module, configured to select at least one event from the event promoting lists to be promoted to the target user.

[0042] A computer apparatus comprises a memory unit, a processor, and computer programs stored in the memory unit executable on the processor, wherein the procedures of any aforementioned method embodiments are performed when the described processor executes the described computer programs.

[0043] A readable computer storage medium with computer programs stored, wherein the procedures of any aforementioned method embodiments are performed when the described computer programs are executed on the described processor.

[0044] In the aforementioned embodiments, the server determines the user type of a target user according to user information of the event promotion target user, acquires a corresponding event promoting list according to a promoting algorithm for the user type, and promotes events to the target user based on the described event promoting list. In particular, different promoting algorithms are applied to users of different user types. By promoting models based on gradient boosting decision tree and logistic regression, the performance of promoting models is greatly improved. The remaining users are covered by promoting algorithm based on user similarity and event popularity. Each user can be applied with the appropriate algorithm for corresponding event promoting, to cover the majority of cases in event promotion and improve event promotion accuracy and diversity, preventing internet resource wastes.
Brief descriptions of the drawings

[0045] Fig. 1 is an application environment diagram of event promoting method in an embodiment;

[0046] Fig. 2 is a flow diagram of event promoting method in an embodiment;

[0047] Fig. 3 is an operation structure diagram of event promoting method in an embodiment;

Date recue/ date received 2021-12-23

[0048] Fig. 4 is a structural diagram of the promoting model based on gradient boosting decision tree and logistic regression in an embodiment;

[0049] Fig. 5 is a flow diagram of procedures for acquiring active user event promoting list in an embodiment;

[0050] Fig. 6 is a flow diagram of procedures for processing event Chinese names in an embodiment;

[0051] Fig. 7 is a flow diagram of procedures for acquiring inactive user event promoting list in an embodiment;

[0052] Fig. 8 is a flow diagram of procedures for acquiring new user event promoting list in an embodiment;

[0053] Fig. 9 is a structure diagram of event promoting device in an embodiment; and

[0054] Fig. 10 is an internal structure diagram of computer apparatus in an embodiment.
Detailed descriptions

[0055] In order to make the objective, the technical protocol, and the advantages of the present invention clearer, the present invention will be explained further in detail precisely below with references to the accompanying drawings. Obviously, the embodiments described below are only a portion of embodiments of the present invention and cannot represent all possible embodiments. Based on the embodiments in the present invention, the other applications by those skilled in the art without any creative works are falling within the scope of the present invention.
Embodiment one

[0056] In the present embodiment, the event promoting method provided in the present invention can be implemented in the application environment shown in Fig. 1. In particular, terminal 102 is connected with the server 104 via a network for communications. The server 104 acquires user information of the target user for event promoting; determines the user type of the target user according to the user information of the target user, wherein the user types includes active user type, inactive user types, and new user types;
and acquires event promoting list for the target user according to promoting algorithm for the user type, then selecting at least one event from the event promoting list to be promoted to the target user viewing web pages or an App via terminal 102. In particular, terminal 102 can be, but is not limited to, a personal computer, a laptop, a smart phone, a tablet, or portable wearable devices. The server 104 can be an independent server or a server cluster composed of multiple servers.

[0057] The procedures of an event promoting method provided in the present embodiment is shown in Fig.
2. The following descriptions are based on implementing the described method in the server shown in Fig.
1.

Date recue/ date received 2021-12-23

[0058] Step 202, acquiring user information of a target user to be promoted with events, and determining user types of the target user based on the user information, wherein user types include active user type, inactive user type, and new user type;

[0059] In particular, at least one of aforementioned target users is involved, wherein the user information of the target user includes user identity information (such as user serial number). In an embodiment, the user information can further include user type information to directly determine the user type of the target user from the user type information, so as for rapidly acquiring a promoting algorithm for event promoting list corresponding to the target user in the following steps.

[0060] In another embodiment, the user information may not directly include user type information. The user type is not determined from the user type information. The user type of the target user is determined from the information content included in the user information or detailed data types. For example, users with frequent access to a web page/App are identified as active users. Besides general user property information, related activity data are also saved for the active users due to frequent access to a web page/App. The inactive users with less access for no access for a long time period to a web page/App may only have user property data and no user activity data. For new users first accessing a web page/App, the user information is normally not complete without user property information or user activity data. In practical applications, the user type classification of target users can be updated periodically based on practical needs.

[0061] In an embodiment, the server can determine the user type of the target user according to the user information of the target user, by means of:

[0062] determining the user type of the described target user as the active user type when the user information is determined to include user property data and user activity data; determining the user type of the described target user as the inactive user type when the user information is determined to include user property data but no user activity data; and determining the user type of the described target user as the new user type when the user information is determined to not include user property data or user activity data.

[0063] In particular, user property information can include user basic information, user value property information, user interest preference information, etc. The user activity data can include exposure data, clicking data, and viewing duration for promotion events. For practical applications, the long-past data may not be valuable for referencing. As a result, the user activity data generally include activity data of the target user within a certain period of time before the current time point, such as within 1 month, 3 months, or a year. In other words, where if an active user has not accessed a web page/App within a long time period before the current time point, the active user may lack user activity data, and consequently the user is identified as an inactive user.

Date recue/ date received 2021-12-23

[0064] In the present embodiment, the user type of the target user is determined by types of information included in the user information. Due to time passing, user types are generally converting. For example, new users can become an active user with frequent access. And the user type conversion is related to types of information included in the user information. Therefore, the user type of the target user can be updated more immediately based on determination of user type of the target user based on types of information included in the user information, gaining more time effectiveness on determination of user types, so as for adopting more appropriate algorithm for acquiring an event promoting listof the target user in the following process, and promoting events to users more accurately.

[0065] Step 204, acquiring event promoting lists corresponding to the target user based promoting algorithm of user types, wherein the promoting algorithm for the active user type is a promoting model based on gradient boosting decision tree and logistic regression; the promoting algorithm for the inactive user type is a user similarity based promoting algorithm; and the promoting algorithm for the new user type is an event popularity-based promoting algorithm.

[0066] In particular, the trained promoting model based on gradient boosting decision tree and logistic regression can be noted as GBDT+LR model. GBDT is the gradient boosting decision tree, and LR is the logistic regression algorithm.

[0067] In particular, a promoting algorithm corresponding to the user type is selected by the use type of the target user, and an event promoting list of the target user is acquired based on the promoting algorithm corresponding to the user type.

[0068] In an embodiment, an event promoting list of the target user is acquired based on the promoting algorithm corresponding to the user type by means of:

[0069] acquiring candidate event set corresponding to the described target user when the user type of the target user is determined as the active user type or new user type, selecting at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user, and obtaining an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions.

[0070] Furthermore, when the user type of the targe user is the active user type, the user information of the target user includes user property information and user activity data. The preference scores to each event type by the target user can be calculated from the user activity data. Then, preference of individual event types is assigned as a section of the described user property data of the described target user, obtaining the current user property data of the target user. Then, the candidate event set corresponding to the target user is acquired to calculate current popularity predicted value for each candidate event in the candidate event set. The candidate events are sorted based on the current popularity predicted values and the event promoting list of the target user is acquired.

Date recue/ date received 2021-12-23

[0071] When the user type of the target user is the inactive user type, user property data of each active user from the currently active user set is acquired. The similarity between the described target user and each active user is calculated based on the described target user property data and the user property data of each active user, to obtain an event promoting list for an active user with the highest similarity with the described target user. The described event promoting list is assigned as the event promoting list for the described target user, wherein the described active users are classified as the active user type. In detail, the similarity calculation methods such as Euclidean distance and the cosine distance can be used to calculate similarity between the described target user and each active user is calculated based on the described inactive user property data.

[0072] In particular, the aforementioned trained promoting model based on gradient boosting decision tree and logistic regression is trained with training data sets including multiple training data combinations and tag information corresponding to each training data combination, wherein each described training combination includes pre-processed user property data of sample users and event property data of sample events, and the described sample users are classified as the active user type.
The user property data includes preference scores for each event type by the described sample users, wherein the described preference scores are obtained by calculation based on user activity data of the described sample users. The described pre-processed event property data includes value-based vectors, and the value-based vectors for the described sample event Chinese names are obtained by segmenting the event Chinese names of the described sample events and converting from the described event Chinese name segmenting results via word2vec model.

[0073] In an embodiment, the operation structure for the event promoting method is shown in Fig. 3. The promoting model is constructed by the following of:

[0074] (1) extracting event content

[0075] Event data for the sample events primarily includes event codes, event Chinese names, event service line, event targets, event status, etc. In the present embodiment, the event Chinese names are majorly processed. First, corresponding word library of event Chinese name is constructed, including customized library and stop sign library. Taking economical events as an example, certain uncommon words but frequently appearing in event lists by service depaifinents, such as "caprice mortgage" or "caprice payment"
can be added into the customized library. For frequently used words without special meaning and punctuation marks, such as comma, period, colon, numbers, "of', "-'s", and "also", are added into the stop sign library. Then, jieba segmentation is used to process the event Chinese names to obtain segmented word series. For example, the Chinese name segmenting result of the event 1 is "bank", "saving"], the described word series replaces the original event Chinese names. Then the word2vec model is applied to the Chinese Date recue/ date received 2021-12-23 name segmenting result from (1) for value conversion. A value-based vector is obtained to achieve the conversion of Chinese names to value-based vectors.

[0076] (2) Data processing and feature engineering

[0077] User property information can include user basic information, user value property information, user interest preference information, etc. The user activity data can include exposure data, clicking data, and viewing duration for promotion events. In particular, the user interest preference information is the preference scores to each event type by the user, wherein the preference scores to each event type can be calculated from the user activity data.

[0078] The user property data of the sample users and the event property data of the sample events are pre-processed. In other words, all features of the user property data and event property data are processed. For example, missing value statistics is performed for each aforementioned feature, wherein features with over 30% missing rate are deleted and refilled by average values and mode value for different feature types, then the features with single value for over 80% are deleted. The type features in the filled features are converted to value-based by StringToIndex function, then perform feature transform of the discrete features by one-hot coding.

[0079] (3) Constructing the promoting model based on gradient boosting decision tree and logistic regression

[0080] Based on user activity data of the sample users, the clicking status to an event by the sample users can be acquired. Then the data from (2) is labelled with the following pre-set rubrics.
10, click = 0 or duration < 5

[0081] label =
1, click = 1 and duration > 5

[0082] In particular, label is the label of a piece of data; click is the clicking status to the event by the user in the present piece of data; click = 0 indicates no clicks; click = /
indicates being clicked; and duration is the time spent on viewing in the unit of seconds. The sample data format is "user-event-label", wherein the user section is the user related information; the event second is the event and associated data (including all events clicked by the user and random picked non-click events). The "user-event" is a training set and label is associated label information for each training data set. By sampling the sample data, equalized training data with positive and negative samples are obtained and selected by a ratio of 8:2 to form a training data set and testing data set. During the promoting model training, the value-based vectors of event Chinese names are obtained via jieba segmentation and word2vec model. The value-based vectors are used as training data of the promoting model, so as for automatically converting new event Chinese names into value-based vectors via word2vec model when new event names appear in the event promoting prediction phase but not in the training phase.

Date recue/ date received 2021-12-23

[0083] The GBDT model is constructed. By training and optimization, finally the GBDT model with parameter setting of maximum depth of 7 using three-fold cross-validation yields the best performance and highest accuracy. The final leaf nodes of the GBDT classification model require the output of softmax function for the classification results. Therefore, after the training data passes the parameter-adjusted GBDT
model, the second last output data, as the leaf node output data, is extracted to form a new vector and assigned as a new feature added to original features of the training data.
Then, these combined features are put into the LR model for obtaining the final classifier via training. The final model achieves 92.1%
accuracy on the testing set. The structural diagram of the promoting model based on gradient boosting decision tree and logistic regression is shown in Fig. 4.

[0084] The traditional methods generally adopt multi-class model construction for N promotion events to users. However, the N-class model has lower accuracy than the binary classification models. Therefore, the embodiment of the present invention converts the N-class into binary problems, to significantly improve model stability. In the meanwhile, the GBDT algorithm can divide or combine space from an aspect of binary trees to obtain high-dimensional properties of the features and nonlinear relationships, exhibiting great capability for regression and generalization. Therefore, to maximize feature relationship extraction and model performance improvement, the GBDT is used to extract features, acquire extracted features by the model, and add to original features. After the feature combination, the LR
model is then used to complete the aforementioned promoting model based on gradient boosting decision tree and logistic regression with better accuracy and diversity.

[0085] Step 206, selecting at least one event from the event promoting lists to be promoted to the target user.

[0086] In particular, the event promoting list can include multiple events, sorted according to priority from highest to the lowest, wherein the promoting priority can predict probability of being clicked for events.
The events with greater predicted probability of being clicked will be prompted to the target user priorly.
In a practical application, when a user accesses a web page/App, the accessed page may include multiple displaying sites for event promoting displaying. Therefore, when promoting events to the target user, multiple events can be selected and displayed for each displaying sites for event promotion.

[0087] The events to be promoted are generally set with validity period. For example, the Christmas promotion events are only valid for the periods before and after the Christmas. To further achieve precise promotion, the events promoted to the users should all be valid events. In a practical application, after acquiring an event promoting list for the target user based on the promoting algorithm of the user type, the described method further includes selecting each valid candidate event from the described event promoting list, so as for obtaining a valid event promoting list based on each valid candidate event. In the meanwhile, the described selection of at least one event from the event promoting lists to be promoted to the target user Date recue/ date received 2021-12-23 includes selecting at least one event from the described valid event promoting lists to be promoted to the target user.

[0088] The present embodiment is applicable to promoting scenarios for economic marketing and promotion events, such as using prompt windows, floating layers and banners.
In practical applications, the event promoting method of the present invention can be combined with service conditions. For example, in a big sale, an event is promoted priorly; another event can only be promoted from Monday to Thursday;
and other service conditions may apply.

[0089] The aforementioned event promoting method first determines the user type of a target user according to user information of the event promotion target user, acquires a corresponding event promoting list according to a promoting algorithm for the user type, and promotes events to the target user based on the described event promoting list. In particular, different promoting algorithms are applied to users of different user types. By the promoting model based on gradient boosting decision tree and logistic regression, the performance of promoting models is greatly improved. The remaining users are covered by a promoting algorithm based on user similarity and event popularity. Each user can be applied with an appropriate algorithm for corresponding event promotion, to cover the majority of cases in event promotion and improve event promotion accuracy and diversity, preventing intemet resource wastes.
Embodiment two

[0090] As shown in Fig. 5, the described selection of at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user and attainment of an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions, include

[0091] Step 302, calculating the preference of each event type by the described target user when the user type of the described target user is the active user type, and assigning preference of individual event types as a section of the described user property data of the described target user.

[0092] In particular, the user activity data can include exposure data, clicking data, and viewing duration for promotion events.

[0093] In detail, the information such as clicked elements and purchased products are classified. The preference of an event type by the target user is calculated according to the following equation to obtain the preference score of the event type. The equation is shown below:

[0094] Scoreui = Ec(Expou, + Cliuc * 5) + Ep Buyup * 10

[0095] In particular, Score is the preference score to an event type i by the user u, wherein the event type i can be insurance, investment, funds, etc. P indicates events belonging to the event type 1, and c represents elements displayed on page for the event p of the event type i during user event promotions. Expo uc is the Date recue/ date received 2021-12-23 time of exposure of the element c to the user u. Clitic is the time of element c being clicked by the user u.
Buy up is the purchase time of the product p by the user u.

[0096] Step 304, pre-processing the user property data for the described target user and event property data of individual events in the candidate event set.

[0097] In an embodiment, the event property data includes event Chinese names.
The pre-processing of event property data of the candidate events in the candidate event set is primarily to process the event Chinese names of candidate events in the candidate event set. As shown in Fig.
6, the step 304 can include the following steps:

[0098] Step 402, segmenting event Chinese names of the described candidate events in the described candidate sets to obtain segment results for Chinese names of each described candidate event.

[0099] Step 404, converting the Chinese name segment results of each described candidate event into values via word2vec model, so as for acquiring value-based vectors corresponding to Chinese names of each described candidate event.

[0100] In particular, the promotion event names are segmented using the accurate mode of the jieba Chinese segmentation combined with the pre-set stop-sign library and the customized library, to acquire segment results of event Chinese names of individual events. The aforementioned stop-sign library and the customized library can adopt the same stop-sign library and the customized library from the embodiment one. Then, the Chinese name segment results of each described candidate event are converted into values via word2vec model, so as for acquiring value-based vectors corresponding to Chinese names of each described candidate event.

[0101] The step 304 further includes the following steps, performing missing-refill and vectorization of all data sections besides the described event Chinese names in the described user property data of the target user and the described candidate event property data in the described candidate set.

[0102] In particular, first, missing value statistics is performed for all data sections besides the described event Chinese names in the described user property data of the target user and the described candidate event property data in the described candidate set. Data sections with over 30%
missing rate are deleted. Value distributions for each data section are counted, and the data sections with single value for over 80% are deleted. Missing values of different data sections are refilled by average values and mode values to ensure no null-value in the data set of all data sections. Then, features of the discrete features in each data section are vectorized by one-hot coding, converting all discrete features into vector format, so as for obtaining data satisfying proper promoting model input formats.

[0103] Step 306, inputting the processed user property data of the described target user and the event property data of individual events in the candidate event set into a trained promoting model based on Date recue/ date received 2021-12-23 gradient boosting decision tree and logistic regression for prediction, and selecting individual candidate events with a predicted value greater than a pre-set threshold.

[0104] Step 308, sorting the individual candidate events with a predicted value greater than a pre-set threshold to obtain an event promoting list corresponding to the target user.

[0105] In detail, when the trained promoting model based on gradient boosting decision tree and logistic regression is used for prediction, the continuous values obtained by the logistic regression model prediction without binary-value process are acquired. Then the predicted values greater than the pre-set threshold are collected and sorted in descending order according to uses dimension groups, for obtaining a user event promoting list. In the present embodiment, based on the practical needs, the prediction threshold can be set as 0.5 or any other arbitrary value.

[0106] The event promoting method provided in the present embodiment adopts the trained promoting model based on gradient boosting decision tree and logistic regression. The described model can take advantage of the outstanding feature extraction by the gradient boosting decision tree model to extract data features, to be combined with the original features as new features. With the logistic regression model, the problems of over-regression by simple gradient boosting decision tree model can be prevented, to further improve model generalization and accuracy and yield more accurate and diversified event promoting results.
In the meanwhile, the jieba segmentation and word2vec model can be used to vectorize event Chinese names to obtain value-based vectors for candidate events. Therefore, when an event name not in the training appears, the promoting model based on gradient boosting decision tree and logistic regression can still acquire a prediction value for the candidate event according to the input value-based vector. The event code based problems of wrong prediction or not being able to predict can be prevented.
Embodiment three

[0107] In the aforementioned embodiment one, when the target user is an inactive user, the user activity data of the target user is missed. The promoting algorithm based on user similarity is adopted, suitable for users with user property data while missing user activity data or significantly lacking user activity data. As shown in Fig. 7, the event promoting list corresponding to the inactive user is acquired by means of:

[0108] Step 502, acquiring user property data of each active user from the currently active user set when the user type of the described target user is the inactive target user types, calculating the similarity between the described target user and each active user based the described target user property data and the user property data of each active user.

[0109] Step 504, obtaining an event promoting list for an active user with the highest similarity with the described target user, and assigning the described event promoting list as the event promoting list for the described target user, wherein the described active users are classified as the active user type.
Date recue/ date received 2021-12-23

[0110] In detail, the cosine similarity between the user property data of the target user with no user activity data and the user property data of each active user is calculated by the following equation:
Ei0it*Int)

[0111] COSO =

[0112] In particular, ui is the ith user feature of the active users, mi is the ith user feature of the no-activity-data user (the target user), and cos0 is the similarity between the target user m and the user u. The equation is used to calculate similarity between the two users in terms of user property dimension. The closer users are indicated by the value of cos0 closer to 1. After obtaining similarity between the target user m and each active user, the most similar user to the target user is identified. The promoting results of the described active user obtained by the trained GBDT+LG based promoting model is assigned to the target user m.

[0113] In detail, the step 504 includes selecting at least one candidate event satisfying pre-set conditions from the candidate event set according to trained promoting model based on gradient boosting decision tree and logistic regression and obtaining an event promoting list based on the candidate events satisfying pre-set conditions for an active user with the highest similarity with the described target user. The aforementioned calculation is for cosine similarity, wherein the most similar user is identified with the maximum value of the cosine similarity. The event promoting list of the most similar user is assigned as the event promoting list for the described target user lacking user activity data. As a possible application, the practical implementation can adopt other similarity calculation method, not limited by the cosine similarity calculation method. The detailed procedures of the step 504 are the same as the procedures for acquiring an event promoting list for the target user in embodiment two, and not further explained in detail.

[0114] The event promoting method provided in the present embodiment allows to calculate similarity between the target user and each active user by user property data of the target user when the target user has the user property information but lacks user activity data or significantly misses user activity data, yielding more accurate event promoting.
Embodiment four

[0115] The promoting algorithm in the embodiment two is based on the active user with user property data and user activity data, and the promoting algorithm in the embodiment three is based on users with user property data. In terms of new users lacking user property data and user activity data, as known as the cold start, the promoting algorithm based on popularity is used for event promoting for the described users. As shown in Fig. 8, the described selection of at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user and attainment of an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions, include Date recue/ date received 2021-12-23

[0116] Step 602, acquiring activity popularity data of each candidate event in the described candidate event set within a pre-set time interval when the target user is classified as the new user type.

[0117] In particular, the event popularity data can be calculated by the clicking data and event participating data of the candidate event within a pre-set time interval by weighted summation. In particular, the clicking data and event participating data of the candidate event within a pre-set time interval are collected individually and the weighted summation is performed to obtain popularity data of each candidate event.
Furthermore, the pre-set time interval can be set as one week to prevent massive computation volume and ensure timeliness of the predicted data.

[0118] The event popularity data also obeys the natural heat transfer and cooling rule. The present embodiment adopts Newton's cooling rule to predict the current popularity score based on the popularity score in the previous 7 days.

[0119] Step 604, calculating current popularity predicted values from the event popularity data of each candidate event in the pre-set time interval via Newton's law of cooling, so as for selecting candidate events satisfying pre-set conditions based on the described current popularity score.
The calculation equation is shown as the following:

[0120] Score = 15core1 * e(-ratio*inte*0

[0121] In particular, Scorei is the popularity score of a candidate event on the ith day, ratio is the cooling coefficient, inte is time interval constant, and Score is the current day popularity score of the candidate event predicted by the previous D days popularity scores. In practical applications, D can be set as 7, ratio is set as .02, and inte is set as 24.

[0122] Step 606, sorting candidate events satisfying pre-set conditions based on the described current popularity score, to obtain an event promoting list corresponding to the described target user.

[0123] In detail, the pre-set conditions can be a range of predicted popularity scores. If the current popularity prediction of a candidate event falls in the range of predicted popularity scores, the described candidate event satisfies the pre-set conditions. However, in practical applications, the range of predicted popularity scores is not mandatory. All candidate events in the candidate event sets are first identified as satisfying pre-set conditions, then sorted in descending order based on the obtained score values according to the step 604 to obtain an event promoting list sorted by popularity scores.
Furthermore, based on the status of each candidate events, the events are selected by validity. The candidate events with valid status are selected to obtain a valid event promoting list corresponding to the target user. Finally, at least one event from the valid event promoting list is selected and promoted to the target user.

[0124] The event promoting method in the present embodiment allows to use a popularity-based promoting algorithm for event promoting to users as new users lacing user property data and user activity data, as known as the cold start. The technical proposal in the present embodiment can predict and sort event Date recue/ date received 2021-12-23 popularity-based on event clicking data and participating status with significant data missing, then promote events to users according to the predicted event popularity.

[0125] To clarify, although the operations in the aforementioned flow diagrams in Fig. 2 - 8 follow the arrow and explained in sequence, these steps are not necessarily performed strictly following the sequence directed by the arrows. Except the clear indications in the present invention, the implementation of these operations should not be restricted by certain sequences. In other wordss, these steps are allowed to follow the other implementation sequences. In the meanwhile, at least some parts of the flow diagrams in Fig. 2 -8 may include multiple sub-steps or multiple phases. These sub-steps or phases are not necessarily performed in a restricted sequence. Indeed, the fulfillment of the sub-parts or phases may be achieved in turns or in an alternative manner.
Embodiment five

[0126] In the present embodiment, as shown in Fig. 9, an event promoting device is provided, comprising:
a primary acquisition module 110, a secondary acquisition module, 120, and an event promoting module, 130. In particular:

[0127] The primary acquisition module 110, is configured to acquire user information of a target user to be promoted with events, and determine the user types of the target user based on the user information, wherein user types include active user type, inactive user type, and new user type;

[0128] the secondary acquisition module 120, is configured to acquire event promoting lists corresponding to the target user based promoting algorithm of user types, wherein the promoting algorithm for the active user type is a promoting model based on gradient boosting decision tree and logistic regression, the promoting algorithm for the inactive user type is a user similarity based promoting algorithm, and the promoting algorithm for the new user type is an event popularity-based promoting algorithm; and

[0129] an event promoting module 130, is configured to select at least one event from the event promoting lists to be promoted to the target user.

[0130] In an embodiment, the primary acquisition module 110, is configured to determine the user type of the described target user as the active user type when the user information is determined to include user property data and user activity data; determine the user type of the described target user as the inactive user type when the user information is determined to include user property data but no user activity data;
and determine the user type of the described target user as the new user type when the user information is determined to not include user property data or user activity data.

[0131] In an embodiment, the secondary acquisition module 120 includes Date recue/ date received 2021-12-23 [O132] a primary list acquisition unit, configured to acquire candidate event set corresponding to the described target user when the user type of the target user is determined as the active user type or new user type, select at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user, and obtain an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions.
[0133] a secondary list acquisition unit, configured to acquire user property data of each active user from the currently active user set when the user type of the described target user is the inactive target user types, calculate the similarity between the described target user and each active user based the described target user property data and the user property data of each active user to obtain an event promoting list for an active user with the highest similarity with the described target user, and assign the described event promoting list as the event promoting list for the described target user, wherein the described active users are classified as the active user type.
[0134] In an embodiment, the primary acquisition unit is configured to calculate the preference of each event type by the described target user, and assign preference of individual event types as a section of the described user property data of the described target user when the user type of the described target user is the active user type; pre-process the user property data for the described target user and event property data of individual events in the candidate event set; input the processed user property data of the described target user and the event property data of individual events in the candidate event set into a trained promoting model based on gradient boosting decision tree and logistic regression for prediction, select individual candidate events with a predicted value greater than a pre-set threshold, and sort the individual candidate events with a predicted value greater than a pre-set threshold to obtain an event promoting list corresponding to the target user; wherein preferably, the described activity property data includes event Chinese name, and the primary acquisition unit is configured to segment event Chinese names of the described candidate events in the described candidate sets to obtain segment results for Chinese names of each described candidate event, and convert the Chinese name segment results of each described candidate event into values via word2vec model, so as for acquiring value-based vectors corresponding to Chinese names of each described candidate event; and perform missing-refill and vectorization of all data sections besides the described event Chinese names in the described user property data of the target user and the described candidate event property data in the described candidate set.
[0135] In an embodiment, the secondary acquisition unit is configured to acquire a candidate event set corresponding to the active user with the highest similarity with the described target user; and select at least one candidate events satisfying pre-set conditions from the described candidate event set via trained promoting model based on gradient boosting decision tree and logistic regression for prediction, to obtain Date recue/ date received 2021-12-23 an event promoting list for an active user with the highest similarity with the described target user based on each described candidate event satisfying pre-set conditions.
[0136] In an embodiment, the primary acquisition unit is further configured to acquire activity popularity data of each candidate event in the described candidate event set within a pre-set time interval when the target user is classified as the new user type; calculate current popularity predicted values from the event popularity data of each candidate event in the pre-set time interval via Newton's law of cooling, so as for selecting candidate events satisfying pre-set conditions based on the described current popularity score;
and sort candidate events satisfying pre-set conditions based on the described current popularity score, to obtain an event promoting list corresponding to the described target user.
[0137] In an embodiment, the described promoting model based on gradient boosting decision tree and logistic regression is trained with training data sets including multiple training data combinations and tag information corresponding to each training data combination, wherein each described training combination includes pre-processed user property data of sample users and event property data of sample events, and the described sample users are classified as the active user type;
the described user property data includes preference scores for each event type by the described sample users, wherein the described preference scores are obtained by calculation based on user activity data of the described sample users;
the described pre-processed event property data includes value-based vectors, and the value-based vectors for the described sample event Chinese names are obtained by segmenting the event Chinese names of the described sample events and converting from the described event Chinese name segmenting results via word2vec model.
[0138] In an embodiment, the described device further includes a valid event selection module, configured to select each valid candidate event from the described event promoting list, so as for obtaining a valid event promoting list based on each valid candidate event;
and the event promoting module 130, configured to select at least one event from the described valid event promoting lists to be promoted to the target user.
[0139] The detailed descriptions of the event promoting device can refer to the specifications of the event promoting method, and are not further explained. Modules in the described event promoting device can be performed by all or in a portion of software, hardware, and the combinations. The aforementioned modules can be embedded as the hardware, independently installed in the processor of the computer apparatus, or stored as the software in the memory of the computer apparatus, so as for calling corresponding operations of aforementioned modules by the processor.
Embodiment six Date recue/ date received 2021-12-23 [0140] In an embodiment of the present invention, a computer apparatus is provided, wherein the described computer apparatus can be a server with the internal structure diagram shown in Fig. 10. The computer apparatus comprises a processor, a memory unit, a network connection port, and a database connected by system bus control. The processor of the computer apparatus is used to provide computation and control.
The memory unit of the computer apparatus includes a nonvolatile storage medium and an internal memory.
The operating system, computer programs, and databases are stored in the nonvolatile storage medium. The internal memory provides the operation environment for the execution of the operating system and the computer programs stored in the nonvolatile storage medium. The database of the computer apparatus is configured to store the message execution results. The network connection port of the computer apparatus is configured for communication with the external terminals via network connection. The execution of the computer apparatus by the processor permits the method of event promotion.
[0141] It is comprehensible for those skilled in the art that the structure shown in Fig. 10 represents only a portion of structure associated with the applications of the present invention. The computer apparatus associated with the applications of the present invention are not restricted or limited by the structure. An exact computer apparatus may include more components or less components than that is shown in the drawings, possibly with combinations of some components or different component layouts.
[0142] Embodiment seven [0143] In an embodiment, a readable computer storage medium with computer programs stored, wherein the procedures of event promoting methods of any one of embodiment one to embodiment four are performed when the described computer programs are executed on the described processor.
[0144] All or portions of the aforementioned procedures are comprehensible for those skilled in the art, and may be achieved by the computer program configured for sending commands to the related hardware.
The computer programs are stored in the computer readable nonvolatile memory unit. When the computer programs are executed on the processor, the aforementioned procedure of the embodiments may be included.
In particular, all the referred memory units, storages, databases, as well as any other media in the embodiments provided in the present invention, may include nonvolatile and/or volatile memory units. The nonvolatile memory units may include read-only memory (ROM), programmable ROM
(PROM), electrical programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory.
The volatile memory units may include random access memory (RAM) or external cache memory. To describe RAM without limiting, RAM may be different formats such as static RAM
(SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced DDRSDRAM (EDDRSDRAM), Synchlink DRAM (SLDRAM), direct rambus RAM (RDRAM), direct rambus dynamic RAM (DRDRAM) and rambus dynamic RAM (RDRAM).

Date recue/ date received 2021-12-23 [O145] The aforementioned technical strategies can be combined as possible. To be concise, not all possible combinations of each technical strategy in the aforementioned embodiments are explained.
However, the combinations of these technical strategies without any conflict shall fall in the protection scope of the present invention.
[0146] The aforementioned contents are preferred embodiments of the present invention, and shall not limit the applications of the present invention. Therefore, all alterations, modifications, equivalence, improvements of the present invention fall within the scope of the present invention.

Date recue/ date received 2021-12-23

Claims (10)

1. An event promotion method, comprises:
acquiring user information of a target user to be promoted with events, and determining user types of the target user based on the user information, wherein user types include active user type, inactive user type, and new user type;
acquiring event promoting lists corresponding to the target user based promoting algorithm of user types, wherein the promoting algorithm for the active user type is a promoting model based on gradient boosting decision tree and logistic regression; the promoting algorithm for the inactive user type is a user similarity based promoting algorithm; and the promoting algorithm for the new user type is an event popularity-based promoting algorithm; and selecting at least one event from the event promoting lists to be promoted to the target user.

2. The method of claim 1, is characterized in that, the described determination of user types of the target user based on the user information, includes:
determining the user type of the described target user as the active user type when the user information is determined to include user property data and user activity data;
determining the user type of the described target user as the inactive user type when the user information is determined to include user property data but no user activity data; and determining the user type of the described target user as the new user type when the user information is determined to not include user property data or user activity data.

3. The described method of claim 2, is characterized in that, the described acquisition of event promoting lists corresponding to the target user based promoting algorithm of user types, includes:
acquiring candidate event set corresponding to the described target user when the user type of the target user is determined as the active user type or new user type, selecting at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user, and obtaining an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions; and acquiring user property data of each active user from the currently active user set when the user type of the described target user is the inactive target user types, calculating the similarity between the described target user and each active user based the described target user property data and the user property data of each active user to obtain an event promoting list for an active user with the highest similarity with the described target user, and assigning the described event promoting list as the event promoting list for the described target user, wherein the described active users are classified as the active user type.

Date recue/ date received 2021-12-23

4. The method of claim 3, is characterized in that, the described selection of at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user and attainment of an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions, include when the user type of the described target user is the active user type, calculating the preference of each event type by the described target user, and assigning preference of individual event types as a section of the described user property data of the described target user;
pre-processing the user property data for the described target user and event property data of individual events in the candidate event set;
inputting the processed user property data of the described target user and the event property data of individual events in the candidate event set into a trained promoting model based on gradient boosting decision tree and logistic regression for prediction, selecting individual candidate events with a predicted value greater than a pre-set threshold, and sorting the individual candidate events with a predicted value greater than a pre-set threshold to obtain an event promoting list corresponding to the target user; wherein preferably, the described activity property data includes event Chinese name, and the described pre-processing of the user property data for the described target user and event property data of individual events in the candidate event set includes:
segmenting event Chinese names of the described candidate events in the described candidate sets to obtain segment results for Chinese names of each described candidate event, and converting the Chinese name segment results of each described candidate event into values via word2vec model, so as for acquiring value-based vectors corresponding to Chinese names of each described candidate event; and performing missing-refill and vectorization of all data sections besides the described event Chinese names in the described user property data of the target user and the described candidate event property data in the described candidate set.

5. The method of claim 3, is characterized in that, before obtaining an event promoting list for an active user with the highest similarity with the described target user, the described method includes:
acquiring a candidate event set corresponding to the active user with the highest similarity with the described target user; and selecting at least one candidate events satisfying pre-set conditions from the described candidate event set via trained promoting model based on gradient boosting decision tree and logistic regression for prediction, to obtain an event promoting list for an active user with the highest Date recue/ date received 2021-12-23 similarity with the described target user based on each described candidate event satisfying pre-set conditions.

6. The method of claim 3, is characterized in that, the described selection of at least one candidate event satisfying pre-set conditions from the described candidate event set corresponding to the user type of the target user and attainment of an event promoting list corresponding to the described target user according to individual candidate event satisfying pre-set conditions, include acquiring activity popularity data of each candidate event in the described candidate event set within a pre-set time interval when the target user is classified as the new user type;
calculating current popularity predicted values from the event popularity data of each candidate event in the pre-set time interval via Newton's law of cooling, so as for selecting candidate events satisfying pre-set conditions based on the described current popularity score;
and sorting candidate events satisfying pre-set conditions based on the described current popularity score, to obtain an event promoting list corresponding to the described target user.

7. The method of claim 1, is characterized in that, the described promoting model based on gradient boosting decision tree and logistic regression is trained with training data sets including multiple training data combinations and tag information corresponding to each training data combination, wherein each described training combination includes pre-processed user property data of sample users and event property data of sample events, and the described sample users are classified as the active user type;
the described user property data includes preference scores for each event type by the described sample users, wherein the described preference scores are obtained by calculation based on user activity data of the described sample users;
the described pre-processed event property data includes value-based vectors, and the value-based vectors for the described sample event Chinese names are obtained by segmenting the event Chinese names of the described sample events and converting from the described event Chinese name segmenting results via word2vec model;
preferably, after acquiring an event promoting list for the described target user from the promoting algorithm corresponding to the described user type, the method further includes:
selecting each valid candidate event from the described event promoting list, so as for obtaining a valid event promoting list based on each valid candidate event; and the described selection of at least one event from the event promoting lists to be promoted to the target user includes selecting at least one event from the described valid event promoting lists to be promoted to the target user.

8. An event promoting device, comprising:
Date recue/ date received 2021-12-23 a primary acquisition module, configured to acquire user information of a target user to be promoted with events, and determine the user types of the target user based on the user information, wherein user types include active user type, inactive user type, and new user type;
a secondary acquisition module, configured to acquire event promoting lists corresponding to the target user based promoting algorithm of user types, wherein the promoting algorithm for the active user type is a promoting model based on gradient boosting decision tree and logistic regression, the promoting algorithm for the inactive user type is a user similarity based promoting algorithm, and the promoting algorithm for the new user type is an event popularity-based promoting algorithm; and an event promoting module, configured to select at least one event from the event promoting lists to be promoted to the target user.

9. A computer apparatus comprises a memory unit, a processor, and computer programs stored in the memory unit executable on the processor, wherein the methods of any one of claims 1 ¨ 7 are performed when the described processor executes the described computer programs.

10. A readable computer storage medium with computer programs stored, wherein the methods of any one of claims 1 ¨ 7 are performed when the described computer programs are executed on the described processor.

Date recue/ date received 2021-12-23