Bill recommendation method and device, storage medium, and electronic device

Through the bill recommendation method based on user access data diversion and feature extraction, the machine learning model is used to calculate the similarity between bills and user interests, and combined with historical and real-time data sorting, the problem of not being able to recommend suitable bills in the existing technology is solved, and the recommendation is improved accuracy and user experience.

CN114329191BActive Publication Date: 2024-11-26JIANGSU YINCHENG NETWORK TECH CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202111546350.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-15
Publication Date
2024-11-26
Estimated Expiration
2041-12-15

AI Technical Summary

Technical Problem

The prior art cannot effectively recommend appropriate bills to users, which affects users' property safety and purchasing efficiency.

Method used

By obtaining target user data based on the preset user access data diversion results, feature extraction and ticket recall are performed, and the recommendation model trained by machine learning calculates the similarity between tickets and user interests, and sorts them with historical purchasing behavior, real-time purchasing data and in-depth models. , recommend suitable notes.

Benefits of technology

It improves the accuracy of bill recommendations and user click-through rate, reduces purchase risks, ensures the safety of users' property, and realizes the recommendation of suitable bills.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114329191B_ABST
    Figure CN114329191B_ABST
Patent Text Reader

Abstract

The present application discloses a ticket recommendation method and device, storage medium, and electronic device. The method includes obtaining target user data based on preset user access data diversion results; extracting features from the target user data to obtain user feature data; recalling the ticket data of interest to the user based on the preset ticket feature data; inputting the ticket data of interest to the user and the preset ticket feature data into a preset recommendation model to obtain a target recommendation result; and obtaining the ticket recommendation result based on the ticket list sorting result in the target recommendation result. The present application solves the technical problem of being unable to recommend suitable tickets to users. The present application can better recommend tickets of interest to users, thereby improving the user's click-through rate and product conversion rate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular, to a bill recommendation method and device, a storage medium, and an electronic device. Background Art

[0002] Users can select different bills for electronic acceptance through the platform.

[0003] Many factors will affect whether the acceptance bill will be deferred, which is also an important factor related to the property safety of users. At the same time, users also need some useful recommendations for reference.

[0004] With respect to the problem in the related art that suitable tickets cannot be recommended to users, no effective solution has been proposed yet. Summary of the invention

[0005] The main purpose of the present application is to provide a bill recommendation method and device, a storage medium, and an electronic device to solve the problem of being unable to recommend suitable bills to users.

[0006] In order to achieve the above objective, according to one aspect of the present application, a bill recommendation method is provided.

[0007] The bill recommendation method according to the present application includes: obtaining target user data based on a preset user access data diversion result, wherein the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data; performing feature extraction on the target user data to obtain user feature data, wherein the user feature data includes preset bill feature data related to the bill, and the preset bill feature data includes at least one of the following: the transaction status of the bill, the financial status of the accepting enterprise, and the public opinion information of the accepting enterprise; based on the preset bill feature data, recalling the bill data of interest to the user; inputting the bill data of interest to the user and the preset bill feature data into a preset recommendation model to obtain a target recommendation result, wherein the preset recommendation model is obtained by machine learning training using multiple groups of data, and each group of data in the multiple groups of data includes: bill data of interest to the user and bill feature data corresponding to the bill data of interest to the user, and the preset recommendation model is used to calculate the similarity between the bill feature data and the bill data of interest to the user based on the bill feature data and the bill data of interest to the user, and convert it into a corresponding score; based on the bill list sorting result in the target recommendation result, the bill recommendation result is obtained.

[0008] Furthermore, obtaining the ticket recommendation result according to the ticket list sorting result in the target recommendation result also includes: determining the ticket list sorting result in the target recommendation result according to a preset scenario, wherein the ticket list sorting result includes, in sequence: a first sorting method based on historical purchase behavior, historical click behavior, and real-time purchase data, and a second sorting method based on a neighborhood model or a deep model.

[0009] Furthermore, the first sorting method based on historical purchase behavior, historical click behavior, and real-time purchase data includes: in the first sorting method of the historical purchase behavior and the historical click row, historical popular data is used to supplement the users whose features are less than a preset number; in the first sorting method of real-time purchase data, the user's purchase feature data is saved in Redis, and the user preference data is updated in real time by executing a preset Flink task, and the features of the current popular data are counted at the same time, which are used to supplement the user data when the features are less than the preset number.

[0010] Furthermore, the second sorting method based on the deep model includes: in the sorting method of the deep model, recalling bill data according to the type of acceptance bank preferred by the user; inputting the bill data recall result into the preset deep recommendation model on the preset modeling service, obtaining the probability of user clicks, and outputting the sorting result of the bill data.

[0011] Furthermore, the deep model builds a model service based on Kubernetes containerization and TensorFlow serving.

[0012] Furthermore, the method of obtaining target user data based on preset user access data diversion results includes: according to the preset recommendation scenario classification, diverting user traffic according to actual needs in each preset recommendation scenario, wherein the preset recommendation scenario includes at least one of the following: empty search recommendation, commercial bill recommendation, and the preset user access data diversion results include 90% control group user data and 10% experimental group user data, and the control group user data does not adopt the recommendation strategy.

[0013] Furthermore, the target user data is subjected to feature extraction to obtain user feature data, including: the preset bill feature data includes at least one of the following: interest accrual days, face value, bill type, accepting enterprise type, bill annualization, accepting enterprise financial situation, and whether the accepting enterprise has negative information; when the continuity of the user feature data does not meet the preset conditions, the preset user feature data with continuity is mapped to different categories to form feature pairs; when user feature data is lacking, the current popular feature data is used to fill in, by selecting the user's recently unused purchase features and click features, and sorting them according to time decay, and then sending these feature data to the user portrait library; when the real-time nature of the user feature data does not meet the preset conditions, the user's real-time purchase features are extracted for recommendation.

[0014] In order to achieve the above objective, according to another aspect of the present application, a bill recommendation device is provided.

[0015] The bill recommendation device according to the present application includes: a diversion module, which is used to obtain target user data based on a preset user access data diversion result, wherein the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data; a feature extraction module, which is used to extract features from the target user data to obtain user feature data, wherein the user feature data includes preset bill feature data related to the bill, and the preset bill feature data includes at least one of the following: the transaction status of the bill, the financial status of the accepting enterprise, and the public opinion information of the accepting enterprise; a bill recall module, which is used to recall the bill data of interest to the user based on the preset bill feature data; a model recommendation module, which is used to input the bill data of interest to the user and the preset bill feature data into a preset recommendation model to obtain a target recommendation result, wherein the preset recommendation model is obtained by machine learning training using multiple sets of data, and each set of data in the multiple sets of data includes: bill data of interest to the user and bill feature data corresponding to the bill data of interest to the user; a result output module, which is used to obtain the bill recommendation result based on the bill list sorting result in the target recommendation result.

[0016] In the embodiments of the present application, the ticket recommendation method and device, storage medium, and electronic device adopt a method of obtaining target user data based on the preset user access data diversion result, and extract features from the target user data to obtain user feature data and recall the ticket data of interest to the user based on the preset ticket feature data, thereby achieving the purpose of inputting the ticket data of interest to the user and the preset ticket feature data into the preset recommendation model to obtain the target recommendation result, thereby achieving the technical effect of obtaining the ticket recommendation result based on the ticket list sorting result in the target recommendation result, and further solving the technical problem of being unable to recommend suitable tickets to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The schematic embodiment drawings and their descriptions of this application are used to explain this application and do not constitute an improper limitation on this application. In the drawings:

[0018] Figure 1 is a schematic diagram of the hardware structure of the bill recommendation method according to an embodiment of the present application;

[0019] Figure 2 is a flowchart of a bill recommendation method according to an embodiment of the present application;

[0020] Figure 3 is a schematic diagram of the structure of a bill recommendation device according to an embodiment of the present application;

[0021] Figure 4 It is a flowchart of a bill recommendation method according to a preferred embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0023] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] In the present application, the terms "upper", "lower", "left", "right", "front", "back", "top", "bottom", "inner", "outer", "middle", "vertical", "horizontal", "lateral", "longitudinal" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings. These terms are mainly used to better describe the present application and its embodiments, and are not used to limit the indicated devices, elements or components to have a specific orientation, or to be constructed and operated in a specific orientation.

[0025] In addition, some of the above terms may be used to express other meanings in addition to indicating orientation or positional relationship. For example, the term "on" may also be used to express a certain dependency or connection relationship in some cases. For those of ordinary skill in the art, the specific meanings of these terms in this application can be understood according to specific circumstances.

[0026] In addition, the terms "installed", "set", "provided with", "connected", "connected", and "socketed" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection, or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be an internal connection between two devices, elements, or components. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0027] Considering that electronic acceptance bill recommendation is a special type of recommendation problem, due to the relatively single platform product, the data features that can be obtained are limited and the data is relatively sparse. However, whether the bill is purchased is not only related to the user's purchasing habits, the annualization of the bill, and the number of days for the bill to accrue interest, but also closely related to the market acceptance of the accepting company and the reputation of the accepting company. Especially for commercial bills, the company's financial status and market information directly affect whether the acceptance bill is deferred, and are important factors related to the user's property safety.

[0028] If suitable bills can be recommended to users, it will not only save users’ time costs in purchasing bills, but also reduce the risks of purchasing bills and protect users’ property safety.

[0029] The bill recommendation method in this application involves steps from data diversion, feature extraction, bill recall to the final bill list sorting. The feature data mainly comes from the transaction status of the bills on the platform, supplemented by the financial status and public opinion information of the acceptance enterprise, etc., which can better recommend bills that users are interested in and improve the user's click-through rate and conversion rate.

[0030] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] like Figure 1 As shown, the hardware structure in the embodiment of the present application includes: a user terminal 400, a database 300 of bill feature data, a bill trading platform 200, and a server 100. The user terminal 400 conducts bill transactions through the bill trading platform 200 and obtains relevant bill recommendation information at the same time. The database 300 of bill feature data stores a large amount of historical transaction data. The server 100 is used to respond to the request of the user terminal 400 on the bill trading platform 200.

[0032] like Figure 2 As shown, the method includes the following steps S201 to S205:

[0033] Step S201, acquiring target user data based on a preset user access data diversion result, wherein the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data;

[0034] Step S202, extracting features from the target user data to obtain user feature data, wherein the user feature data includes preset bill feature data related to the bill, and the preset bill feature data includes at least one of the following: bill transaction status, financial status of the accepting enterprise, and public opinion information of the accepting enterprise;

[0035] Step S203, recalling the ticket data that the user is interested in according to the preset ticket feature data;

[0036] Step S204, inputting the ticket data of interest to the user and the preset ticket feature data into a preset recommendation model to obtain a target recommendation result, wherein the preset recommendation model is obtained by machine learning training using multiple sets of data, each set of data in the multiple sets of data includes: the ticket data of interest to the user and the ticket feature data corresponding to the ticket data of interest to the user, and the preset recommendation model is used to calculate the similarity between the ticket feature data and the ticket data of interest to the user based on the similarity between the two and convert it into a corresponding score;

[0037] Step S205, obtaining the bill recommendation result according to the sorting result of the bill list in the target recommendation result.

[0038] From the above description, it can be seen that the present application achieves the following technical effects:

[0039] The method of obtaining target user data based on the preset user access data diversion result is adopted. By extracting features from the target user data, user feature data is obtained, and based on the preset ticket feature data, the ticket data of interest to the user is recalled. The purpose of inputting the ticket data of interest to the user and the preset ticket feature data into the preset recommendation model to obtain the target recommendation result is achieved, thereby achieving the technical effect of obtaining the ticket recommendation result based on the ticket list sorting result in the target recommendation result, and further solving the technical problem of being unable to recommend suitable tickets to users.

[0040] In the above step S201, the target user data is obtained based on the preset user access data diversion result. The user access data after diversion is more targeted and can also meet the requirements of different scenarios. According to the diversion result of the user access data, the target user data is further obtained.

[0041] As an optional implementation, the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data. It can be understood that the data is divided into control group user data and experimental group user data in the initial stage based on the AB TEST method. For example, 10% of users are selected to enter the experimental group, and the remaining 90% of users enter the control group, that is, no recommendation strategy is adopted.

[0042] As a preferred embodiment of this embodiment, the target user data is obtained based on the preset user access data diversion results, including: according to the preset recommendation scenario classification, user traffic is diverted according to actual needs in each preset recommendation scenario, wherein the preset recommendation scenario includes at least one of the following: empty search recommendation, commercial bill recommendation, and the preset user access data diversion results include 90% control group user data and 10% experimental group user data, and the control group user data does not adopt the recommendation strategy.

[0043] In the above step S202, feature extraction is performed on the target user data, and the target user-related tickets are focused on to obtain user feature data.

[0044] As an optional implementation, the user characteristic data includes preset ticket characteristic data related to the ticket. The preset ticket characteristic data related to the ticket covers factors of different dimensions.

[0045] As an optional implementation, the preset bill characteristic data includes at least one of the following: bill transaction status, financial status of the accepting enterprise, and public opinion information of the accepting enterprise. The bill transaction status, financial status of the accepting enterprise, and public opinion information of the accepting enterprise are taken as the main considerations.

[0046] As a preferred embodiment of the present invention, the feature extraction of the target user data to obtain the user feature data includes: the preset bill feature data includes at least one of the following: interest accrual days, par value, bill type, accepting enterprise type, bill annualization, accepting enterprise financial situation, and whether the accepting enterprise has negative information;

[0047] It can also solve the problem that when the continuity of user feature data does not meet the preset conditions, the preset user feature data with continuity is mapped to different categories to form feature pairs;

[0048] In the case of lack of user feature data, current popular feature data is used to fill in the gaps. By selecting the user's recently used purchase features and click features, and sorting them according to time decay, these feature data are sent to the user portrait library respectively;

[0049] When the real-time nature of user feature data does not meet the preset conditions, the user's real-time purchase features are extracted for recommendation.

[0050] In the above step S203, the ticket data that the user is interested in is recalled from the database according to the preset ticket feature data. For example, the user's latest five purchase features can be obtained from the ES user portrait system, and then 200 tickets that meet the partial feature conditions are screened from a large number of ticket libraries according to partial features.

[0051] In the above step S204, the user's interested bill data and the preset bill feature data are input into the preset recommendation model to obtain the target recommendation result.

[0052] As an optional implementation, the preset recommendation model is obtained by machine learning training using multiple sets of data, each of which includes: ticket data of interest to the user and ticket feature data corresponding to the ticket data of interest to the user. The training is performed based on the ticket data of interest to the user and the ticket feature data corresponding to the ticket data of interest to the user as a group.

[0053] As a preferred implementation, the preset recommendation model is used to calculate the similarity between the ticket feature data and the ticket data that the user is interested in, and convert it into a corresponding score. It can be understood that the higher the score, the more likely it is to be recommended to the user.

[0054] In the above step S205, the ticket recommendation result is obtained according to the sorting result of the ticket list in the target recommendation result.

[0055] As an optional implementation, the sorting results of the ticket list can be sorted in different dimensions according to different scenarios and stages.

[0056] As a preferred embodiment of the present invention, the ticket recommendation result is obtained according to the ticket list sorting result in the target recommendation result, and also includes: determining the ticket list sorting result in the target recommendation result according to a preset scenario, wherein the ticket list sorting result includes in sequence: a first sorting method based on historical purchase behavior, historical click behavior, and real-time purchase data, and a second sorting method based on a neighborhood model or a deep model.

[0057] It should be noted that different sorting methods can be used for different scenarios in different diverted user traffic data.

[0058] As a preferred embodiment of the present embodiment, the first sorting method based on historical purchase behavior, historical click behavior, and real-time purchase data includes: in the first sorting method of the historical purchase behavior and the historical click row, historical popular data is used to supplement the users whose features are less than a preset number; in the first sorting method of real-time purchase data, the user's purchase feature data is saved in Redis, and the user preference data is updated in real time by executing a preset Flink task, and the features of the current popular data are counted at the same time, which are used to supplement the user data when the features are less than the preset number.

[0059] As a preferred embodiment of the present invention, the second sorting method based on the deep model includes: in the sorting method of the deep model, the bill data is recalled according to the type of acceptance bank preferred by the user; the bill data recall result is input into the preset deep recommendation model on the preset modeling service to obtain the probability of user clicks, and the sorting result of the bill data is output.

[0060] As a preferred embodiment of this embodiment, the deep model builds a model service based on Kubernetes containerization and TensorFlowserving.

[0061] As a preferred embodiment of the present invention, the first sorting method of real-time purchase data includes: in the neighborhood model-based fine sorting, a spark task is executed each time, a user rating matrix is ​​constructed based on user clicks + purchase behavior, a bill similarity matrix is ​​constructed based on the user rating matrix, the latest preference vector is obtained based on the user's current preference and the similarity matrix, the preference vector is used to recall the database, and a recommended bill list is obtained, and the recall list is scored based on annualization and list similarity. Although this method can achieve good results in e-commerce, in bill recommendation, due to the limited number of users and the single product category, the recommended bills cannot meet the user's short-term memory, and the recommendation effect is not good.

[0062] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0063] According to an embodiment of the present application, a bill recommendation device for implementing the above method is also provided. Figure 3 As shown, the device comprises:

[0064] A diversion module 301 is used to obtain target user data based on a preset user access data diversion result, wherein the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data;

[0065] A feature extraction module 302 is used to extract features from the target user data to obtain user feature data, wherein the user feature data includes preset bill feature data related to the bill, and the preset bill feature data includes at least one of the following: bill transaction status, financial status of the accepting enterprise, and public opinion information of the accepting enterprise;

[0066] The bill recall module 303 is used to recall the bill data that the user is interested in according to the preset bill feature data;

[0067] The model recommendation module 304 is used to input the ticket data of interest to the user and the preset ticket feature data into a preset recommendation model to obtain a target recommendation result, wherein the preset recommendation model is obtained by machine learning training using multiple sets of data, and each set of data in the multiple sets of data includes: the ticket data of interest to the user and the ticket feature data corresponding to the ticket data of interest to the user;

[0068] The result output module 305 is used to obtain the bill recommendation result according to the bill list sorting result in the target recommendation result.

[0069] The diversion module 301 obtains target user data based on the preset user access data diversion result. The user access data after diversion is more targeted and can also meet the requirements of different scenarios. According to the diversion result of the user access data, the target user data is further obtained.

[0070] As an optional implementation, the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data. It can be understood that the data is divided into control group user data and experimental group user data in the initial stage based on the AB TEST method. For example, 10% of users are selected to enter the experimental group, and the remaining 90% of users enter the control group, that is, no recommendation strategy is adopted.

[0071] As a preferred embodiment of this embodiment, the target user data is obtained based on the preset user access data diversion results, including: according to the preset recommendation scenario classification, user traffic is diverted according to actual needs in each preset recommendation scenario, wherein the preset recommendation scenario includes at least one of the following: empty search recommendation, commercial bill recommendation, and the preset user access data diversion results include 90% control group user data and 10% experimental group user data, and the control group user data does not adopt the recommendation strategy.

[0072] The feature extraction module 302 extracts features from the target user data, focusing on the target user's related tickets to obtain user feature data.

[0073] As an optional implementation, the user characteristic data includes preset ticket characteristic data related to the ticket. The preset ticket characteristic data related to the ticket covers factors of different dimensions.

[0074] As an optional implementation, the preset bill characteristic data includes at least one of the following: bill transaction status, financial status of the accepting enterprise, and public opinion information of the accepting enterprise. The bill transaction status, financial status of the accepting enterprise, and public opinion information of the accepting enterprise are taken as the main considerations.

[0075] As a preferred embodiment of the present invention, the feature extraction of the target user data to obtain the user feature data includes: the preset bill feature data includes at least one of the following: interest accrual days, par value, bill type, accepting enterprise type, bill annualization, accepting enterprise financial situation, and whether the accepting enterprise has negative information;

[0076] It can also solve the problem that when the continuity of user feature data does not meet the preset conditions, the preset user feature data with continuity is mapped to different categories to form feature pairs;

[0077] In the case of lack of user feature data, current popular feature data is used to fill in the gaps. By selecting the user's recently used purchase features and click features, and sorting them according to time decay, these feature data are sent to the user portrait library respectively;

[0078] When the real-time nature of user feature data does not meet the preset conditions, the user's real-time purchase features are extracted for recommendation.

[0079] The bill recall module 303 recalls the bill data that the user is interested in from the database according to the preset bill feature data. For example, the user's latest five purchase features can be obtained from the ES user portrait system, and then 200 bills that meet some feature conditions can be screened from a large number of bill libraries according to some features.

[0080] The model recommendation module 304 inputs the user's interested bill data and the preset bill feature data into the preset recommendation model to obtain the target recommendation result.

[0081] As an optional implementation, the preset recommendation model is obtained by machine learning training using multiple sets of data, each of which includes: ticket data of interest to the user and ticket feature data corresponding to the ticket data of interest to the user. The training is performed based on the ticket data of interest to the user and the ticket feature data corresponding to the ticket data of interest to the user as a group.

[0082] As a preferred implementation, the preset recommendation model is used to calculate the similarity between the ticket feature data and the ticket data that the user is interested in, and convert it into a corresponding score. It can be understood that the higher the score, the more likely it is to be recommended to the user.

[0083] The result output module 305 obtains the bill recommendation result according to the sorting result of the bill list in the target recommendation result.

[0084] As an optional implementation, the sorting results of the ticket list can be sorted in different dimensions according to different scenarios and stages.

[0085] As a preferred embodiment of the present invention, the ticket recommendation result is obtained according to the ticket list sorting result in the target recommendation result, and also includes: determining the ticket list sorting result in the target recommendation result according to a preset scenario, wherein the ticket list sorting result includes in sequence: a first sorting method based on historical purchase behavior, historical click behavior, and real-time purchase data, and a second sorting method based on a neighborhood model or a deep model.

[0086] It should be noted that different sorting methods can be used for different scenarios in different diverted user traffic data.

[0087] Obviously, those skilled in the art should understand that the above modules or steps of the present application can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the present application is not limited to any specific combination of hardware and software.

[0088] In order to better understand the above-mentioned bill recommendation method process, the above-mentioned technical solution is explained below in combination with a preferred embodiment, but it is not used to limit the technical solution of the embodiment of the present invention.

[0089] The bill recommendation method in the embodiment of the present application is a first in the field of bills. It can reasonably divide user groups, capture user preferences for bills, and obtain the value of bills. After going online, due to the convenience of TensorFlow serving, it can quickly deploy and easily replace newly trained model files. After being combined with Kubernetes, it can ensure the stability of the recommendation service and can adapt to the network environment.

[0090] like Figure 4 FIG. 1 is a flow chart of a bill recommendation method in an embodiment of the present application, and the specific process includes the following steps:

[0091] Step S401, user access.

[0092] Step S402, obtaining the branch number where the user is located, and performing ABTEST diversion.

[0093] Step S403: whether the user is an experimental branch user.

[0094] User diversion is to verify the improvement of recommendation effects by different recommendation methods.

[0095] In ABTEST diversion, recommendation scenarios are set according to different situations, such as empty search recommendation and commercial bill recommendation, and user traffic is diverted according to actual needs within each recommendation scenario. For example, in the empty search recommendation of the "Enterprise Hall", 10% of users are selected to enter the experimental group, and the remaining 90% of users enter the control group, that is, no recommendation strategy is adopted.

[0096] Step S404: if yes, obtain user features.

[0097] Focus on the features of the user's related bills, such as the number of interest-bearing days, par value, bill type, acceptor type, bill annualization, acceptor's financial situation, and whether there is negative information. In order to prevent some continuous features from having a significant impact on the recommendation effect due to changes in values, and users are sensitive to some features, some continuous features are first mapped to different categories and combined to form feature pairs.

[0098] For users who lack features, current popular features are used to fill in. Since user choices have long-term or short-term memory, the user's five most recent purchase features and click features are first selected, sorted by time decay, and the features are sent to the user portrait respectively.

[0099] Since the features are not real-time enough, the results fluctuate greatly. Then, we extract the users’ real-time purchase features for recommendation.

[0100] Step S405: Recall the tickets that the user is interested in.

[0101] The user's historical purchasing behavior is used as the basis for recommendation. When a user visits the "Qiqi Hall", the interface will be called to obtain the user's five most recent purchase features from the ES user portrait system. Then, based on some features, 200 bills that meet some feature conditions are screened out from a large bill library.

[0102] Step S406, input the ticket and user characteristics into the algorithm service.

[0103] Step S407: generating recommendation results.

[0104] The sorting goes through several stages. In the case of empty search in the "Qiqi Hall", we adopted rough sorting of historical purchase behavior, rough sorting of historical click behavior, rough sorting of real-time purchase data, fine sorting based on neighborhood models, and fine sorting based on deep models.

[0105] In the case of rough sorting of historical purchase behavior and historical click behavior, historical popular data is used to complete the users with less than 5 features. In the case of real-time purchase data, the purchase features are saved in redis, and a Flink task is executed to update the user preference data in real time. At the same time, the current popular features are counted to complete the user data with less than 5 features. Then the recalled tickets and user features will be sent to the model deployed on Tenseorflow serving. The model service will calculate the similarity between the ticket features and user preferences and convert them into corresponding scores. Subsequent operations will sort the ticket list according to the scores, extract the 5 tickets with the highest scores, and return them.

[0106] In the neighborhood model-based fine ranking, a Spark task is executed each time to build a user rating matrix based on user click + purchase behavior, and a bill similarity matrix based on the user rating matrix. The user's current preference is multiplied by the similarity matrix to obtain the latest preference vector, and the preference vector is used to recall the database to obtain the recommended bill list. The recall list is scored based on annualization and list similarity. Although this method can achieve good results in e-commerce, it is not effective in bill recommendation due to the limited number of users and the single product category. The recommended bills cannot meet the user's short-term memory, and the recommendation effect is not good.

[0107] In the deep model sorting, the bank type preferred by the user is recalled, and then sent to the DeepFm model on tensorflow serving to calculate the probability of the user's click, and then sorted from large to small according to the probability, and finally the top five bills are intercepted for output. The model is divided into two parts, Fm and Deep. The Fm part is a low-order combination of features, taking into account breadth and feature cross-over, and the Deep part is a high-order combination of features with generalization ability.

[0108] In order to facilitate training and online deployment, this article adopts Kubernetes containerization + TensorFlow serving to build model services. TensorFlow serving has its own REST and GRPC interface services, without the need to write server-side code. It can deploy multiple models at the same time, automatically detect and load new model files, and can complete hot updates of models without terminating the service. Kubernetes containerized deployment is convenient and efficient, and can automatically pack, self-repair, automatically achieve horizontal expansion, automatically achieve service discovery and load balancing, and can also achieve automatic release and rollback. Through Kubernetes, multiple services can share storage (PV) and quickly connect model training services to AI online services.

[0109] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A bill recommendation method, characterized in that: include: Based on the preset user access data diversion result, the target user data is acquired, wherein the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data; Extracting features from the target user data to obtain user feature data, wherein the user feature data includes preset bill feature data related to the bill, and the preset bill feature data includes at least one of the following: transaction status of the bill, financial status of the accepting enterprise, and public opinion information of the accepting enterprise; The step of extracting features from the target user data to obtain user feature data includes: The preset bill characteristic data includes at least one of the following: interest-bearing days, par value, bill type, accepting enterprise type, bill annualization, accepting enterprise financial situation, and whether the accepting enterprise has negative information; When the continuity of the user feature data does not meet the preset conditions, the preset user feature data with continuity is mapped to different categories to form feature pairs; In the case of lack of user feature data, current popular feature data is used to fill in the gaps. By selecting the user's recently used purchase features and click features, and sorting them according to time decay, these feature data are sent to the user portrait library respectively; When the real-time nature of user feature data does not meet the preset conditions, extract the user's real-time purchase features for recommendation; Recalling the ticket data that the user is interested in according to the preset ticket feature data; Inputting the ticket data of interest to the user and the preset ticket feature data into a preset recommendation model to obtain a target recommendation result, wherein the preset recommendation model is obtained by machine learning training using multiple sets of data, each set of data in the multiple sets of data includes: the ticket data of interest to the user and the ticket feature data corresponding to the ticket data of interest to the user, and the preset recommendation model is used to calculate the similarity between the ticket feature data and the ticket data of interest to the user based on the similarity between the two and convert it into a corresponding score; The bill recommendation result is obtained according to the sorting result of the bill list in the target recommendation result.

2. The method according to claim 1, characterized in that The obtaining the bill recommendation result according to the bill list sorting result in the target recommendation result further includes: The ticket list sorting result is determined in the target recommendation result according to a preset scenario, wherein the ticket list sorting result includes, in sequence: a first sorting method based on historical purchase behavior, historical click behavior, and real-time purchase data, and a second sorting method based on a neighborhood model or a deep model.

3. The method according to claim 2, characterized in that The first sorting method based on historical purchase behavior, historical click behavior, and real-time purchase data includes: In the first sorting method of the historical purchase behaviors and the historical click rows, for users whose features are less than a preset number, historical hot data is used to supplement; In the first sorting method for real-time purchase data, the user purchase feature data is saved in Redis, and the user preference data is updated in real time by executing a preset Flink task. At the same time, the features of the current popular data are counted to supplement the user data when the preset number of features is insufficient.

4. The method according to claim 2, characterized in that: The second sorting method based on the depth model includes: In the sorting method of the deep model, the bill data is recalled according to the type of accepting bank preferred by the user; The ticket data recall result is input into a preset deep recommendation model on a preset modeling service to obtain the probability of user clicks, and the ranking result of the ticket data is output.

5. The method according to claim 4, characterized in that The deep model is based on Kubernetes containerization and TensorFlow serving to build model services.

6. The method according to claim 1, characterized in that The step of obtaining target user data based on the preset user access data diversion result includes: According to the classification of preset recommendation scenarios, user traffic is diverted according to actual needs in each preset recommendation scenario, wherein the preset recommendation scenarios include at least one of the following: empty search recommendation, commercial bill recommendation, and the preset user access data diversion results include 90% control group user data and 10% experimental group user data, and the control group user data does not adopt the recommendation strategy.

7. A bill recommendation device, characterized in that: include: A diversion module, configured to obtain target user data based on a preset user access data diversion result, wherein the preset user access data diversion result includes control group user data and experimental group user data, and the experimental group user data is used as the target user data; A feature extraction module is used to extract features from the target user data to obtain user feature data, wherein the user feature data includes preset bill feature data related to the bill, and the preset bill feature data includes at least one of the following: transaction status of the bill, financial status of the accepting enterprise, and public opinion information of the accepting enterprise; The step of extracting features from the target user data to obtain user feature data includes: The preset bill characteristic data includes at least one of the following: interest-bearing days, par value, bill type, accepting enterprise type, bill annualization, accepting enterprise financial situation, and whether the accepting enterprise has negative information; When the continuity of the user feature data does not meet the preset conditions, the preset user feature data with continuity is mapped to different categories to form feature pairs; In the case of lack of user feature data, current popular feature data is used to fill in the gaps. By selecting the user's recently used purchase features and click features, and sorting them according to time decay, these feature data are sent to the user portrait library respectively; When the real-time nature of user feature data does not meet the preset conditions, extract the user's real-time purchase features for recommendation; A bill recall module, used to recall the bill data of interest to the user according to the preset bill feature data; A model recommendation module, used for inputting the ticket data of interest to the user and the preset ticket feature data into a preset recommendation model to obtain a target recommendation result, wherein the preset recommendation model is obtained by machine learning training using multiple sets of data, and each set of data in the multiple sets of data includes: the ticket data of interest to the user and the ticket feature data corresponding to the ticket data of interest to the user; A result output module is used to obtain the bill recommendation result according to the bill list sorting result in the target recommendation result.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

9. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Citation Information

Patent Citations

  • Data processing method for bill score calculation and related device

    CN114330847A

  • Computer system and method for analyzing data sets and providing personalized recommendations

    US20120095863A1

  • Object recommendation method and apparatus, storage medium and terminal device

    US20210157860A1

  • Recommender and remediation system for enterprise service management

    US20210342536A1

  • Machine learning-facilitated data entry

    US20210342738A1