TWM625395U - Item Recommendation Device Using Combinatorial Algorithm - Google Patents

Abstract

The present invention discloses an item recommendation device using a combination algorithm, which includes a front-end device and a running device. The front-end device is used for receiving recommendation algorithm requirements and input information. The running device is used to obtain algorithm running information from the algorithm running configuration library according to the recommended algorithm requirements, wherein the algorithm running information records one or more running modules to be run; the running device is from the data and model database, according to the algorithm The running information selects at least one trained model and at least one trained data of the running module, and obtains at least one running parameter, so as to establish the running program code of the running module; and the running device executes the running of the running module based on the input information The program code is used to generate at least one recommended item and transmit it to the front-end device.

Description

Item Recommendation Device Using Combinatorial Algorithm

The present invention relates to an item recommendation device using a combined algorithm, in particular, to an item recommendation device that modulates the algorithm and then uses it in combination according to requirements.

Many websites will use recommended positions on the webpage to provide users with other relevant information when users browse the webpage, such as other products related to the webpage content, advertisement content relevant to the user, etc. In this way, the revenue of the website is increased. The recommendation algorithm is the protagonist in determining the recommended content, and it is also the key to the difficulty of implementation.

In one known technique, each recommendation algorithm in a recommendation system is implemented as independently developed. In this case, since there are various user contact points, and each user contact point needs to constantly change and test different algorithms, there is a demand for a large number of algorithms, but the independent development of each algorithm requires Processes are often time-consuming and labor-intensive. In addition to easily causing the overall development schedule to be delayed, it also represents a large demand for human resources.

In another prior art technique, the use of shared databases and shared code is further improved. However, this method can only partially reduce the repetitive development time. The model training and serving of each algorithm still need to be developed independently. Moreover, if complex methods such as deep learning are used, Model modeling or hybrid recommendation algorithms, it is more time-consuming to develop and adjust an algorithm at this time.

Therefore, there is a real need to improve the conventional techniques to obtain the desired recommendation algorithm more efficiently.

In order to solve the above-mentioned problem, an item recommendation device using a combined algorithm is developed in the embodiment of the present invention, and the algorithm is modularized, so as to quickly respond to the needs of the recommendation algorithm and change with the change of the recommended site plan of the website. , thereby saving a lot of time and human resources required for development.

Specifically, an embodiment of the present invention develops an item recommendation device using a combination algorithm, which outputs at least one recommended item according to a received recommendation algorithm requirement and an input information, and the item recommendation device includes a front-end device, is used for receiving the recommendation algorithm requirement and the input information, wherein the input information includes a necessary information required to execute a recommendation algorithm required by the recommendation algorithm requirement; and a running device, the signal is connected to the front-end device and used for executing : Obtain an algorithm operation information from an algorithm operation configuration library according to the requirements of the recommended algorithm, wherein the algorithm operation information records one or more operation modules to be executed; from a data and model database, according to the The algorithm operation information selects at least one trained model and at least one trained data of the operation module, and obtains at least one operation parameter, thereby establishing an operation program code of the operation module, wherein the data and the model database are storing a plurality of trained data and a plurality of trained models; and executing the running program code of the running module based on the input information to generate the at least one recommended item, and transmit the at least one recommended item to the front-end device .

According to an embodiment, the algorithm operation information records the operation modules to be executed, and the operation device is further configured to construct an operation pipeline process according to a module association relationship between the operation modules; wherein According to the running pipeline process, the running program codes of the running modules are executed sequentially or in parallel to generate the at least one recommended item.

According to yet another embodiment, wherein the algorithm operation information includes an operation program information, an operation parameter information, a trained data information, an output data information and an output data for each of the operation modules to be executed Format information, and includes a module association between the running modules.

According to yet another embodiment, the algorithm operation information further records at least one execution result of at least one of the other operation modules required by the operation module, and the creation of the operation program code of the operation module It is also associated with the execution result of the other running modules that the running module needs to use.

According to another embodiment, the item recommendation device further includes the data and model database, which is connected to the running device by a signal; and the algorithm running configuration library, which is connected to the running device by a signal.

According to yet another embodiment, the item recommendation device further includes a training device for obtaining one or more training module information from a training module configuration library according to a training operation information record, wherein the training operation information record is to be trained one or more training modules of the training module; obtain a model to be trained, a training parameter and a required training input data of the training module from a raw data database according to the training module information, and generate the training accordingly a training program code of the module; and executing the training program code of the training module to generate a trained model of the training module and a trained data for storage in the data and model database.

According to yet another embodiment, wherein the training operation information records the training modules to be trained, and the training operation information records a module association relationship between the training modules, the training device is further configured to The module association relationship between the training modules constructs a training pipeline process; wherein according to the training pipeline process, the training program codes of the training modules are executed sequentially or in parallel to generate the training modules the trained models and the trained data of the group.

According to yet another embodiment, the training operation information further records at least one execution result of at least one of the other training modules required by the training module, and the creation of the training program code of the training module is also The execution result of the other training modules that the training module needs to use.

According to yet another embodiment, the input information is at least one of a user information, a user behavior history information, a service page information and a device information.

According to another embodiment, the item recommendation device further includes the training module configuration library, which is connected to the training device by signals; and the original data database, which is connected to the training device by signals.

Combining the technical features of the above-mentioned embodiments, the project recommendation device of the present invention can specifically have the following technical effects: by modularizing each recommendation algorithm, and respectively generating the trained data and/or models of each operating module, therefore When the algorithm needs to be run, it can be selected from the trained data and/or model and combined into the required algorithm through the control configuration configuration, which saves a lot of algorithm development time and algorithm adjustment test cycle time.

In order to describe the embodiments of the present invention more concretely, the following description is supplemented with the accompanying drawings.

When the user is browsing the service page information of the service page of the website, and the service page is planned to have a recommended position to present the recommended item, there is a need to output the recommended item through the recommendation algorithm, so as to provide the recommended item during the browsing process. user, and guide the user to further view the recommended items. For example, it may be that the user enters the shopping website through a user device such as a mobile phone or a computer, and browses the introduction page of the product in the shopping website. At this time, the recommended site may be, for example, the recommended site for related products , Bought and bought recommended spots, recommended spots with purchased products, recommended spots based on browsing history, recommended spots of popular products, recommended spots of products from the same store and other recommended spots related to purchasing behavior, and through such way to present the recommended items to the user. The novel technology is an item recommendation device that generates an algorithm by a combination method when a recommendation algorithm is required due to the recommended placement.

Please refer to FIG. 1 , which shows an item recommendation apparatus using a combinatorial algorithm according to an embodiment of the present invention. The project recommendation device 100 using a combination algorithm of the present invention includes a front-end device 101 , an operation device 102 , an algorithm operation configuration library 103 , and a data and model database 104 . In other embodiments, the algorithm operation configuration library 103 and the data and model database 104 may be external components of the project recommendation device 100 , that is, the minimum components to implement the project recommendation device 100 are the front-end device 101 , the operation device 101 and the operation device 102 .

The front-end device 101 is used for receiving recommendation algorithm requirements and input information. The recommendation algorithm requirement as described above occurs when the recommendation algorithm needs to be executed to present at least one recommended item, and the input information includes necessary information required to execute the recommendation algorithm.

Input information may be implemented as user information, user behavior history information, service page information, and/or device information. For example, user information can be information such as interest, age, race, occupation, membership level, and bound credit card; user behavior history information can be information such as historical consumption amount, Internet time, Internet time, purchased products, advertisements, etc. Information such as click habits, click advertisement types, types of user devices habitually used, orientation of handheld devices when browsing web pages, etc.; service page information can be information such as product introduction content on shopping sites, information provided on portal sites, etc.; The information can be, for example, the mobile phone model and computer model used by the user, so as to know various information such as operating system, browser type, screen size, and resolution.

The running device 102 obtains algorithm running information from the algorithm running configuration library 103 according to the requirements of the recommended algorithm, and the algorithm running information records the running modules to be run, wherein the number of running modules can be changed according to different requirements. One or more, unlimited.

The data and model database 104 stores a plurality of trained data and a plurality of trained models, and the running device 102 selects at least one trained model of the running module from the data and model database 104 according to the algorithm running information The model and at least one trained data are obtained, and at least one running parameter is obtained, thereby establishing the running program code of the running module.

Afterwards, the running device 102 executes the running program code of the running module based on the input information to generate at least one recommended item, and transmit the at least one recommended item to the front-end device 101 .

The algorithm operation information includes operation program information, operation parameter information, trained data information, output data information and output data format information of the operation module to be run; in addition, when implemented as multiple operation modules, The algorithm operation information also includes a module association relationship among the plurality of operation modules.

Accordingly, the running device 102 further constructs a serving pipeline according to the module association relationship among the plurality of running modules, and the iso-running program codes of the plurality of running modules are generated according to the running pipeline process. are executed sequentially or in parallel to generate at least one recommended item.

For example, a possible situation is that each running module does not have any dependency on each other, and in this case, each running module can be executed independently, so parallel or sequential execution can be selected. Another possible situation is that one or more of the input information required by one of the running modules is the output data generated by the other one or more running modules after executing the running results. In this case, the running module must wait for It can only be executed after other running modules are executed, so it needs to be executed in sequence. For example, when the information on the service page is the product iphone 11, if you want to provide a "look and see" recommended location, the operation pipeline process of the recommendation algorithm will be, first from the trained data "the vector of the product generated according to the click behavior" Start with "representation", use the vector KNN algorithm, the output result is the 100 most similar products of iphone 11, and then further filter the 100 most similar products based on the trained data "the user has purchased products in the last 14 days", and get the output The result is less than 100 products. Finally, the trained model "Personalized Deep Learning Sorting Based on User Click Behavior" is used to sort the above-mentioned less than 100 products to select the most suitable 10 product recommendation items; in addition, If you choose the "see and buy" recommendation algorithm, the execution pipeline process is similar to the above "see and see", you only need to change the last used trained module to the trained module "based on the user's purchase behavior" Personalized Deep Learning Ranking".

Accordingly, the algorithm operation information will further record at least one execution result of at least one of the other operation modules used by the operation module, and the creation of the operation program code of the operation module is also related to the operation module required by the operation module. The execution results of other running modules used.

Next, please refer to FIG. 2 , which shows an item recommendation apparatus using a combination algorithm according to another embodiment of the present invention. In this embodiment, different from FIG. 1 , the item recommendation apparatus 100 using the combination algorithm also further It further includes a training device 201 , a training module configuration library 202 and a raw data database 203 .

As described above, the data and model database 104 stores a plurality of trained data and a plurality of trained models for use by the operating device, and the training device 201 is used to generate the trained data and the trained models.

First, the training device 201 obtains the training module information from the training module configuration library 202 according to the training operation information, and the training operation information records the training modules to be trained, wherein the number of the training modules can be changed according to different requirements One or more, unlimited.

Next, according to the training module information of the training module, the training device 201 obtains the model to be trained, the training parameters and the required training input data of the training module from the original data database 203, and generates the training program of the training module accordingly. code.

The training input information may include, but is not limited to, for example, commodity information, commodity thesaurus, commodity classification information, order information, and user behavior information, such as click, browse, add to shopping cart, and purchase.

Afterwards, the training device 201 executes the training program code of the training module to generate the trained model and the trained data of the training module, and sends them to the data and model database 202 for storage.

When there are multiple training modules, the training operation information will further record the module association relationship between the multiple training modules, and the training device will construct a training pipeline according to the module association relationship between the multiple training modules training pipeline, and the training program codes of multiple training modules are executed sequentially or in parallel according to the training pipeline process, so as to generate multiple trained models and multiple trained models of multiple training modules material. For example, a possible situation is that each training module does not have any dependency on each other. In this case, each training module can be executed independently, so parallel or sequential execution can be selected. Another possibility is that one or more of the input data required by one of the training modules are already trained data or models generated after the other training modules execute the training results. In this case, the training module has to wait for other training modules. It can only be executed after the training module is executed.

Accordingly, the training operation information will further record at least one execution result of at least one of the other training modules required by the training module, and the establishment of the training program code of the training module is also associated with the training module. Execution results of other training modules to be used.

Here, the plurality of training data may include, but are not limited to, for example, vector representations of commodities generated according to click behaviors, vector representations generated according to purchase behaviors, consumer vector representations generated according to consumer behaviors, Purchased products in 14 days, statistics of hot-selling products in a certain category in a certain period of time, discount ratio of each product in a certain category, ranking of similar products of a certain product, ranking of related products in a search string, etc., without limitation . In addition, multiple trained models may include, but are not limited to, for example, content-based, collaborative filtering, association rule-based, utility-based, knowledge-based, deep learning-based, combined recommendation and other models, and the types and types of trained models are also not limited This new type.

It should be noted that the above-mentioned training system is implemented only when there is a training requirement, and is not a necessary program. In addition, there are different implementation options for the proposal of the training operation information. , or can be executed automatically by scheduling, such as once a day, once a week, when an event occurs, etc. There are various possibilities.

To sum up, the present invention, by modularizing the algorithm and generating the trained data and model separately, can combine the required algorithm through the control configuration when the requirements of the recommended algorithm change. , and then generate the required recommended items, which can not only achieve the purpose of rapid response, but also make the application of time and human resources more efficient, which is a very advantageous content.

The present invention is only disclosed in preferred embodiments herein. However, any person skilled in the art should understand that the above-mentioned embodiments are only used to describe the present invention and are not intended to limit the scope of the claimed patent rights of the present invention. All changes or substitutions that are equal or equivalent to the above embodiments should be construed as being covered within the spirit or scope of the present invention. Therefore, the protection scope of this new model shall be defined by the following patent application scope.

100: Project Recommendation Device 101: Front-end device 102: Running device 103: Algorithm Run Configuration Library 104: Data and Model Database 201: Training Device 202: Training module configuration library 203: Raw Data Repository

In order to make the above-mentioned and other objects, features, advantages and embodiments of the present invention more clearly understood, the accompanying drawings are described as follows: FIG. 1 is a schematic diagram of an item recommendation apparatus using a combination algorithm according to an embodiment of the present invention. FIG. 2 is a schematic diagram of an item recommendation apparatus using a combination algorithm according to another embodiment of the present invention.

100: Project Recommendation Device

101: Front-end device

102: Running device

103: Algorithm Run Configuration Library

104: Data and Model Database

Claims (10)

An item recommendation device using a combination algorithm, which outputs at least one recommended item according to a received recommendation algorithm requirement and an input information, and the item recommendation device includes: a front-end device for receiving the recommendation algorithm requirement and the input information, wherein the input information includes a necessary information required to execute a recommendation algorithm required by the recommendation algorithm requirement; and A running device, the signal is connected to the front-end device, and is used to execute: Obtaining algorithm operation information from an algorithm operation configuration library according to the recommendation algorithm requirements, wherein the algorithm operation information records one or more operation modules to be executed; From a data and model database, select at least one trained model and at least one trained data of the operation module according to the algorithm operation information, and obtain at least one operation parameter, thereby establishing an operation of the operation module program code, wherein the data and model database stores a plurality of trained data and a plurality of trained models; and Execute the running program code of the running module based on the input information to generate the at least one recommended item, and transmit the at least one recommended item to the front-end device. The project recommendation device of claim 1, wherein the algorithm operation information records the operation modules to be executed, and the operation device is further configured to execute: constructing a running pipeline process according to a module association relationship between the running modules; Wherein, according to the running pipeline process, the running program codes of the running modules are executed sequentially or in parallel to generate the at least one recommended item. The project recommendation device of claim 2, wherein the algorithm operation information includes an operation program information, an operation parameter information, a trained data information, an output data of each of the operation modules to be executed Information and an output data format information, and including a module association between the running modules. The project recommendation device as claimed in claim 2, wherein the algorithm operation information further records at least one execution result of at least one of the other operation modules required by the operation module, and the operation module's The creation of the running program code is also associated with the execution result of the other running modules that the running module needs to use. The item recommendation device according to claim 2, further comprising: the data and model database, the signal is connected to the running device; and The algorithm runs the configuration library, and the signal is connected to the running device. The item recommendation device according to claim 2, further comprising: a training device for performing: obtaining one or more training module information from a training module configuration library according to a training operation information, wherein the training operation information records one or more training modules to be trained; Obtain a model to be trained, a training parameter and a required training input data of the training module from an original data database according to the training module information, and generate a training program code of the training module accordingly; and The training program code of the training module is executed to generate a trained model of the training module and a trained data for storage in the data and model database. The item recommendation device according to claim 6, wherein the training operation information records the training modules to be trained, and the training operation information records a module association relationship between the training modules, the training device More for execution: constructing a training pipeline process according to the module association between the training modules; Wherein, according to the training pipeline process, the training program codes of the training modules are executed sequentially or in parallel to generate the trained models and the trained data of the training modules. The item recommendation device according to claim 7, wherein the training operation information further records at least one execution result of at least one of the other training modules required by the training module, and the training of the training module The creation of program code is also associated with the execution results of the other training modules that the training module needs to use. The item recommendation device according to claim 1, wherein the input information is at least one of a user information, a user behavior history information, a service page information and a device information. The item recommendation device as described in claim 6, further comprising: the training module configuration library, the signal is connected to the training device; and The original data database, the signal is connected to the training device.

Images