Deep-model inference apparatus, and material recommendation method and apparatus

By deploying a deep model inference device in a deep learning recommendation system and utilizing a K8S cluster and Tensorflow-Serving architecture for multi-path recall and ranking, the problem of improving revenue, click-through rate, and ARPU in existing technologies is solved, resulting in better recommendation performance.

WO2024230642A1PCT designated stage expired Publication Date: 2024-11-14SHANGHAI 2345 NETWORK TECH
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2024/091189
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-08
Filing Date
2024-05-06
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

Existing technologies in deep learning recommendation systems are insufficient to further improve key metrics such as revenue, click-through rate, and ARPU.

Method used

It employs a deep model inference device, deploys model management, routing, recommendation, and feature service modules through a K8S cluster, and combines the Tensorflow-Serving architecture to perform multi-path recall, fusion, coarse and precise ranking, and recommend materials in conjunction with business strategies.

Benefits of technology

It improved various metrics of the recommendation system, such as revenue, click-through rate, and ARPU, and optimized the user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2024091189_14112024_PF_FP_ABST
    Figure CN2024091189_14112024_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of material recommendation. Disclosed are a deep-model inference apparatus, and a material recommendation method and apparatus. The deep-model inference apparatus comprises a model management service module, a routing service module and a recommendation service module, which are deployed in a K8S cluster, wherein the routing service module, the recommendation service module and a feature service module are in communication connection with the model management service module. The material recommendation method comprises: performing multi-path recall from an item library, so as to obtain multiple paths of recall results; merging the multiple paths of recall results, so as to obtain a merged recall result; performing coarse ranking; performing fine ranking; performing re-ranking; and outputting a result. In the technical solution of the present application, recommendation request information is combined with user features and material features, TensorFlow-Serving is used to obtain recommended-material ranking information, and finally, materials which are obtained by performing screening by means of an algorithm are pushed to the user, such that various indicators recommended by an algorithm, such as revenue, click rate and ARPU value, are further increased.
Need to check novelty before this filing date? Find Prior Art

Description

A deep model inference device, material recommendation method and device Technical Field

[0001] The present invention relates to the technical field of material recommendation, and in particular to a deep model inference device, a material recommendation method and a device. The deep model inference device is based on Tensorflow-Serving. Background Art

[0002] With the continuous advancement of deep learning in various fields such as image processing, language processing, and ad click-through rate prediction, current technological developments have begun exploring the practical application of deep learning technology at the business level. Currently, existing algorithm prediction platforms primarily focus on providing end-to-end personalized recommendations for a wide range of news, images, videos, apps, and other content to project partners across various business models within a company. Throughout this entire recommendation process, the powerful end-to-end fitting capabilities provided by deep models significantly enhance the final recommendation results.

[0003] In the current market, it is still necessary to continue to develop deep models to further improve various indicators of algorithm recommendation, such as revenue, click-through rate, ARPU value, etc.

[0004] Summary of the Invention

[0005] This application provides a deep model reasoning device, a material recommendation method and a device to solve the above-mentioned technical problems that the existing technology does not meet the existing needs.

[0006] According to one aspect of the present application, an embodiment provides a deep model inference device, including: a model management service module, a routing service module, a recommendation service module and a feature service module deployed in a K8S cluster; the routing service module, the recommendation service module and the feature service module are communicatively connected to the model management service module; the K8S cluster is communicatively connected to the TF-Serving architecture via http and grpc protocols.

[0007] In one embodiment, the routing service module periodically pulls the latest TF-Serving instance information.

[0008] According to one aspect of the present application, an embodiment provides a material recommendation method, comprising the following steps:

[0009] S1. Perform multi-way recall from the item library based on some user characteristics to obtain multi-way recall results;

[0010] S2. Fusing the multiple recall results to obtain a mixed recall result;

[0011] S3. According to the deep model inference device of claim 1 or 2, roughly sorting the recall results obtained in S2;

[0012] S4. According to the deep model reasoning device of claim 1 or 2, accurate sorting;

[0013] S5. Reorder based on improving user experience;

[0014] S6. Output the results.

[0015] In one embodiment, in step S1, the multi-channel recall results include one or more of personalized recall results, popular recall results, and random recall results.

[0016] In one embodiment, in step S2, fusion is performed based on the ratio of the content libraries and the ratio of the algorithms configured for each recall channel.

[0017] In one embodiment, the control source ratio is output in steps S3 and S4.

[0018] In one embodiment, in step S5, corresponding business policies are used for re-arrangement.

[0019] In one embodiment, the business strategy includes one or more of de-reading, de-duplication, scattering, diversity guarantee, and insertion of fixed-type items.

[0020] According to one aspect of the present application, an embodiment provides a material recommendation device, including:

[0021] The recall module is used to perform multi-way recall from the item library based on partial user characteristics and obtain multi-way recall results;

[0022] A fusion module, configured to fuse the multiple recall results to obtain a mixed recall result;

[0023] A coarse sorting module, configured to roughly sort the recall results obtained in the fusion module according to the above-mentioned deep model inference device;

[0024] A precise sorting module, used for precise sorting based on the above-mentioned deep model inference device;

[0025] Rearrangement module, used to reorder based on improving user experience;

[0026] and

[0027] Output module, used to output results.

[0028] In one embodiment, the fusion module performs fusion according to the ratio of content libraries and the ratio of algorithms configured for each recall channel.

[0029] The technical solutions in the aforementioned embodiments of this application are developed specifically for specific recommendation needs. In this technical solution, recommendation request information is combined with user and item characteristics, and TensorFlow Serving is used to obtain recommended item ranking information. Ultimately, the algorithm-selected items are pushed to the user. This further improves various algorithmic recommendation metrics, such as revenue, click-through rate, and ARPU. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] FIG1 is a schematic diagram of calling an inference service of a deep model inference device according to an embodiment;

[0032] FIG2 is a flow chart of a material recommendation method according to an embodiment;

[0033] FIG3 is a flowchart of deep (ranking) model reasoning in one embodiment;

[0034] FIG4 is a schematic structural diagram of a material recommendation device in an embodiment. DETAILED DESCRIPTION

[0035] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this 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.

[0036] In order to enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0037] It should be noted that the terms "first", "second" etc. in this application 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 in appropriate circumstances, so that the embodiments of the application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, the process, method, system, product or equipment that includes 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 equipment.

[0038] It should be understood that when an element (such as a layer, film, region, or substrate) is described as being "on" another element, the element may be directly on the other element or intervening elements may be present. Moreover, in this application, when an element is described as being "connected to" another element, the element may be "directly connected to" the other element or "connected to" the other element through a third element.

[0039] Example 1

[0040] Referring to Figure 1, one embodiment provides a deep model inference device, comprising: a model management service module, a routing service module, a recommendation service module, and a feature service module deployed in a Kubernetes cluster; the routing service module, the recommendation service module, and the feature service module are communicatively connected to the model management service module; and the Kubernetes cluster is communicatively connected to the TF-Serving architecture via HTTP and GRPC protocols. In other words, the routing service, recommendation service, model management service, and feature service are all deployed in the Kubernetes cluster.

[0041] Among them, Kubernetes (often referred to as "K8s") is an open source container orchestration system for automatically deploying, scaling, and managing containerized applications. Kubernetes manages applications by organizing containers into logical units (called "Pods"). It can automatically distribute containers across multiple nodes in the cluster and expand or reduce the number of application replicas as needed. Kubernetes also provides various features, such as load balancing, automatic recovery, automatic expansion and contraction, automatic rolling upgrades, etc., to ensure high availability and elasticity of containerized applications.

[0042] In one embodiment, the routing service module periodically pulls the latest TF-Serving instance information.

[0043] In one embodiment, the feature service (module) obtains the user feature data and material data required by the model management service (module) by calling the cql statement provided by the Cassandra database.

[0044] Example 2

[0045] Referring to FIG2 , an embodiment provides a material recommendation method, including the following steps:

[0046] S1. Based on partial user characteristics, perform a multi-way recall from the item library to obtain multiple recall results; preferably, the multi-way recall results include one or more of personalized recall results, popular recall results, and random recall results. Specifically, referring to FIG2 , based on partial user characteristics, a small number of items of potential user interest are quickly retrieved from a massive item library; where C1, C2, C3, and C4 represent different types of recall results, such as personalized recall results, popular recall results, and random recall results.

[0047] S2. Merge the multiple recall results in S1 to obtain a mixed recall result. Preferably, the merging is performed based on the content library ratio (which can be understood as the control source ratio) and the algorithm ratio of each recall configuration.

[0048] S3. According to the deep model inference device of claim 1 or 2, the recall results obtained in S2 are roughly sorted (Pre-Rank). Preferably, the recall results are roughly sorted based on a simple model using a small number of user and item features, further reducing the number of selected items while ensuring a certain degree of accuracy. This is to prevent the number of items returned in the user recall process from being too large, resulting in the sorting process being slow to keep up. Therefore, a rough sorting process is added between the recall and the refined sorting process.

[0049] S4. According to the deep model inference device of claim 1 or 2, accurate ranking (Rank), wherein any features and complex models can be used to perform personalized ranking of items as accurately as possible.

[0050] S5. Re-rank based on improving user experience;

[0051] S6. Output the results.

[0052] In one embodiment, the control source ratio needs to be output in steps S3 and S4.

[0053] In one embodiment, in step S5, a corresponding business strategy (driven by technical product strategy) is used for reordering. Preferably, the business strategy includes one or more of: removing read items, removing duplicates, breaking up, ensuring diversity, and inserting items of fixed types. Of course, in step S5, various corresponding technologies can also be used for reordering.

[0054] Example 3

[0055] Please refer to Figures 1 and 3. An embodiment provides a flowchart of deep (sorting) model reasoning. The flowchart describes the specific content in words and also illustrates the specific process. Please understand it in conjunction with Figure 1 and no further details will be given.

[0056] Example 4

[0057] Please refer to Figure 4. One embodiment provides a material recommendation device, including: a recall module 10, a fusion module 20, a rough sorting module 30, a fine sorting module 40, a rearrangement module 50, and an output module 60. The recall module 10 is used to perform multiple recalls from the item library based on partial user features to obtain multiple recall results; the fusion module 20 is used to fuse the multiple recall results to obtain a mixed recall result; the rough sorting module 30 is used to roughly sort the recall results obtained in the fusion module according to the deep model inference device of the above embodiment; the fine sorting module 40 is used to accurately sort according to the deep model inference device of the above embodiment; the rearrangement module 50 is used to re-sort based on improving user experience; and the output module 60 is used to output the results.

[0058] In one embodiment, the fusion module 20 performs fusion according to the ratio of content libraries and the ratio of algorithms configured for each recall channel.

[0059] In one embodiment, the rough sorting module 30 and the fine sorting module 40 need to output a control source ratio, which can be understood as a content library ratio.

[0060] In one embodiment, the rearrangement module 50 performs rearrangement using corresponding business strategies, preferably including one or more of: removing read items, removing duplicates, breaking up, ensuring diversity, and inserting items of a fixed type.

[0061] Example 5

[0062] Based on the same inventive concept, an embodiment of the present application provides a readable storage medium, on which computer instructions are stored; wherein, when the computer instructions are executed by a processor, the material recommendation method described in the above embodiment is implemented.

[0063] The one or more computer instructions mentioned above may form a program.

[0064] The above program can be executed in a processor or stored in a memory (or computer-readable medium), which includes permanent and non-permanent, removable and non-removable media that can implement information storage by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, tape, disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0065] These computer programs can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce computer-implemented processing, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram, and corresponding different steps can be implemented through different modules.

[0066] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A deep model inference device, characterized in that: include: Model management service module, routing service module, recommendation service module and feature service module deployed in the K8S cluster; The routing service module, the recommendation service module and the feature service module are communicatively connected with the model management service module; the K8S cluster is communicatively connected with the TF-Serving architecture via http and grpc protocols.

2. A deep model inference device according to claim 1, characterized in that: The routing service module periodically pulls the latest TF-Serving instance information.

3. A material recommendation method, characterized in that: The following steps are involved: S1, based on some user features, perform multi-way recall from the item library to obtain multi-way recall results; S2, fusing the multiple recall results to obtain a mixed recall result; S3, according to the deep model reasoning device in claim 1 or 2, roughly sorting the recall results obtained in S2; S4, according to the deep model reasoning device in claim 1 or 2, accurate sorting; S5, re-ranked based on improving user experience; S6, output results.

4. A material recommendation method according to claim 3, characterized in that: In step S1, the multi-channel recall results include one or more of personalized recall results, popular recall results and random recall results.

5. A material recommendation method according to claim 3, characterized in that: In step S2, fusion is performed according to the ratio of the content library and the ratio of the algorithm configured for each recall channel.

6. A material recommendation method according to claim 3, characterized in that: In steps S3 and S4, the control source ratio is output.

7. A material recommendation method according to claim 3, characterized in that: In step S5, corresponding business strategies are adopted for rearrangement.

8. A material recommendation method according to claim 7, characterized in that: The business strategy includes one or more of de-reading, de-duplication, scattering, diversity guarantee and fixed-type item insertion.

9. A material recommendation device, characterized in that: include: The recall module is used to perform multi-channel recall from the item library based on some user characteristics and obtain multi-channel recall results; A fusion module, used to fuse the multiple recall results to obtain a mixed recall result; A rough sorting module, used for roughly sorting the recall results obtained in the fusion module according to the deep model inference device of claim 1 or 2; A precise sorting module, used for precise sorting according to the deep model inference device of claim 1 or 2; Rearrangement module, used to reorder based on improving user experience; and Output module, used to output results.

10. A material recommendation device according to claim 9, characterized in that: The fusion module performs fusion according to the proportion of the content library and the proportion of the algorithm configured for each recall.

Citation Information

Patent Citations

  • Inference service system based on Kubernetes

    CN111629061A

  • Method and device for sorting multiple recall results

    CN114595406A

  • Model reasoning service deployment method, model reasoning method and device

    CN115617358A

  • Depth model reasoning device and material recommendation method and device

    CN116611483A

  • Deep Learning Job Scheduling Method and System and Related Device

    US20210011762A1