Content recommendation method and related equipment

By pre-training the content recommendation model in the cloud and fine-tuning it in real time on the terminal, the problems of lag in response to changes in user interests and excessive consumption of terminal resources are solved, the accuracy and user experience of content recommendation are improved, and different terminal types are adapted to.

CN119025748APending Publication Date: 2024-11-26TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202310610316.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-26
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing content recommendation method responds to lag when user interests change, and the model timeliness problem leads to unfriendly long-tail users, inaccurate recommendations, and excessively consumed terminal resources, which is a local optimal problem.

Method used

Pre-train the initial content recommendation network model in the cloud, and use the terminal's deep model inference ability to adjust the model in real time, combine the user behavior data collected by the terminal for fine-tuning, realize end-cloud collaboration, reduce end-cloud interaction, and improve response speed and recommendation accuracy.

Benefits of technology

It realizes rapid response to changes in user interests, improves the accuracy and user experience of content recommendations, reduces terminal resource consumption, avoids local optimization, adapts to different terminal types, and improves cost-effectiveness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025748A_ABST
    Figure CN119025748A_ABST
Patent Text Reader

Abstract

The invention discloses a content recommendation method and related equipment, and the method comprises the steps: a terminal can obtain an initial content recommendation network model issued by a server, and the initial content recommendation network model is trained on the server; collecting object behavior data; based on the object behavior data, performing model adjustment on the initial content recommendation network model to obtain a content recommendation network model; obtaining a recommendation candidate set generated by the server; and in response to an on-end rearrangement triggering operation, performing candidate set rearrangement on the recommendation candidate set based on the content recommendation network model, and obtaining and displaying a content recommendation result. The accuracy of content recommendation can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a content recommendation method and related equipment. Background Art

[0002] When a content recommendation system recommends content, it will send and present different types of content to the client for users to choose from. During this process, the recommendation system will capture changes in user interests and recommend content that better suits the user's interests. Therefore, whether the recommendation system can respond immediately when user interests change is very important for the effectiveness and reaction of the recommendation. The existing content recommendation method directly trains the model in the cloud and sends it to the terminal device for inference. However, due to the timeliness of the model itself, it is not friendly to long-tail users and the content recommendation is not accurate enough. Summary of the invention

[0003] The embodiments of the present application provide a content recommendation method and related devices, which can improve the accuracy of content recommendation.

[0004] The present application embodiment provides a content recommendation method, which is applied to a terminal and includes:

[0005] Obtaining an initial content recommendation network model sent by a server, wherein the initial content recommendation network model is trained on the server;

[0006] Collecting subject behavior data;

[0007] Based on the object behavior data, adjusting the initial content recommendation network model to obtain a content recommendation network model;

[0008] Obtaining a recommendation candidate set generated by the server;

[0009] In response to the on-end reordering trigger operation, the recommendation candidate set is reordered based on the content recommendation network model to obtain and display a content recommendation result.

[0010] Accordingly, an embodiment of the present application provides a content recommendation device, applied to a terminal, including:

[0011] A first acquisition unit is used to acquire an initial content recommendation network model sent by a server, wherein the initial content recommendation network model is trained on the server;

[0012] A collection unit, used for collecting object behavior data;

[0013] An adjustment unit, configured to adjust the initial content recommendation network model based on the object behavior data to obtain a content recommendation network model;

[0014] A second acquisition unit, used to acquire a recommendation candidate set generated by the server;

[0015] A display unit is used to respond to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model, and obtain and display the content recommendation result.

[0016] Optionally, in some embodiments of the present application, the adjustment unit may include a storage subunit, a trigger subunit, and an adjustment subunit, as follows:

[0017] A storage subunit, used for storing the object behavior data into an object behavior library;

[0018] a triggering subunit, configured to trigger a model adjustment request when the object behavior data in the object behavior library is accumulated to meet a model adjustment condition;

[0019] The adjustment subunit is used to perform model adjustment on the initial content recommendation network model based on the model adjustment request to obtain a content recommendation network model.

[0020] Optionally, in some embodiments of the present application, the adjustment subunit may include a model adjustment subunit and an evaluation subunit, as follows:

[0021] A model adjustment subunit, configured to perform model adjustment on the initial content recommendation network model based on the model adjustment request to obtain an adjusted network model;

[0022] The evaluation subunit is used to perform model evaluation on the adjusted network model to obtain an evaluation result, and determine the content recommendation network model based on the evaluation result.

[0023] Optionally, in some embodiments of the present application, the model adjustment subunit can be used to perform sample sampling on the object behavior library to obtain a sample set; based on the sample set, perform model adjustment on the initial content recommendation network model to obtain an adjusted network model.

[0024] Optionally, in some embodiments of the present application, the evaluation subunit can be used to perform test set sampling on the object behavior library to obtain a test set; obtain a basic network model; based on the test set, perform model evaluation on the adjusted network model and the basic network model to obtain an evaluation result; based on the evaluation result, determine a content recommendation network model from the adjusted network model and the basic network model.

[0025] Optionally, in some embodiments of the present application, the evaluation subunit can be used to obtain a first loss result of the adjusted network model on the test set, and to obtain a second loss result of the basic network model on the test set; based on the first loss result and the second loss result, perform model evaluation on the trained network model and the basic network model to obtain an evaluation result.

[0026] The present application embodiment provides a content recommendation method, which is applied to a server and includes:

[0027] Obtaining a basic content recommendation network model, and using basic samples to perform model training on the basic content recommendation network model to obtain an initial content recommendation network model;

[0028] Obtaining object feature data from a terminal;

[0029] Based on the object feature data, updating the initial content recommendation network model;

[0030] The initial content recommendation network model is sent to the terminal so that the terminal can adjust the initial content recommendation network model, thereby performing content recommendation at the terminal in real time.

[0031] Accordingly, an embodiment of the present application provides a content recommendation device, which is applied to a server and includes:

[0032] A training unit, used to obtain a basic content recommendation network model, and perform model training on the basic content recommendation network model using basic samples to obtain an initial content recommendation network model;

[0033] A third acquisition unit, used to acquire object feature data from a terminal;

[0034] An updating unit, configured to update the initial content recommendation network model based on the object feature data;

[0035] The sending unit is used to send the initial content recommendation network model to the terminal so that the terminal can adjust the initial content recommendation network model, thereby performing content recommendation in real time at the terminal.

[0036] Optionally, in some embodiments of the present application, the sending unit can be used to compress the initial content recommendation network model to obtain a compressed initial content recommendation network model; split the compressed initial content recommendation network model, and use a content distribution network to send the model to the terminal.

[0037] Optionally, in some embodiments of the present application, the content recommendation device can also be used to obtain network model parameters based on the initial content recommendation network model; generate a recommendation candidate set based on the network model parameters; and send the recommendation candidate set to the terminal so that the terminal displays content recommendation results based on the recommendation candidate set.

[0038] An electronic device provided in an embodiment of the present application includes a processor and a memory, wherein the memory stores a plurality of instructions, and the processor loads the instructions to execute the steps in the content recommendation method provided in the embodiment of the present application.

[0039] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps in the content recommendation method provided in the embodiment of the present application are implemented.

[0040] In addition, an embodiment of the present application also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the steps in the content recommendation method provided in the embodiment of the present application.

[0041] The embodiment of the present application provides a content recommendation method and related equipment, wherein the terminal can obtain an initial content recommendation network model sent by the server, and the initial content recommendation network model is trained on the server; collect object behavior data; adjust the initial content recommendation network model based on the object behavior data to obtain a content recommendation network model; obtain the recommendation candidate set generated by the server; in response to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model to obtain and display the content recommendation result. The server can obtain a basic content recommendation network model, and use basic samples to train the basic content recommendation network model to obtain an initial content recommendation network model; obtain object feature data from the terminal; update the initial content recommendation network model based on the object feature data; send the initial content recommendation network model to the terminal so that the terminal adjusts the initial content recommendation network model, thereby performing content recommendation on the terminal in real time. The present application proposes a new end-cloud collaboration framework, which upgrades and improves the existing end-cloud collaboration mode, and fully utilizes the global user and content information in the cloud by pre-training the initial content recommendation network model in the cloud. In addition, the present application also utilizes the terminal's deep model reasoning capabilities to pre-position real-time reasoning calculations and decision-making to the terminal. The terminal uses the collected real-time object behavior data to fine-tune the network model from the server, and then performs data reasoning and decision-making on the terminal side in real time, thereby reducing end-cloud interaction, so that the end-to-end response time can be reduced to hundreds of milliseconds, improving the overall cost-effectiveness, and avoiding excessive consumption of terminal resources and data sparseness as the model becomes larger, causing the network model to enter the local optimum. This can also solve problems such as the inability to rearrange low-end models and insufficient end re-arrangement candidate sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 is a scenario diagram of a content recommendation method provided in an embodiment of the present application;

[0044] Figure 2 is a first flow chart of the content recommendation method provided by an embodiment of the present application;

[0045] Figure 3 is a second flow chart of the content recommendation method provided in an embodiment of the present application;

[0046] Figure 4is a third flow chart of the content recommendation method provided in an embodiment of the present application;

[0047] Figure 5 is a schematic diagram of a system including a terminal-cloud collaboration module provided in an embodiment of the present application;

[0048] Figure 6 is a server-terminal interaction flow chart provided in an embodiment of the present application;

[0049] Figure 7 is a flow chart of model adjustment on a terminal provided in an embodiment of the present application;

[0050] Figure 8 It is a schematic diagram of model training provided in an embodiment of the present application;

[0051] Fig. 9 It is an overall system diagram provided by the embodiment of the present application;

[0052] Fig.10 is a first structural diagram of a content recommendation device provided in an embodiment of the present application;

[0053] Fig.11 is a second structural diagram of the content recommendation device provided in an embodiment of the present application;

[0054] Fig.12 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions 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 those skilled in the art without creative work are within the scope of protection of this application.

[0056] The embodiment of the present application provides a content recommendation method and related equipment, and the related equipment may include a content recommendation device, an electronic device, a computer-readable storage medium, and a computer program product. The content recommendation device can be specifically integrated in an electronic device. It can be understood that the content recommendation method of this embodiment is jointly executed by a terminal and a server. Among them, the terminal may include a mobile phone, an intelligent voice interaction device, an intelligent home appliance, a vehicle terminal, an aircraft, a tablet computer, a laptop computer, or a personal computer (PC, Personal Computer), etc. A client may also be set on the terminal, and the client may be an application client or a browser client, etc. Among them, the server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. In the content recommendation method or device disclosed in the present application, multiple servers can be composed of a blockchain, and the server is a node on the blockchain.

[0057] like Figure 1 As shown, the content recommendation system provided by the embodiment of the present application includes a terminal and a server. Among them, the terminal can obtain the initial content recommendation network model sent by the server, and the initial content recommendation network model is trained on the server; collect object behavior data; based on the object behavior data, adjust the model of the initial content recommendation network model to obtain the content recommendation network model; obtain the recommendation candidate set generated by the server; in response to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model to obtain and display the content recommendation result.

[0058] Among them, the server can obtain a basic content recommendation network model, and use basic samples to perform model training on the basic content recommendation network model to obtain an initial content recommendation network model; obtain object feature data from the terminal; based on the object feature data, update the initial content recommendation network model; send the initial content recommendation network model to the terminal so that the terminal can adjust the model of the initial content recommendation network model, thereby performing content recommendation at the terminal in real time.

[0059] The following are detailed descriptions. It should be noted that the description order of the following embodiments is not intended to limit the preferred order of the embodiments. This embodiment will be described from the perspective of a content recommendation device. Figure 2 As shown, the specific process of applying the content recommendation method to the terminal can be as follows:

[0060] 201. Obtain an initial content recommendation network model sent by a server, where the initial content recommendation network model is trained on the server.

[0061] Among them, terminal intelligence is to use terminal computing power as an extension of the cloud system. While solving problems such as network bandwidth and user information, it can provide feedback to users in a faster and more fine-grained manner, thereby improving the overall user experience of content recommendation. For example, terminal intelligence is specifically implemented in the technical solution of this application through terminal adjustments performed by a deep learning reasoning framework, where the deep learning reasoning framework belongs to the basic component of the terminal. For example, the technical solution of this application can use TNN as a deep learning reasoning framework, where TNN is a high-performance, lightweight reasoning framework for mobile terminals, which strengthens support for mobile devices and performance optimization, and has many advantages such as cross-platform, high performance, model compression, code tailoring, and good scalability.

[0062] In one embodiment, the initial content recommendation network model on the terminal is a network model that is personalized and adjusted using data on the terminal according to the initial content recommendation network model that needs to be fine-tuned and sent by the server, as well as a global shared parameter matrix related to the initial content recommendation network model and other important global features. Figure 6 The figure shows the overall interface of the terminal fine-tuning training and the server-side collaboration and the relationship flow chart of each module.

[0063] For example, each time the terminal is started, it can check whether there is a new engine library and initial content recommendation network model data in the cloud, where the engine library includes the TNN model and configuration, and the initial content recommendation network model data includes the distilled and compressed initial content recommendation network model file updated by the server to be distributed. If there is a new engine library and initial content recommendation network model data in the cloud, download them locally. Then, after the user starts the terminal, the terminal can send the path of the obtained model and configuration file to the terminal engine, which will be responsible for initializing it.

[0064] 202. Collect object behavior data.

[0065] Among them, the object behavior data is the behavior that is public to the user and can be clearly perceived by other users, such as commenting, liking, forwarding, collecting, modifying status, modifying signature, etc.

[0066] For example, in order to better realize terminal intelligence, the terminal can collect object behavior data in real time, where the object behavior data will be processed as sample data during the network model adjustment. Because the terminal uses the real-time collected object behavior data to fine-tune the initial content recommendation network model, the terminal can perceive the user's intention in seconds and make corresponding decisions. At the same time, it requests the end-cloud system to provide feedback that matches the user's intention in real time, thereby achieving the purpose of improving the user's willingness to click and browse, and overall changing the user's perception.

[0067] 203. Based on the object behavior data, the initial content recommendation network model is adjusted to obtain a content recommendation network model.

[0068] In one embodiment, if Figure 7 As shown, the terminal can adjust the initial content recommendation network model based on the object behavior data to obtain the content recommendation network model. For example, the terminal can fine-tune the initial content recommendation network model based on TNN. The fine-tuning process uses the object behavior data collected by the terminal in real time and the model feature data related to the initial content recommendation network model from the server.

[0069] By learning real-time object behavior data, the terminal realizes real-time incremental training on the terminal, achieving the optimal fusion between the global and local. This method can alleviate the problems of network models being difficult to learn and difficult to capture the long-tail behavior of users on the server side, achieving the effect of thousands of models for thousands of people on the terminal; on the other hand, since the terminal fine-tunes the network model, the focus is on fine-tuning the fully connected layer after feature extraction, which is not only highly real-time, but also saves a lot of training and computing resources, and has low performance loss, because there is no need to train the entire network model.

[0070] Optionally, in one embodiment, the step of “adjusting the initial content recommendation network model based on the object behavior data to obtain the content recommendation network model” may include:

[0071] storing the object behavior data into the object behavior library;

[0072] When the object behavior data in the object behavior library are accumulated to meet the model adjustment conditions, a model adjustment request is triggered;

[0073] Based on the model adjustment request, the initial content recommendation network model is adjusted to obtain a content recommendation network model.

[0074] For example, the data sequence of the object behavior data collected by the user in real time recently can be stored in the object behavior library, and a terminal adjustment counter can be maintained on the terminal. When the terminal adjustment counter accumulates to a preset number of times, that is, when the object behavior data in the object behavior library accumulates to meet the model adjustment conditions, the model adjustment request can be triggered. Among them, the interactive behavior between the user and the terminal, as well as the completion behavior, can be used as object behavior data and accumulated in the terminal adjustment counter. Among them, the preset number of times pre-set by the terminal adjustment counter can be set according to the number of user evaluations per day and the expected network model update frequency, such as the update frequency can be fine-tuned by day or by week.

[0075] In addition, in one embodiment, since the number of object behavior data in the object behavior library stored on the terminal needs to be greater than the sample threshold, invalid adjustments caused by overly sparse samples can be avoided. Therefore, a sample access mechanism can be set to control the number of samples, wherein the sample threshold can be set to 64. At the same time, it should also be ensured that the ratio of positive and negative samples is not unbalanced, such as the ratio between positive and negative samples cannot exceed 4:1. If the ratio of positive and negative samples is unbalanced, the network model inference result will shift toward the dominant label, thereby affecting the adjustment effect.

[0076] Optionally, in one embodiment, the step of “adjusting the initial content recommendation network model based on the model adjustment request to obtain the content recommendation network model” may include:

[0077] Based on the model adjustment request, the initial content recommendation network model is adjusted to obtain an adjusted network model;

[0078] The adjusted network model is evaluated to obtain an evaluation result, and based on the evaluation result, a content recommendation network model is determined.

[0079] For example, Figure 7 As shown, the processing of the network model on the terminal is divided into model adjustment and model evaluation, wherein the model adjustment is to fine-tune the initial content recommendation network model using the object behavior data collected in real time to obtain the adjusted network model; and the model evaluation is to determine whether to update the basic network model by evaluating the adjusted network model and the basic network model.

[0080] Optionally, in one embodiment, the step of “adjusting the initial content recommendation network model based on the model adjustment request to obtain an adjusted network model” may include:

[0081] Sampling the object behavior library to obtain a sample set;

[0082] Based on the sample set, the initial content recommendation network model is adjusted to obtain an adjusted network model.

[0083] The sample set includes positive samples and negative samples. Positive samples mainly include positive interactions and object behavior data such as completion. Positive interactions here can include playing, liking, recommending, sharing, forwarding and other behaviors. Completion here means that the completion rate of playback must meet certain conditions, such as a completion rate of more than 10%, which means that it cannot be content that users quickly swipe. Pure quick swiping behavior can be classified as a negative sample.

[0084] In one embodiment, the object behavior data in the object behavior library may be sampled and divided by a method of sequential extraction and fixed ratio division. For example, the object behavior data in the first 70% of the sequence in the object behavior library may be used as a sample set, and the object behavior data in the last 30% of the sequence in the object behavior library may be used as a test set, and so on.

[0085] Optionally, in one embodiment, the step of “performing a model evaluation on the adjusted network model to obtain an evaluation result, and determining a content recommendation network model based on the evaluation result” may include:

[0086] Sampling the test set of the object behavior library to obtain a test set;

[0087] Get the basic network model;

[0088] Based on the test set, the adjusted network model and the basic network model are evaluated to obtain the evaluation results;

[0089] Based on the evaluation result, a content recommendation network model is determined from the adjusted network model and the basic network model.

[0090] For example, the object behavior library can be sampled to obtain a test set and obtain the basic network model stored in the model library on the terminal. Then, based on the sampled test set, the adjusted network model and the basic network model can be evaluated to obtain the evaluation result. If the evaluation result shows that the evaluation is passed, the pointer loaded by the model library on the terminal can be updated to point to the adjusted network model, and the basic network model can be released. By adjusting the pointer of the network model loading address, the network model can be seamlessly switched without the user's perception. If the evaluation result shows that the evaluation is not passed, the adjusted network model is directly released.

[0091] Optionally, in one embodiment, the step of “performing model evaluation on the adjusted network model and the basic network model based on the test set to obtain evaluation results” may include:

[0092] Obtain a first loss result of the adjusted network model on the test set, and obtain a second loss result of the basic network model on the test set;

[0093] Based on the first loss result and the second loss result, a model evaluation is performed on the adjusted network model and the basic network model to obtain an evaluation result.

[0094] For example, after obtaining the adjusted network model, the first loss result of the adjusted network model on the test set and the second loss result of the basic network model on the test set can be evaluated. If the value of the first loss result is smaller than the value of the second loss result, the adjustment can be evaluated to be effective, and the network model can be updated to complete the incremental adjustment.

[0095] In one embodiment, the end-cloud collaboration in the technical solution of the present application is reflected in that the server sends the initial content recommendation network model to the terminal, and the terminal can synchronously update the adjusted user personalized vector data to the cloud, so that the cloud can maintain global information and use the powerful computing power of the cloud to extract commonalities, thereby achieving better global cold start and generalization effects.

[0096] 204. Obtain a recommendation candidate set generated by the server.

[0097] 205. In response to the on-end reordering trigger operation, the recommendation candidate set is reordered based on the content recommendation network model to obtain and display the content recommendation result.

[0098] For example, when a user's click operation triggers a re-arrangement on the terminal, the terminal can use the TNN architecture to re-arrange the recommendation candidate set from the server, obtain the content recommendation results, and display the content recommendation results.

[0099] The recommendation system in the prior art includes a solution that directly deploys the model learned on the cloud to the terminal, but due to the timeliness of the model itself, this method is not friendly to some long-tail users; the recommendation system in the prior art also includes training the model directly on the terminal, but due to its huge amount of calculation, it has special requirements on the computing power of the terminal. As the network model becomes larger, the excessive amount of calculation will lead to rapid consumption of terminal resources and hot mobile phones. At the same time, when the terminal data is sparse, it is impossible to avoid the model entering the local optimum, affecting the overall recommendation effect. In addition, the content recommendation system in the prior art will adopt a paging mechanism. Before paging, the changes in user preferences, the content recommendation system's perception of the user, and the timing of the content recommendation system's adjustment of the content cannot be adapted, which will cause the content recommended by the content recommendation system to not be what the user currently wants to see, thereby causing a decrease in the user's willingness to browse and consume.

[0100] Firstly, the technical solution of the present application can adapt to various terminal types with different capabilities through cloud-based training and fine-tuning dynamic deployment, effectively control the recommended resource consumption of the cloud and the terminal, achieve more efficient collaboration, and use the terminal computing power as an extension of the cloud system. While solving problems such as network bandwidth and user information, it can provide feedback to users at a faster speed and enhance the user's personalized experience. At the same time, the localization of calculations can alleviate the problem of cloud computing overload, reduce communication costs, and centralized operation and maintenance risks, and enhance user interaction and experience; secondly, the technical solution of the present application is more friendly to mid- and long-tail users, not only making full use of the global information of the cloud, but also making full use of the personalized real-time interactive information of terminal users to avoid entering the local optimum, so that the overall joint effect can be effectively guaranteed; thirdly, the technical solution of the present application fine-tunes the dynamic and regular update mechanism of the model on the terminal device, which can achieve seamless switching and reduce the user's direct perception. It is more friendly to low-end models and the overall cost performance of the project is higher.

[0101] In general, the technical solution of the present application can solve the problems of system feedback delay and user perception delay in the recommendation system in the prior art. By adding real-time reasoning capabilities to the terminal side model, the technical solution of the present application has the ability to perceive user intentions in seconds to make decisions and provide feedback that matches the intentions, thereby achieving the purpose of improving user willingness to click and browse and overall changing the user's perception.

[0102] As can be seen from the above, this embodiment is applied to the terminal, and the terminal can obtain the initial content recommendation network model sent by the server, and the initial content recommendation network model is trained on the server; collect object behavior data; adjust the model of the initial content recommendation network model based on the object behavior data to obtain the content recommendation network model; obtain the recommendation candidate set generated by the server; in response to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model to obtain and display the content recommendation result. This application proposes a new framework for end-cloud collaboration, which upgrades and improves the existing end-cloud collaboration mode, and fully utilizes the global user and content information on the cloud by pre-training the initial content recommendation network model on the cloud. In addition, the present application also utilizes the terminal's deep model reasoning capabilities to pre-position real-time reasoning calculations and decision-making to the terminal. The terminal uses the collected real-time object behavior data to fine-tune the network model, and then performs data reasoning and decision-making on the terminal side in real time, thereby reducing end-cloud interaction, so that the end-to-end response time can be reduced to hundreds of milliseconds, improving the overall cost-effectiveness, and avoiding excessive consumption of terminal resources and data sparseness as the model becomes larger, causing the network model to enter the local optimum. This can also solve problems such as the inability to rearrange low-end models and insufficient end re-arrangement candidate sets.

[0103] This embodiment will be described from the perspective of a content recommendation device. Figure 3 As shown, the specific process of applying the content recommendation method to the server can be as follows:

[0104] 301. Obtain a basic content recommendation network model, and use basic samples to perform model training on the basic content recommendation network model to obtain an initial content recommendation network model.

[0105] For example, the server may obtain a basic content recommendation network model, and use basic samples to perform model training on the basic content recommendation network model to obtain an initial content recommendation network model, wherein the basic samples may be sample data on the conventional recommendation side.

[0106] 302. Obtain object feature data from the terminal.

[0107] In one embodiment, the end-cloud collaboration in the technical solution of the present application is reflected in that the server sends the initial content recommendation network model to the terminal, and the terminal can synchronously update the adjusted user personalized vector data to the cloud, so that the cloud can maintain global information and use the powerful computing power of the cloud to extract commonalities, thereby achieving better global cold start and generalization effects.

[0108] 303. Update the initial content recommendation network model based on the object feature data.

[0109] For example, Figure 8 As shown, the server can obtain the user personalized vector data that has been adjusted from the terminal and use it as a feature to update the weight of the initial content recommendation network model.

[0110] 304. Send the initial content recommendation network model to the terminal so that the terminal can adjust the initial content recommendation network model, thereby performing content recommendation in the terminal in real time.

[0111] For example, the server can send the initial content recommendation network model to the terminal, including initializing the global shared parameter matrix, and the terminal can use it to fine-tune the network model, thereby making content recommendations in real time on the terminal. Since the global shared parameter matrix is ​​frozen on the terminal, training and parameter updates can only be performed on the server.

[0112] Optionally, in one embodiment, the step of “issuing the initial content recommendation network model to the terminal” may include:

[0113] Compressing the initial content recommendation network model to obtain a compressed initial content recommendation network model;

[0114] The compressed initial content recommendation network model is split into two parts, and the model is sent to the terminal using the content distribution network.

[0115] For example, Figure 6 As shown, the server can use the distillation method to compress the network model to obtain a compressed initial content recommendation network model, and then split the model for deployment, and use CDN (content distribution network) to send the model to the terminal.

[0116] Optionally, in one embodiment, the content recommendation method may further include:

[0117] Recommend the network model based on the initial content and obtain network model parameters;

[0118] Generate a recommendation candidate set based on network model parameters;

[0119] The recommendation candidate set is sent to the terminal so that the terminal can display content recommendation results based on the recommendation candidate set.

[0120] For example, Figure 6 As shown, the server can obtain network model parameters based on the initial content recommendation network model, and use the network model parameters to recall, coarse sorting, and fine sorting operations on the content, thereby generating a recommendation candidate set, and then send the recommendation candidate set to the terminal so that the terminal can display content recommendation results based on the recommendation candidate set.

[0121] Since the network model has been iteratively updated for a long time on the server side, a large number of basic scene-related features and multiple joint objectives including clicks and duration have been established. If these large-scale dimensional features and multi-objective optimization are directly reused on the terminal, it will cause huge computing overhead and storage and transmission pressure on the terminal side. However, if only the server side is used to output the estimated results, it will inevitably affect the cross-expression ability of the end-cloud features. In addition, the iteration and update of the models on both sides of the end-cloud will also incur a large maintenance cost.

[0122] Therefore, the technical solution of the present application adopts a joint end-cloud training method to introduce a large number of cloud feature cross-signals and multi-target high-order representations into the terminal side for use. In addition, the network model is distilled and compressed on the server side, and optimization steps such as deployment are performed to send the network model and related features to the terminal, which can further reduce the space occupied by the network model on the terminal side, as well as the impact from power consumption or performance.

[0123] As can be seen from the above, this embodiment is applied to the server, and the server can obtain the basic content recommendation network model, and use the basic samples to perform model training on the basic content recommendation network model to obtain the initial content recommendation network model; obtain object feature data from the terminal; update the initial content recommendation network model based on the object feature data; send the initial content recommendation network model to the terminal so that the terminal can adjust the model of the initial content recommendation network model, thereby performing content recommendation in real time at the terminal. This application proposes a new end-cloud collaboration framework, which upgrades and improves the existing end-cloud collaboration mode, and fully utilizes the global user and content information in the cloud by pre-training the initial content recommendation network model in the cloud. In addition, the present application also utilizes the terminal's deep model reasoning capabilities to pre-position real-time reasoning calculations and decision-making to the terminal. The terminal uses the collected real-time object behavior data to fine-tune the network model, and then performs data reasoning and decision-making on the terminal side in real time, thereby reducing end-cloud interaction, so that the end-to-end response time can be reduced to hundreds of milliseconds, improving the overall cost-effectiveness, and avoiding excessive consumption of terminal resources and data sparseness as the model becomes larger, causing the network model to enter the local optimum. This can also solve problems such as the inability to rearrange low-end models and insufficient end re-arrangement candidate sets.

[0124] This embodiment will be described from the perspective of a content recommendation device. Fig. 9 As shown, the modules of the content recommendation system and the functions of each module can be as follows:

[0125] like Fig. 9 The figure shows a flow chart of a method and system for recommending content in information streams based on cloud training and terminal fine-tuning deployment with dynamic updates. Fig. 9 The relationship between the various modules and their functional description:

[0126] 1. Content production

[0127] Content producers upload graphic content or video content through the content production end, and the above content can be used as the content source of the distributed content. The content production end first obtains the interface address of the upload server through communication with the upstream and downstream content interface servers, and then publishes the content.

[0128] 2. Content consumption

[0129] The content consumption end obtains the content access entry address and consumes the content through communication with the upstream and downstream content interface servers. The content consumption end can also collect the user's current network environment and object behavior data, and report it to the statistics reporting interface server. For video content, the content consumption end can report the video playback time, cache time, and various interactive behaviors for the video content, such as comments, forwarding, sharing, collection, likes, reports, etc.

[0130] 3. Uplink and downlink content interface server

[0131] The uplink and downlink content interface servers communicate directly with the content production end and store the content obtained from the content production end in the content database.

[0132] 4. Content export service

[0133] The content export service is a set of access services that are geographically deployed near users. The content export service communicates with the recommendation cloud system to obtain the results of recommendation distribution and sends the results of recommendation distribution to the content consumption end.

[0134] 5. Content Database

[0135] The content database is the core database of content. The meta information of the content published by the content producer can be stored in the content database. The information in the content database will be read during the manual review process, and the results and status of the manual review will be sent back to the content database for storage. The processing results obtained by the deduplication service will also be stored in the content database.

[0136] 6. Dispatch Center

[0137] The dispatch center is responsible for the entire dispatch process of content flow, receiving the incoming content through the upstream and downstream content interface servers, and then obtaining the content's meta-information from the content database. It can also dispatch duplicate removal services and manual review systems.

[0138] 7. Statistics reporting interface server

[0139] The statistics reporting interface server may receive statistics data reports from the uplink and downlink content interface servers, and write the reported statistics data into the statistics database.

[0140] 8. Terminal local processing system and terminal TNN inference engine

[0141] The terminal local processing system and the terminal TNN inference engine can collect and persist the object behavior data of local users, and process the samples of the local training end model accordingly. In addition, the terminal local processing system and the terminal TNN inference engine can also realize the division of labor and cooperation with the cloud recommendation model and auxiliary modules.

[0142] The terminal local processing system and the terminal TNN reasoning engine can check whether there is a new engine library (including TNN model and configuration) in the cloud and the server-side updated distilled and compressed network model file to be distributed each time the terminal is started. If there is, it will be downloaded locally. After the user starts the terminal, the terminal will give the model and its configuration file path to the terminal engine, which will be responsible for initialization. When the user's click behavior triggers a re-arrangement on the terminal, the terminal can use the TNN architecture to re-arrange the recommended candidate set and display the results.

[0143] 9. Terminal storage system

[0144] The terminal storage system is used to save different versions of the TNN engine obtained from the server, and accept the network model and important parameters sent by the server after training and compression in the cloud. It can also save the object behavior data collected by the terminal and the user's personalized vector.

[0145] 10. Statistical database

[0146] The statistical database is used to receive statistical data from content consumption terminals, and provide data support for subsequent statistical analysis and mining of user interests. Through communication with the cloud recommendation model and auxiliary modules, user-related data can be provided to them to realize the content recommendation function in the cloud.

[0147] 11. Recommend cloud system

[0148] The recommendation cloud system can implement a variety of content recall and sorting algorithms, such as collaborative recall, classification, theme recall, user history behavior, user interest point recall, etc. The recommendation cloud system usually recommends content based on core goals such as the click-through rate of the recommendation and the length of user time, using steps such as recall, rough sorting, fine sorting, and mixed sorting, and sends the results to the terminal.

[0149] 12. Cloud-based recommendation model and auxiliary modules

[0150] The cloud-based recommendation model and auxiliary modules are the corresponding recommendation model algorithm modules used to implement the recommendation of each cloud-based system. They use the information reported by the terminal and the global information of the server to complete the training and regular update of the relevant models. At the same time, the cloud-based model can also be distilled and compressed. According to the above description, the global parameters of the large model and important model feature information after relevant processing are sent to the terminal. The terminal fine-tunes and dynamically updates the model based on this, realizing the linkage and collaborative work of the terminal and the cloud.

[0151] As can be seen from the above, this embodiment is applied to the terminal, and the terminal can obtain the initial content recommendation network model sent by the server, and the initial content recommendation network model is trained on the server; collect object behavior data; adjust the initial content recommendation network model based on the object behavior data to obtain the content recommendation network model; obtain the recommendation candidate set generated by the server; in response to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model to obtain and display the content recommendation result. The server can obtain the basic content recommendation network model, and use the basic sample to train the basic content recommendation network model to obtain the initial content recommendation network model; obtain object feature data from the terminal; update the initial content recommendation network model based on the object feature data; send the initial content recommendation network model to the terminal so that the terminal adjusts the initial content recommendation network model, thereby performing content recommendation in real time on the terminal. This application proposes a new end-cloud collaboration framework, which upgrades and improves the existing end-cloud collaboration mode, and fully utilizes the global user and content information in the cloud by pre-training the initial content recommendation network model in the cloud. In addition, the present application also utilizes the terminal's deep model reasoning capabilities to pre-position real-time reasoning calculations and decision-making to the terminal. The terminal uses the collected real-time object behavior data to fine-tune the network model, and then performs data reasoning and decision-making on the terminal side in real time, thereby reducing end-cloud interaction, so that the end-to-end response time can be reduced to hundreds of milliseconds, improving the overall cost-effectiveness, and avoiding excessive consumption of terminal resources and data sparseness as the model becomes larger, causing the network model to enter the local optimum. This can also solve problems such as the inability to rearrange low-end models and insufficient end re-arrangement candidate sets.

[0152] According to the method described in the above embodiment, the following will be further described in detail by taking the content recommendation device specifically integrated into an electronic device as an example.

[0153] The present application embodiment provides a content recommendation method, such as Figure 4 As shown, the specific process of the content recommendation method can be as follows:

[0154] 401. The server obtains a basic content recommendation network model, and uses basic samples to perform model training on the basic content recommendation network model to obtain an initial content recommendation network model.

[0155] For example, the server can send the initial content recommendation network model to the terminal, including the initialization of the global shared parameter matrix, so that the terminal can use it to fine-tune the model. Since the global shared parameter matrix is ​​frozen on the terminal, model training and parameter updates need to be performed in the cloud. Figure 5 As shown, the main steps in the technical solution of this application are in the end-cloud collaboration module.

[0156] 402. The terminal obtains an initial content recommendation network model sent by the server.

[0157] For example, each time the terminal is started, it will check whether there is a new engine library (including TNN model and configuration) and EageFiles file (the initial content recommendation network model file after distillation and compression updated by the server to be distributed) in the cloud. If so, it will be downloaded locally; after the user starts it, the terminal will give the model and configuration file path to the terminal engine, and the terminal engine will be responsible for initialization.

[0158] 403. The terminal collects the object behavior data.

[0159] For example, in order to better realize terminal intelligence, the terminal can use the object behavior center, and these data will be processed as sample data during the terminal model adjustment process.

[0160] 404. The terminal adjusts the initial content recommendation network model based on the object behavior data to obtain a content recommendation network model.

[0161] For example, a terminal adjustment counter can be maintained on the terminal. When the object behavior data accumulates to the set number of times, the model adjustment can be triggered and the adjusted network model can be obtained. Then the basic network model stored in the terminal model library can be obtained, and the loss of the adjusted network model and the basic network model on the test set can be evaluated. If the loss of the adjusted network model is lower, the evaluation adjustment is effective, the pointer loaded by the model library is updated to point to the adjusted network model, and the basic network model is released; if it fails the evaluation, the adjusted network model is directly released.

[0162] In one embodiment, after the terminal completes model learning, it can report the adjusted user personalized vector data so that the cloud can maintain global information. Then the server can use the personalized vector reported by the terminal as a feature to update the weight of the cloud model.

[0163] 405. The terminal obtains a recommendation candidate set generated by the server.

[0164] 406. In response to the on-end reordering trigger operation, the terminal reorders the recommendation candidate set based on the content recommendation network model to obtain and display the content recommendation result.

[0165] For example, when a user's click behavior triggers a re-ranking on the terminal, the terminal can use the TNN architecture to re-rank the recommendation candidate set and return the results to the screen.

[0166] As can be seen from the above, this embodiment can obtain the basic content recommendation network model through the server, and use the basic samples to train the basic content recommendation network model to obtain the initial content recommendation network model; obtain the initial content recommendation network model issued by the server through the terminal; collect object behavior data through the terminal; adjust the initial content recommendation network model based on the object behavior data through the terminal to obtain the content recommendation network model; obtain the recommendation candidate set generated by the server through the terminal; and rearrange the recommendation candidate set based on the content recommendation network model in response to the rearrangement trigger operation on the terminal through the terminal to obtain and display the content recommendation result. This application proposes a new end-cloud collaboration framework, which upgrades and improves the existing end-cloud collaboration mode, and fully utilizes the global user and content information on the cloud by pre-training the initial content recommendation network model in the cloud. In addition, the present application also utilizes the terminal's deep model reasoning capabilities to pre-position real-time reasoning calculations and decision-making to the terminal. The terminal uses the collected real-time object behavior data to fine-tune the network model, and then performs data reasoning and decision-making on the terminal side in real time, thereby reducing end-cloud interaction, so that the end-to-end response time can be reduced to hundreds of milliseconds, improving the overall cost-effectiveness, and avoiding excessive consumption of terminal resources and data sparseness as the model becomes larger, causing the network model to enter the local optimum. This can also solve problems such as the inability to rearrange low-end models and insufficient end re-arrangement candidate sets.

[0167] In order to better implement the above method, the present application embodiment also provides a content recommendation device, which is applied to a terminal, such as Fig.10 As shown, the content recommendation device may include a first acquisition unit 1001, a collection unit 1002, an adjustment unit 1003, a second acquisition unit 1004, and a display unit 1005, as follows:

[0168] The first acquisition unit 1001 is used to acquire an initial content recommendation network model sent by a server, where the initial content recommendation network model is trained on the server;

[0169] A collecting unit 1002, used to collect object behavior data;

[0170] An adjustment unit 1003 is used to adjust the initial content recommendation network model based on the object behavior data to obtain a content recommendation network model;

[0171] A second acquisition unit 1004 is used to acquire a recommendation candidate set generated by the server;

[0172] The display unit 1005 is used to respond to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model, and obtain and display the content recommendation result.

[0173] Optionally, in some embodiments of the present application, the adjustment unit 1003 may include a storage subunit, a triggering subunit, and an adjustment subunit, as follows:

[0174] A storage subunit, used for storing the object behavior data into an object behavior library;

[0175] a triggering subunit, configured to trigger a model adjustment request when the object behavior data in the object behavior library is accumulated to meet a model adjustment condition;

[0176] The adjustment subunit is used to perform model adjustment on the initial content recommendation network model based on the model adjustment request to obtain a content recommendation network model.

[0177] Optionally, in some embodiments of the present application, the adjustment subunit may include a model adjustment subunit and an evaluation subunit, as follows:

[0178] A model adjustment subunit, configured to perform model adjustment on the initial content recommendation network model based on the model adjustment request to obtain an adjusted network model;

[0179] The evaluation subunit is used to perform model evaluation on the adjusted network model to obtain an evaluation result, and determine the content recommendation network model based on the evaluation result.

[0180] Optionally, in some embodiments of the present application, the model adjustment subunit can be used to perform sample sampling on the object behavior library to obtain a sample set; based on the sample set, perform model adjustment on the initial content recommendation network model to obtain an adjusted network model.

[0181] Optionally, in some embodiments of the present application, the evaluation subunit can be used to perform test set sampling on the object behavior library to obtain a test set; obtain a basic network model; based on the test set, perform model evaluation on the adjusted network model and the basic network model to obtain an evaluation result; based on the evaluation result, determine a content recommendation network model from the adjusted network model and the basic network model.

[0182] Optionally, in some embodiments of the present application, the evaluation subunit can be used to obtain a first loss result of the adjusted network model on the test set, and to obtain a second loss result of the basic network model on the test set; based on the first loss result and the second loss result, perform model evaluation on the trained network model and the basic network model to obtain an evaluation result.

[0183] The present application also provides a content recommendation device, which is applied to a server, such as Fig.11As shown, the content recommendation device may include a training unit 1101, a third acquisition unit 1102, an updating unit 1103, and a sending unit 1104, as follows:

[0184] The training unit 1101 is used to obtain a basic content recommendation network model, and perform model training on the basic content recommendation network model using basic samples to obtain an initial content recommendation network model;

[0185] The third acquisition unit 1102 is used to acquire object feature data from the terminal;

[0186] An updating unit 1103, configured to update the initial content recommendation network model based on the object feature data;

[0187] The sending unit 1104 is used to send the initial content recommendation network model to the terminal so that the terminal can adjust the initial content recommendation network model, thereby performing content recommendation in real time at the terminal.

[0188] Optionally, in some embodiments of the present application, the sending unit 1104 can be used to compress the initial content recommendation network model to obtain a compressed initial content recommendation network model; split the compressed initial content recommendation network model, and use a content distribution network to send the model to the terminal.

[0189] Optionally, in some embodiments of the present application, the content recommendation device can also be used to obtain network model parameters based on the initial content recommendation network model; generate a recommendation candidate set based on the network model parameters; and send the recommendation candidate set to the terminal so that the terminal displays content recommendation results based on the recommendation candidate set.

[0190] As can be seen from the above, this embodiment is applied to the terminal, and the terminal can obtain the initial content recommendation network model sent by the server, and the initial content recommendation network model is trained on the server; collect object behavior data; adjust the initial content recommendation network model based on the object behavior data to obtain the content recommendation network model; obtain the recommendation candidate set generated by the server; in response to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model to obtain and display the content recommendation result. The server can obtain the basic content recommendation network model, and use the basic sample to train the basic content recommendation network model to obtain the initial content recommendation network model; obtain object feature data from the terminal; update the initial content recommendation network model based on the object feature data; send the initial content recommendation network model to the terminal so that the terminal adjusts the initial content recommendation network model, thereby performing content recommendation in real time on the terminal. This application proposes a new end-cloud collaboration framework, which upgrades and improves the existing end-cloud collaboration mode, and fully utilizes the global user and content information in the cloud by pre-training the initial content recommendation network model in the cloud. In addition, the present application also utilizes the terminal's deep model reasoning capabilities to pre-position real-time reasoning calculations and decision-making to the terminal. The terminal uses the collected real-time object behavior data to fine-tune the network model, and then performs data reasoning and decision-making on the terminal side in real time, thereby reducing end-cloud interaction, so that the end-to-end response time can be reduced to hundreds of milliseconds, improving the overall cost-effectiveness, and avoiding excessive consumption of terminal resources and data sparseness as the model becomes larger, causing the network model to enter the local optimum. This can also solve problems such as the inability to rearrange low-end models and insufficient end re-arrangement candidate sets.

[0191] The present application also provides an electronic device, such as Fig.12 As shown, it shows a schematic diagram of the structure of the electronic device involved in the embodiment of the present application, specifically:

[0192] The electronic device may include components such as a processor 1201 with one or more processing cores, a memory 1202 with one or more computer-readable storage media, a power supply 1203, and an input unit 1204. Those skilled in the art will appreciate that Fig.12 The electronic device structure shown in the figure does not constitute a limitation on the electronic device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0193] The processor 1201 is the control center of the electronic device, which uses various interfaces and lines to connect various parts of the entire electronic device, and executes various functions of the electronic device and processes data by running or executing software programs and / or modules stored in the memory 1202, and calling data stored in the memory 1202. Optionally, the processor 1201 may include one or more processing cores; preferably, the processor 1201 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 1201.

[0194] The memory 1202 can be used to store software programs and modules. The processor 1201 executes various functional applications and content recommendations by running the software programs and modules stored in the memory 1202. The memory 1202 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 1202 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage device. Accordingly, the memory 1202 may also include a memory controller to provide the processor 1201 with access to the memory 1202.

[0195] The electronic device also includes a power supply 1203 for supplying power to each component. Preferably, the power supply 1203 can be logically connected to the processor 1201 through a power management system, so as to manage charging, discharging, power consumption and other functions through the power management system. The power supply 1203 can also include any components such as one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, and power status indicators.

[0196] The electronic device may further include an input unit 1204, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0197] Although not shown, the electronic device may further include a display unit, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 1201 in the electronic device will load the executable files corresponding to the processes of one or more application programs into the memory 1202 according to the following instructions, and the processor 1201 will run the application programs stored in the memory 1202, thereby realizing various functions, as follows:

[0198] The terminal can obtain an initial content recommendation network model sent by a server, wherein the initial content recommendation network model is trained on the server; collect object behavior data; adjust the initial content recommendation network model based on the object behavior data to obtain a content recommendation network model; obtain a recommendation candidate set generated by the server; in response to an on-terminal reordering trigger operation, reorder the recommendation candidate set based on the content recommendation network model to obtain and display content recommendation results.

[0199] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0200] As can be seen from the above, this embodiment is applied to the terminal, and the terminal can obtain the initial content recommendation network model sent by the server, and the initial content recommendation network model is trained on the server; collect object behavior data; adjust the initial content recommendation network model based on the object behavior data to obtain the content recommendation network model; obtain the recommendation candidate set generated by the server; in response to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model to obtain and display the content recommendation result. The server can obtain the basic content recommendation network model, and use the basic sample to train the basic content recommendation network model to obtain the initial content recommendation network model; obtain object feature data from the terminal; update the initial content recommendation network model based on the object feature data; send the initial content recommendation network model to the terminal so that the terminal adjusts the initial content recommendation network model, thereby performing content recommendation in real time on the terminal. This application proposes a new end-cloud collaboration framework, which upgrades and improves the existing end-cloud collaboration mode, and fully utilizes the global user and content information in the cloud by pre-training the initial content recommendation network model in the cloud. In addition, the present application also utilizes the terminal's deep model reasoning capabilities to pre-position real-time reasoning calculations and decision-making to the terminal. The terminal uses the collected real-time object behavior data to fine-tune the network model, and then performs data reasoning and decision-making on the terminal side in real time, thereby reducing end-cloud interaction, so that the end-to-end response time can be reduced to hundreds of milliseconds, improving the overall cost-effectiveness, and avoiding excessive consumption of terminal resources and data sparseness as the model becomes larger, causing the network model to enter the local optimum. This can also solve problems such as the inability to rearrange low-end models and insufficient end re-arrangement candidate sets.

[0201] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0202] To this end, an embodiment of the present application provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps in any content recommendation method provided in the embodiment of the present application. For example, the instructions can execute the following steps:

[0203] The terminal can obtain an initial content recommendation network model sent by a server, wherein the initial content recommendation network model is trained on the server; collect object behavior data; adjust the initial content recommendation network model based on the object behavior data to obtain a content recommendation network model; obtain a recommendation candidate set generated by the server; in response to an on-terminal reordering trigger operation, reorder the recommendation candidate set based on the content recommendation network model to obtain and display content recommendation results.

[0204] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0205] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0206] Since the instructions stored in the computer-readable storage medium can execute the steps in any content recommendation method provided in the embodiments of the present application, the beneficial effects that can be achieved by any content recommendation method provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0207] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program including computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method provided in various optional implementations of the above-mentioned content recommendation aspect.

[0208] The above is a detailed introduction to a content recommendation method and related equipment provided in an embodiment of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A content recommendation method, characterized in that: Applied to terminals, including: Obtaining an initial content recommendation network model sent by a server, wherein the initial content recommendation network model is trained on the server; Collecting subject behavior data; Based on the object behavior data, adjusting the initial content recommendation network model to obtain a content recommendation network model; Obtaining a recommendation candidate set generated by the server; In response to the on-end reordering trigger operation, the recommendation candidate set is reordered based on the content recommendation network model to obtain and display a content recommendation result.

2. The content recommendation method according to claim 1, characterized in that: The step of adjusting the initial content recommendation network model based on the object behavior data to obtain the content recommendation network model includes: storing the object behavior data in an object behavior library; When the object behavior data in the object behavior library are accumulated to meet a model adjustment condition, a model adjustment request is triggered; Based on the model adjustment request, the initial content recommendation network model is adjusted to obtain a content recommendation network model.

3. The content recommendation method according to claim 2, characterized in that: The step of adjusting the initial content recommendation network model based on the model adjustment request to obtain a content recommendation network model includes: Based on the model adjustment request, adjusting the initial content recommendation network model to obtain an adjusted network model; Performing model evaluation on the adjusted network model to obtain an evaluation result, and determining a content recommendation network model based on the evaluation result.

4. The content recommendation method according to claim 3, characterized in that: The step of adjusting the initial content recommendation network model based on the model adjustment request to obtain an adjusted network model includes: Sampling the object behavior library to obtain a sample set; Based on the sample set, the initial content recommendation network model is adjusted to obtain an adjusted network model.

5. The content recommendation method according to claim 3, characterized in that: The step of performing a model evaluation on the adjusted network model to obtain an evaluation result, and determining a content recommendation network model based on the evaluation result, includes: Performing test set sampling on the object behavior library to obtain a test set; Get the basic network model; Based on the test set, performing model evaluation on the adjusted network model and the basic network model to obtain an evaluation result; Based on the evaluation result, a content recommendation network model is determined from the adjusted network model and the basic network model.

6. The content recommendation method according to claim 5, characterized in that: The step of performing model evaluation on the adjusted network model and the basic network model based on the test set to obtain evaluation results includes: Obtain a first loss result of the adjusted network model on the test set, and obtain a second loss result of the basic network model on the test set; Based on the first loss result and the second loss result, a model evaluation is performed on the trained network model and the basic network model to obtain an evaluation result.

7. A content recommendation method, characterized in that: Applicable to servers, including: Obtaining a basic content recommendation network model, and using basic samples to perform model training on the basic content recommendation network model to obtain an initial content recommendation network model; Obtaining object feature data from a terminal; Based on the object feature data, updating the initial content recommendation network model; The initial content recommendation network model is sent to the terminal so that the terminal can adjust the initial content recommendation network model, thereby performing content recommendation at the terminal in real time.

8. The content recommendation method according to claim 7, characterized in that: The sending the initial content recommendation network model to the terminal includes: Performing model compression on the initial content recommendation network model to obtain a compressed initial content recommendation network model; The compressed initial content recommendation network model is split into models, and the model is sent to the terminal using a content distribution network.

9. The content recommendation method according to claim 7, characterized in that: The method further comprises: Based on the initial content recommendation network model, obtaining network model parameters; Based on the network model parameters, generating a recommendation candidate set; The recommendation candidate set is sent to the terminal so that the terminal displays content recommendation results based on the recommendation candidate set.

10. A content recommendation device, characterized in that: Applied to terminals, including: A first acquisition unit is used to acquire an initial content recommendation network model sent by a server, wherein the initial content recommendation network model is trained on the server; A collection unit, used for collecting object behavior data; An adjustment unit, configured to adjust the initial content recommendation network model based on the object behavior data to obtain a content recommendation network model; A second acquisition unit, used to acquire a recommendation candidate set generated by the server; A display unit is used to respond to the rearrangement trigger operation on the terminal, rearrange the recommendation candidate set based on the content recommendation network model, and obtain and display the content recommendation result.

11. A content recommendation device, characterized in that: Applicable to servers, including: A training unit, used to obtain a basic content recommendation network model, and perform model training on the basic content recommendation network model using basic samples to obtain an initial content recommendation network model; A third acquisition unit, used to acquire object feature data from a terminal; An updating unit, configured to update the initial content recommendation network model based on the object feature data; The sending unit is used to send the initial content recommendation network model to the terminal so that the terminal can adjust the initial content recommendation network model, thereby performing content recommendation in real time at the terminal.

12. An electronic device, characterized in that: It comprises a memory and a processor; the memory stores an application program, and the processor is used to run the application program in the memory to perform the operations in the content recommendation method according to any one of claims 1 to 6.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the content recommendation method according to any one of claims 1 to 6.

14. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps of the content recommendation method according to any one of claims 1 to 6 are implemented.

Citation Information

Cited By

  • Front-end page dynamic recommendation method and device, storage medium and computer equipment

    CN121092774A