Target recommendation method and device

Abstract

The present application relates to the field of information processing technology and provides a target recommendation method and device. The method comprises: bidirectionally connecting a large language model and a click prediction model to obtain a target recommendation model, and training the target recommendation model based on a target recommendation task; obtaining target user information of a target user; processing the target user information according to a prompt word mechanism, and inputting the target user information processed according to the prompt word mechanism into a trained target recommendation model; converting the target user information processed according to the prompt word mechanism into a target user feature vector using a large language model; determining a recommendation candidate sequence based on the target user feature vector using a click prediction model; screening the targets to be recommended in the recommendation candidate sequence using the large language model to obtain a recommendation sequence; and recommending the target to the target user according to the recommendation sequence.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a target recommendation method and device. Background Art

[0002] In the information age, users are flooded with a wide variety of information, including stock information, national policies, entertainment news, sports news, and more. However, users prefer to spend their time on information they find interesting and ignore irrelevant information. Therefore, recommendation models are widely deployed to automatically infer people's preferences and provide high-quality recommendation services. However, existing recommendation models often recommend information that deviates from user preferences, wasting user time and leaving the user experience in need of further improvement. Summary of the Invention

[0003] In view of this, embodiments of the present application provide a target recommendation method, device, electronic device, and computer-readable storage medium to solve the problem of inaccurate recommendation information of recommendation models in the prior art.

[0004] In a first aspect of an embodiment of the present application, a target recommendation method is provided, comprising: bidirectionally connecting a large language model and a click prediction model to obtain a target recommendation model, and training the target recommendation model based on a target recommendation task; obtaining target user information of a target user, wherein the target user information comprises: search term history information, personal basic information, historical conversation information, and historical interaction behavior information; processing the target user information according to a prompt word mechanism, and inputting the target user information processed according to the prompt word mechanism into a trained target recommendation model: converting the target user information processed according to the prompt word mechanism into a target user feature vector through a large language model; determining a recommendation candidate sequence based on the target user feature vector through a click prediction model, wherein the recommendation candidate sequence comprises multiple targets to be recommended; screening the targets to be recommended in the recommendation candidate sequence through a large language model to obtain a recommendation sequence, wherein the recommendation sequence comprises multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence; and recommending targets to the target user according to the recommendation sequence.

[0005] According to a second aspect of an embodiment of the present application, a target recommendation device is provided, comprising: a training module configured to bidirectionally connect a large language model and a click prediction model to obtain a target recommendation model, and train the target recommendation model based on a target recommendation task; an acquisition module configured to obtain target user information of a target user, wherein the target user information includes: search term history information, personal basic information, historical conversation information, and historical interaction behavior information; a processing module configured to process the target user information according to a prompt word mechanism, and input the target user information processed according to the prompt word mechanism into the trained target recommendation model; and a conversion module configured to obtain the target user information of a target user, wherein the target user information includes: search term history information, personal basic information, historical conversation information, and historical interaction behavior information. It is configured to convert the target user information processed according to the prompt word mechanism into a target user feature vector through a large language model; the determination module is configured to determine a recommendation candidate sequence based on the target user feature vector through a click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended; the screening module is configured to screen the targets to be recommended in the recommendation candidate sequence through the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence; the recommendation module is configured to recommend targets to the target user according to the recommendation sequence.

[0006] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0007] According to a fourth aspect of the embodiments of the present application, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps of the above method are implemented.

[0008] Compared with the prior art, the embodiments of the present application have the following advantages: because the embodiments of the present application obtain a target recommendation model by bidirectionally connecting a large language model and a click prediction model, and train the target recommendation model based on the target recommendation task; obtain target user information of the target user, wherein the target user information includes: search term history information, personal basic information, historical conversation information, and historical interaction behavior information; process the target user information according to the prompt word mechanism, and input the target user information processed according to the prompt word mechanism into the trained target recommendation model; convert the target user information processed according to the prompt word mechanism into a target user feature vector through the large language model; determine a recommendation candidate sequence based on the target user feature vector through the click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended; screen the targets to be recommended in the recommendation candidate sequence through the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence; and recommend targets to the target user according to the recommendation sequence. Therefore, the above technical means can solve the problem of inaccurate recommendation information of the recommendation model in the prior art, thereby improving the accuracy of the recommendation information recommended by the recommendation model and enhancing user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0010] Figure 1 This is a flowchart of a target recommendation method provided in an embodiment of the present application;

[0011] Figure 2 This is a flowchart of a target recommendation model training method provided in an embodiment of the present application;

[0012] Figure 3 This is a schematic diagram of the structure of a target recommendation device provided in an embodiment of the present application;

[0013] Figure 4 This is a structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0014] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0015] Figure 1 This is a flowchart of a target recommendation method provided in an embodiment of the present application. Figure 1 The target recommendation method can be executed by a computer or a server, or by software on a computer or a server. Figure 1 As shown, the target recommendation method includes:

[0016] S101, bidirectionally connect the large language model and the click prediction model to obtain a target recommendation model, and train the target recommendation model based on the target recommendation task;

[0017] S102, obtaining target user information of the target user, wherein the target user information includes: search term history information, personal basic information, historical conversation information, and historical interaction behavior information;

[0018] S103: Process the target user information according to the prompt word mechanism, and input the target user information processed according to the prompt word mechanism into the trained target recommendation model:

[0019] S104, converting the target user information processed according to the prompt word mechanism into a target user feature vector through a large language model;

[0020] S105, determining a recommendation candidate sequence based on the target user feature vector using a click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended;

[0021] S106, screening the targets to be recommended in the recommendation candidate sequence using the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence;

[0022] S107: Recommend the target to the target user according to the recommendation sequence.

[0023] The click prediction model is the CTR model (CTR, short for Click Through Rate). The large language model can use the ChatGPT model (Chat Generative Pre-trained Transformer). This embodiment of the application utilizes a large language model and a click prediction model to construct a target recommendation model. This can be understood as using the large language model to enhance the recommendation effectiveness of the click prediction model.

[0024] Basic personal information includes information such as the target user's age, gender, and hobbies. Historical search term information includes information about search terms used by the target user. Historical conversation information includes information about conversations between the target user and the large language model. Historical interaction behavior information includes information about the interaction between the target user and the recommended target. Interaction behaviors include collection, sharing, clicks, and reading time.

[0025] The prompt word mechanism is a mechanism called a prompt. By processing target user information according to the prompt word mechanism, instructions are provided for converting the target user information into a relevant embedding. Therefore, the target user information processed according to the prompt word mechanism contains the original target user information and instructions for converting the target user information into a relevant embedding. Therefore, a large language model can be used to convert the target user information processed according to the prompt word mechanism into a target user feature vector, which is the relevant embedding vector. This process is an application of the prompt mechanism, a common method, so it will not be detailed here.

[0026] In this embodiment, the target user information processed according to the prompt word mechanism is fed into a trained target recommendation model, that is, into a large language model, which outputs a target user feature vector. The target user feature vector is then fed into a click prediction model, which outputs a recommendation candidate sequence. The recommendation candidate sequence is then fed into the large language model, which outputs a recommendation sequence. This means that the large language model is used to enhance the recommendation effect of the click prediction model, thereby improving the accuracy of the recommendation model's recommendations.

[0027] According to the technical solution provided by the embodiment of the present application, the large language model and the click prediction model are bidirectionally connected to obtain a target recommendation model, and the target recommendation model is trained based on the target recommendation task; the target user information of the target user is obtained, wherein the target user information includes: search term history information, personal basic information, historical conversation information and historical interaction behavior information; the target user information is processed according to the prompt word mechanism, and the target user information processed according to the prompt word mechanism is input into the trained target recommendation model: the target user information processed according to the prompt word mechanism is converted into a target user feature vector by the large language model; based on the target user feature vector, a recommendation candidate sequence is determined by the click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended; the targets to be recommended in the recommendation candidate sequence are screened by the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence; and targets are recommended to the target user according to the recommendation sequence. Therefore, the above technical means can solve the problem of inaccurate recommendation information of the recommendation model in the existing technology, thereby improving the accuracy of the recommendation information recommended by the recommendation model and improving user satisfaction.

[0028] Furthermore, after inputting the target user information processed according to the prompt word mechanism into the trained target recommendation model, the method also includes: using the large language model as a vector generator to convert the search word history information, personal basic information, historical conversation information and historical interaction behavior information processed according to the prompt word mechanism into search word history feature vectors, personal basic feature vectors, historical conversation feature vectors and historical interaction behavior feature vectors respectively through the large language model; based on the search word history feature vectors, personal basic feature vectors, historical conversation feature vectors and historical interaction behavior feature vectors, determining the recommendation candidate sequence through the click prediction model; wherein, the target user feature vector includes: search word history feature vectors, personal basic feature vectors, historical conversation feature vectors and historical interaction behavior feature vectors.

[0029] In some embodiments: after the large language model converts and obtains the target user feature vector, the large language model is controlled to transmit the target user feature vector to the click prediction model through a unidirectional connection from the large language model to the click prediction model; after the click prediction model determines a recommendation candidate sequence based on the target user feature vector, the click prediction model is controlled to transmit the recommendation candidate sequence to the large language model through a unidirectional connection from the click prediction model to the large language model; after the large language model screens and obtains a recommendation sequence from the recommendation candidate sequence, the recommendation sequence is used as the output of the target recommendation model; the bidirectional connection between the large language model and the click prediction model includes: a unidirectional connection from the large language model to the click prediction model and a unidirectional connection from the click prediction model to the large language model.

[0030] In other words, the internal working process of the large language model is: large language model, click prediction model, large language model. Within the large language model, there is a unidirectional connection from the large language model to the click prediction model, and a unidirectional connection from the click prediction model to the large language model. This connection is called a bidirectional connection between the language model and the click prediction model.

[0031] Furthermore, in the process of obtaining a recommendation sequence by screening the targets to be recommended in the recommendation candidate sequence through the large language model: the targets to be recommended in the recommendation sequence are also sorted through the large language model, wherein the targets to be recommended that are closer to the front in the recommendation sequence are recommended to the target user earlier; and the large language model is also used to generate recommendation reasons for the targets to be recommended in the recommendation sequence.

[0032] The large language model can realize multiple functions or roles, and can sort the recommended targets in the recommendation sequence according to the knowledge that the large language model itself has learned, and generate recommendation reasons for the recommended targets in the recommendation sequence.

[0033] Furthermore, the target recommendation model is trained based on the target recommendation task, including: simultaneously training the target recommendation model based on multiple recommendation tasks, wherein the target recommendation task includes multiple recommendation tasks, each recommendation task corresponds to a recommendation field, and the trained target recommendation model can be used for target recommendation in various recommendation fields.

[0034] For example, in an online shopping scenario, the recommendation task is a product recommendation task, and the recommendation target or the target to be recommended is the product; in a news reading scenario, the recommendation task is a text recommendation task, and the recommendation target or the target to be recommended is the text; in a video watching scenario, the recommendation task is a video recommendation task, and the recommendation target or the target to be recommended is the video; in a music listening scenario, the recommendation task is a music recommendation task, and the recommendation target or the target to be recommended is music; in a wallpaper recommendation scenario, the recommendation task is a wallpaper recommendation task, and the recommendation target or the target to be recommended is the wallpaper, etc.

[0035] Furthermore, the target recommendation model is trained based on the target recommendation task, including: sorting multiple recommendation tasks in a preset order, wherein the target recommendation task includes multiple recommendation tasks, each recommendation task corresponds to a recommendation field, and the preset order is pure text recommendation task, pure audio recommendation task, pure image recommendation task and complex information recommendation task; the target recommendation model is trained in sequence according to the sorted multiple recommendation tasks, wherein the target recommendation model is trained in sequence using the transfer learning method, and the trained target recommendation model can be used for target recommendation in various recommendation fields.

[0036] For example, there are four recommendation tasks: news recommendation, movie recommendation, music recommendation, and wallpaper recommendation. The news, music, wallpaper, and movie recommendation tasks correspond to the pure text, pure audio, pure image, and complex information recommendation tasks, respectively. The complex information recommendation task refers to a recommendation target that includes audio, image, and text information. The sorted recommendation tasks are, in order, news recommendation, music, wallpaper, and movie recommendation. Using transfer learning, the target recommendation model is trained sequentially. After training the target recommendation model for the news recommendation task, the knowledge learned by the target recommendation model is transferred to the training of the target recommendation model for the music recommendation task; after training the target recommendation model for the news and music recommendation tasks, the knowledge learned by the target recommendation model is transferred to the training of the target recommendation model for the wallpaper recommendation task; and after training the target recommendation model for the news, music, and wallpaper recommendation tasks, the knowledge learned by the target recommendation model is transferred to the training of the target recommendation model for the movie recommendation task.

[0037] Figure 2 This is a flow chart of a target recommendation model training method provided by an embodiment of the present application. Figure 2 Shown, including:

[0038] S201, bidirectionally connect the large language model and the click prediction model to obtain a target recommendation model;

[0039] S202, obtaining training data, wherein the training data includes user information of multiple users, the user information including: search term history information, personal basic information, historical conversation information, and historical interaction behavior information;

[0040] S203: Process each piece of user information according to the prompt word mechanism, and input each piece of user information processed according to the prompt word mechanism into the target recommendation model:

[0041] S204, converting each piece of user information processed according to the prompt word mechanism into a user feature vector corresponding to each user through a large language model;

[0042] S205, based on the user feature vector corresponding to each user, determine a recommendation candidate sequence corresponding to each user through a click prediction model, where the recommendation candidate sequence includes multiple targets to be recommended;

[0043] S206: Filter the recommended targets in the recommendation candidate sequence corresponding to each user using the large language model to obtain a recommendation sequence corresponding to each user, wherein the recommendation sequence includes multiple recommended targets, and the number of recommended targets in the recommendation sequence is less than the number of recommended targets in the recommendation candidate sequence;

[0044] S207 , training the target recommendation model based on the target recommendation task according to the target to be recommended and the label in the recommendation sequence corresponding to each user.

[0045] The label corresponding to each user is the recommended target of the user marked in advance.

[0046] Furthermore, the target recommendation model is trained based on the target recommendation task, including: converting user information in the training data into user feature vectors through a large language model, transmitting the user feature vectors to the click prediction model, and introducing a dropout mechanism in the click prediction model to enhance the cold start capability of the click prediction model.

[0047] All of the above optional technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here.

[0048] The following are device embodiments of the present application, which can be used to implement the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0049] Figure 3 Schematic diagram of a target recommendation device provided in an embodiment of the present application. Figure 3 As shown, the target recommendation device includes:

[0050] The training module 301 is configured to bidirectionally connect the large language model and the click prediction model to obtain a target recommendation model, and train the target recommendation model based on the target recommendation task;

[0051] The acquisition module 302 is configured to acquire target user information of the target user, wherein the target user information includes: search term history information, personal basic information, historical conversation information and historical interaction behavior information;

[0052] The processing module 303 is configured to process the target user information according to the prompt word mechanism and input the target user information processed according to the prompt word mechanism into the trained target recommendation model:

[0053] The conversion module 304 is configured to convert the target user information processed according to the prompt word mechanism into a target user feature vector through a large language model;

[0054] The determination module 305 is configured to determine a recommendation candidate sequence based on the target user feature vector and a click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended;

[0055] A screening module 306 is configured to screen the targets to be recommended in the recommendation candidate sequence using the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence;

[0056] The recommendation module 307 is configured to make target recommendations to the target user according to the recommendation sequence.

[0057] According to the technical solution provided by the embodiment of the present application, the large language model and the click prediction model are bidirectionally connected to obtain a target recommendation model, and the target recommendation model is trained based on the target recommendation task; the target user information of the target user is obtained, wherein the target user information includes: search term history information, personal basic information, historical conversation information and historical interaction behavior information; the target user information is processed according to the prompt word mechanism, and the target user information processed according to the prompt word mechanism is input into the trained target recommendation model: the target user information processed according to the prompt word mechanism is converted into a target user feature vector by the large language model; based on the target user feature vector, a recommendation candidate sequence is determined by the click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended; the targets to be recommended in the recommendation candidate sequence are screened by the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence; and targets are recommended to the target user according to the recommendation sequence. Therefore, the above technical means can solve the problem of inaccurate recommendation information of the recommendation model in the existing technology, thereby improving the accuracy of the recommendation information recommended by the recommendation model and improving user satisfaction.

[0058] Optionally, the conversion module 304 is also configured to use the large language model as a vector generator to convert the search term history information, personal basic information, historical conversation information and historical interaction behavior information processed according to the prompt word mechanism into search term history feature vectors, personal basic feature vectors, historical conversation feature vectors and historical interaction behavior feature vectors respectively through the large language model; based on the search term history feature vectors, personal basic feature vectors, historical conversation feature vectors and historical interaction behavior feature vectors, the recommendation candidate sequence is determined through the click prediction model; wherein, the target user feature vector includes: search term history feature vectors, personal basic feature vectors, historical conversation feature vectors and historical interaction behavior feature vectors.

[0059] Optionally, the conversion module 304 is further configured to, after the large language model converts the target user feature vector, control the large language model to transmit the target user feature vector to the click prediction model through a unidirectional connection from the large language model to the click prediction model; after the click prediction model determines the recommendation candidate sequence based on the target user feature vector, control the click prediction model to transmit the recommendation candidate sequence to the large language model through a unidirectional connection from the click prediction model to the large language model; after the large language model screens the recommendation sequence from the recommendation candidate sequence, use the recommendation sequence as the output of the target recommendation model; the bidirectional connection between the large language model and the click prediction model includes: a unidirectional connection from the large language model to the click prediction model and a unidirectional connection from the click prediction model to the large language model.

[0060] Optionally, the screening module 306 is also configured to, in the process of screening the to-be-recommended targets in the recommendation candidate sequence by the large language model to obtain the recommendation sequence: also sort the to-be-recommended targets in the recommendation sequence by the large language model, wherein the to-be-recommended targets that are closer to the front in the recommendation sequence are recommended to the target user earlier; and also generate recommendation reasons for the to-be-recommended targets in the recommendation sequence by the large language model.

[0061] Optionally, the training module 301 is also configured to simultaneously train the target recommendation model based on multiple recommendation tasks, wherein the target recommendation task includes multiple recommendation tasks, each recommendation task corresponds to a recommendation field, and the trained target recommendation model can be used for target recommendation in various recommendation fields.

[0062] Optionally, the training module 301 is also configured to sort multiple recommendation tasks in a preset order, wherein the target recommendation task includes multiple recommendation tasks, each recommendation task corresponds to a recommendation field, and the preset order is pure text recommendation task, pure audio recommendation task, pure image recommendation task and complex information recommendation task; the target recommendation model is trained in sequence according to the sorted multiple recommendation tasks, wherein the target recommendation model is trained in sequence using the transfer learning method, and the trained target recommendation model can be used for target recommendation in various recommendation fields.

[0063] Optionally, the training module 301 is further configured to bidirectionally connect the large language model and the click prediction model to obtain a target recommendation model; obtain training data, wherein the training data includes user information of multiple users, and the user information includes: search term history information, personal basic information, historical conversation information and historical interaction behavior information; process each piece of user information according to the prompt word mechanism, and input each piece of user information processed according to the prompt word mechanism into the target recommendation model: convert each piece of user information processed according to the prompt word mechanism into a user feature vector corresponding to each user through the large language model; based on the user feature vector corresponding to each user, determine the recommendation candidate sequence corresponding to each user through the click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended; screen the targets to be recommended in the recommendation candidate sequence corresponding to each user through the large language model to obtain a recommendation sequence corresponding to each user, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence; train the target recommendation model based on the target recommendation task according to the targets to be recommended and labels in the recommendation sequence corresponding to each user.

[0064] Optionally, the training module 301 is also configured to convert user information in the training data into user feature vectors through a large language model, transmit the user feature vectors to the click prediction model, and introduce a dropout mechanism into the click prediction model to enhance the cold start capability of the click prediction model.

[0065] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0066] Figure 4 Schematic diagram of the electronic device 4 provided in the embodiment of the present application. Figure 4 As shown, the electronic device 4 of this embodiment includes: a processor 401, a memory 402, and a computer program 403 stored in the memory 402 and executable by the processor 401. When the processor 401 executes the computer program 403, the steps of the above-described method embodiments are implemented. Alternatively, when the processor 401 executes the computer program 403, the functions of the modules / units in the above-described device embodiments are implemented.

[0067] The electronic device 4 may be a desktop computer, a notebook, a PDA, a cloud server, or other electronic device. The electronic device 4 may include but is not limited to a processor 401 and a memory 402. Those skilled in the art will appreciate that Figure 4 This is merely an example of the electronic device 4 and does not limit the electronic device 4 . The electronic device 4 may include more or fewer components than shown in the figure, or different components.

[0068] The processor 401 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0069] Memory 402 can be an internal storage unit of electronic device 4, such as a hard drive or memory of electronic device 4. Memory 402 can also be an external storage device of electronic device 4, such as a plug-in hard drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. Memory 402 can also include both an internal storage unit of electronic device 4 and an external storage device. Memory 402 is used to store computer programs and other programs and data required by the electronic device.

[0070] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0071] If the integrated module / unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program can include computer program code, which can be in source code form, object code form, executable file, or some intermediate form. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, removable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signals, telecommunications signals, and software distribution media. It should be noted that the content included in computer-readable media can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, based on legislation and patent practice, computer-readable media does not include electric carrier signals and telecommunications signals.

[0072] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A target recommendation method, characterized in that: include: bidirectionally connecting the large language model and the click prediction model to obtain a target recommendation model, and training the target recommendation model based on the target recommendation task; Obtain target user information of the target user, wherein the target user information includes: search term history information, personal basic information, historical conversation information, and historical interaction behavior information; The target user information is processed according to the prompt word mechanism, and the target user information processed according to the prompt word mechanism is input into the trained target recommendation model: Converting the target user information processed according to the prompt word mechanism into a target user feature vector through the large language model; Based on the target user feature vector, determining a recommendation candidate sequence through the click prediction model, wherein the recommendation candidate sequence includes multiple targets to be recommended; Filtering the targets to be recommended in the recommendation candidate sequence using the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of the targets to be recommended in the recommendation sequence is less than the number of the targets to be recommended in the recommendation candidate sequence; Target recommendations are made to the target user according to the recommendation sequence.

2. The method according to claim 1, characterized in that After inputting the target user information processed according to the prompt word mechanism into the trained target recommendation model, the method further includes: The large language model is used as a vector generator to convert the search term history information, personal basic information, historical conversation information, and historical interaction behavior information processed according to the prompt word mechanism into a search term history feature vector, a personal basic feature vector, a historical conversation feature vector, and a historical interaction behavior feature vector, respectively, through the large language model; Determining the recommendation candidate sequence by the click prediction model based on the search term history feature vector, the personal basic feature vector, the historical conversation feature vector, and the historical interaction behavior feature vector; The target user feature vector includes: the search term history feature vector, the personal basic feature vector, the historical conversation feature vector, and the historical interaction behavior feature vector.

3. The method according to claim 1, characterized in that The method further comprises: After the large language model converts the target user feature vector, controlling the large language model to transmit the target user feature vector to the click prediction model through a unidirectional connection from the large language model to the click prediction model; After the click prediction model determines the recommendation candidate sequence based on the target user feature vector, controlling the click prediction model to transmit the recommendation candidate sequence to the large language model via a unidirectional connection from the click prediction model to the large language model; After the large language model screens the recommendation sequence from the recommendation candidate sequences to obtain the recommendation sequence, the recommendation sequence is used as the output of the target recommendation model; The bidirectional connection between the large language model and the click prediction model includes: a unidirectional connection from the large language model to the click prediction model and a unidirectional connection from the click prediction model to the large language model.

4. The method according to claim 1, wherein In the process of obtaining the recommendation sequence by screening the to-be-recommended targets in the recommendation candidate sequence using the large language model: The large language model is also used to sort the targets to be recommended in the recommendation sequence, wherein the targets to be recommended that are closer to the front in the recommendation sequence are recommended to the target user earlier; The large language model is also used to generate recommendation reasons for the target to be recommended in the recommendation sequence.

5. The method according to claim 1, wherein Training the target recommendation model based on the target recommendation task includes: The target recommendation model is trained simultaneously based on multiple recommendation tasks, wherein the target recommendation task includes multiple recommendation tasks, each recommendation task corresponds to a recommendation field, and the trained target recommendation model can be used for target recommendation in various recommendation fields.

6. The method according to claim 1, characterized in that Training the target recommendation model based on the target recommendation task includes: Sorting multiple recommendation tasks in a preset order, wherein the target recommendation task includes multiple recommendation tasks, each recommendation task corresponds to a recommendation field, and the preset order is pure text recommendation task, pure audio recommendation task, pure image recommendation task, and complex information recommendation task; The target recommendation model is trained in sequence according to the sorted multiple recommendation tasks, wherein the target recommendation model is trained in sequence using a transfer learning method, and the trained target recommendation model can be used for target recommendation in various recommendation fields.

7. The method according to claim 1, characterized in that Training the target recommendation model based on the target recommendation task includes: The user information in the training data is converted into a user feature vector through the large language model, and the user feature vector is transmitted to the click prediction model. The dropout mechanism is introduced into the click prediction model to enhance the cold start capability of the click prediction model.

8. A target recommendation device, characterized in that: include: a training module configured to bidirectionally connect the large language model and the click prediction model to obtain a target recommendation model, and train the target recommendation model based on the target recommendation task; An acquisition module is configured to acquire target user information of a target user, wherein the target user information includes: search term history information, personal basic information, historical conversation information, and historical interaction behavior information; A processing module is configured to process the target user information according to a prompt word mechanism, and input the target user information processed according to the prompt word mechanism into the trained target recommendation model: A conversion module configured to convert the target user information processed according to the prompt word mechanism into a target user feature vector through the large language model; a determination module configured to determine a recommendation candidate sequence based on the target user feature vector and through the click prediction model, wherein the recommendation candidate sequence includes a plurality of targets to be recommended; a screening module configured to screen the targets to be recommended in the recommendation candidate sequence using the large language model to obtain a recommendation sequence, wherein the recommendation sequence includes multiple targets to be recommended, and the number of targets to be recommended in the recommendation sequence is less than the number of targets to be recommended in the recommendation candidate sequence; The recommendation module is configured to make target recommendations to the target user according to the recommendation sequence.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images