Network model training method, recommendation method and system

Abstract

The invention provides a network model training method, recommendation method and system, and the training method comprises the steps: obtaining a sample data set, obtaining a first prediction recommendation result corresponding to the sample data set based on a large language model, obtaining a second prediction recommendation result corresponding to the sample data set based on the large language model and a recommendation model, the parameters of the preset network are updated with minimization of a first loss function between the first prediction recommendation result and a preset first label and minimization of a second loss function between the second prediction recommendation result and a preset second label as training targets, a target network model is obtained, and the preset first label is determined based on the rich intention of the sample user. The preset first label of the large language model is determined by adopting the rich intention, and the ability of the large language model to distinguish the deeper user intention of the sample user can be trained, so that the target network model obtained by training can accurately and effectively make recommendation for the user.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a training method, recommendation method, and system for a network model. Background Art

[0002] With the development of Internet technology, recommendation services have been widely used in different fields and scenarios.

[0003] In the related art, a network model is pre-trained to provide recommendation services to users based on the network model. With the development of artificial intelligence, the network model can be a large language model (LLMs). For example, a large language model is trained based on a sample data set to train the recommendation capability of the large language model.

[0004] However, due to challenges such as limited data processing capabilities and slow inference speed, using large language models as independent network models for recommendation has been criticized.

[0005] It is worth noting that the contents of the above-mentioned related technologies are merely information known to the inventor personally, which does not mean that the above-mentioned information has entered the public domain before the filing date of this specification, nor does it mean that it can become the prior art of this specification. Summary of the invention

[0006] This specification provides a network model training method, recommendation method, and system to avoid at least one of the above-mentioned technical problems.

[0007] In a first aspect, the present specification provides a method for training a network model, the method being applied to a preset network, the preset network including a large language model and a recommendation model; the method comprising:

[0008] Obtaining a sample data set, wherein the sample data set includes sample behavior data of a sample user;

[0009] Obtaining a first predicted recommendation result corresponding to the sample data set based on the large language model;

[0010] Obtaining a second predicted recommendation result corresponding to the sample data set based on the large language model and the recommendation model; and

[0011] The parameters of the preset network are updated with the training objectives of minimizing a first loss function between the first predicted recommendation result and a preset first label, and a second loss function between the second predicted recommendation result and a preset second label to obtain a target network model, wherein the preset first label is determined based on the rich intent of the sample user, and the rich intent is used to characterize deeper needs and preferences than the explicit intention of the sample user.

[0012] In a second aspect, this specification provides a recommendation method, comprising:

[0013] In response to a recommendation request corresponding to a target user, obtaining historical behavior data of the target user;

[0014] The historical behavior data is input into a target network model, and a target recommendation result corresponding to the recommendation request is output, wherein the target network model is trained based on the method described in the first aspect.

[0015] In a third aspect, the present specification provides a network model training system, wherein the training system is applied to a preset network, the preset network includes a large language model and a recommendation model; the training system includes:

[0016] At least one storage medium storing at least one instruction set for training the preset network;

[0017] At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, it reads the at least one instruction set and executes the method as described in the first aspect according to the instructions of the at least one instruction set.

[0018] In a fourth aspect, this specification provides a recommendation system, including:

[0019] at least one storage medium storing at least one set of instructions for making recommendations;

[0020] At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, it reads the at least one instruction set and executes the method as described in the second aspect according to the instructions of the at least one instruction set.

[0021] In a fifth aspect, the present specification provides a computer-readable non-temporary storage medium, wherein the computer-readable non-temporary storage medium stores at least one instruction set, and the at least one instruction set is executed by at least one processor to implement the method described in the first aspect and the second aspect.

[0022] It can be seen from the above technical solutions that the training method, recommendation method, and system of the network model provided in this specification include: obtaining a sample data set, wherein the sample data set includes sample behavior data of sample users, obtaining a first predicted recommendation result corresponding to the sample data set based on a large language model, and obtaining a second predicted recommendation result corresponding to the sample data set based on the large language model and the recommendation model, and updating the parameters of the preset network with the first loss function between the first predicted recommendation result and the preset first label and the second loss function between the second predicted recommendation result and the preset second label as the training target to obtain a target network model, wherein the preset first label is determined based on the rich intent of the sample user, and the rich intent is used to characterize needs and preferences that are deeper than the explicit intent of the sample user. By using the rich intent to determine the preset first label of the large language model, the ability of the large language model to distinguish the deeper user intent of the sample user is trained, thereby training the recommendation model to determine the recommendation result based on the deeper user intent of the sample user. In addition, since the rich intent is used to determine the first preset label, rather than directly as a feature, the risk of feature leakage can be avoided. Moreover, under end-to-end learning, through gradient backpropagation, the large language model can be guided by the recommendation model's recommendation capabilities, reducing the reliance on highly complex and resource-intensive large language models. Accordingly, in the application stage, when the recommendation system makes recommendations based on the target network model, the recommendation results can be highly consistent with the user's deep-level needs and preferences, thereby improving the accuracy and reliability of the recommendations and improving the user experience.

[0023] The training method, recommendation method, and other functions of the network model provided in this specification will be partially listed in the following description. The creative aspects of the training method, recommendation method, and system of the network model provided in this specification can be fully explained by practicing or using the methods, devices, and combinations described in the following detailed examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of this specification, 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 this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 A schematic diagram of an application scenario of the recommended method provided in the embodiments of this specification;

[0026] Figure 2 A schematic diagram of the structure of the training method and recommendation system provided in the embodiments of this specification;

[0027] Figure 3A flowchart of a training method provided in an embodiment of this specification;

[0028] Figure 4 A schematic diagram of the principle of the training method provided in the embodiment of this specification;

[0029] Figure 5 A schematic diagram of the distribution obtained by using the Zero-shot method provided in the embodiments of this specification;

[0030] Figure 6 A schematic diagram of the distribution obtained by using the GRec method provided in the embodiments of this specification;

[0031] Figure 7 The embodiment of this specification provides the use of EGRec label Schematic diagram of the distribution obtained by the method;

[0032] Figure 8 This is a schematic diagram of the distribution obtained by using the EGRec method provided in the embodiments of this specification. DETAILED DESCRIPTION

[0033] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this specification. Instead, they are merely examples of devices and methods consistent with some aspects of this specification as detailed in the appended claims.

[0034] It should be understood that the terms "including" and "having" and any variations thereof in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device comprising a series of components is not necessarily limited to those components explicitly listed, but may include other components not explicitly listed or inherent to these products or devices.

[0035] In the embodiments of this specification, the term "and / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0036] The term "plurality" in the embodiments of the present specification refers to two or more than two, and other quantifiers are similar thereto.

[0037] The terms "first", "second", "third", etc. in this specification are used to distinguish similar or similar objects or entities, and do not necessarily mean to limit a specific order or sequence, unless otherwise indicated. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, for example, they can be implemented in an order other than those given in the diagrams or descriptions of the embodiments of this specification.

[0038] The term "unit / module" as used in this specification refers to any known or later developed hardware, software, firmware, artificial intelligence, fuzzy logic or combination of hardware and / or software code that is capable of performing the functions associated with that element.

[0039] To facilitate readers' understanding of this specification, the application scenarios of this specification are now introduced.

[0040] The technical solution provided in this specification is applicable to a recommendation scenario of recommending a target object to a user. The target object may be a commodity or a service. For example, the target object may be a book or a video.

[0041] Take books as an example:

[0042] When a user logs in to the e-book reading platform, a recommendation request may be triggered. Accordingly, the e-book reading platform may execute the technical solution provided in this specification to determine and output books recommended to the user.

[0043] Take the target object as video as an example:

[0044] When a user logs in to a video application (Application, APP), he can enter a search term in the search box of the video APP to trigger a recommendation request. Accordingly, the video APP can execute the technical solution provided in this specification to determine and output videos recommended to the user.

[0045] It is worth noting that the above examples are only used to illustrate the recommendation scenarios to which the technical solutions of this specification can be applied, and should not be understood as limiting the recommendation scenarios.

[0046] In practical applications, network models are usually pre-trained to provide recommendation services to users based on the network models. For example, with the advancement of natural language processing, especially the development of large language models, large language models are widely used to extract semantic information from text data to enhance the effect of recommendations. However, due to challenges such as limited data processing capabilities and slow inference speed, using large language models as independent network models for recommendations has been criticized.

[0047] Therefore, we can use the large model recommendation (LLM for Recommendation, LLM4Rec) method to provide recommendation services to users. Among them, the large model recommendation mainly includes two parts: the large language model and the recommendation model. There are two main types of large model recommendations, one is the large language model embedding + recommendation model (CTRL), and the other is the large language model output text + recommendation model (KAR).

[0048] In the CTRL framework, the feature extraction capability of the large language model is mainly used for recommendation, that is, the large language model mainly acts as a feature extractor. For example, the large language model mainly extracts feature embeddings (embeddings), which can also be called feature vectors, or simply features, from user or project data. These feature embeddings are then used by the recommendation model to generate corresponding recommendation results. That is, the recommendation model uses feature embeddings to make recommendation predictions, obtains and outputs recommendation results recommended to users. Relatively speaking, CTRL utilizes the encoding capabilities of the large language model, but does not fully tap the generation potential of the large language model.

[0049] In the KAR framework, the large language model can generate user-preferred tokens based on user or project data. Tokens capture potential features through deep semantic mining. Correspondingly, the recommendation model uses a text encoder (such as BERT, etc.) to parse the hidden vector of the token for downstream recommendation tasks. In other words, compared to CTRL, KAR also utilizes the generation capability of the large language model.

[0050] However, in the above two training methods, both are based on explicit intent, where explicit intent can be understood as the needs or goals clearly expressed by the user, such as the occurrence of a click.

[0051] For example, during the training phase, the large language model may be trained first, and after the training of the large language model is completed, the recommendation model may be trained based on the large language model.

[0052] When training the large language model, the label of the large language model is determined based on the explicit intent. For example, if the explicit intent is that the click occurs, the label is 1; if the explicit intent is that the click does not occur, the label is 0. Accordingly, after the large language model determines the prediction result based on the input, the large language model can be fine-tuned based on the difference between the prediction result and the label.

[0053] However, the above-mentioned method based on explicit intent training cannot reveal deeper user intentions, and therefore cannot accurately capture user interests, resulting in low accuracy of recommendations in the application stage.

[0054] It is worth noting that the contents of the above-mentioned related technologies are merely information known to the inventor personally, which does not mean that the above-mentioned information has entered the public domain before the filing date of this specification, nor does it mean that it can become the prior art of this specification.

[0055] In order to avoid at least one of the above problems, this specification proposes a technical concept that is the result of creative labor: determining the label of a large language model based on a rich intent, so as to use the rich intent to perform end-to-end training on a preset network including a large language model and a recommendation model, so as to more effectively fine-tune the preset network so that it has a stronger ability to identify user interests. Among them, rich intent is a concept relative to explicit intent, which is used to characterize user intent at a deeper level than explicit intent. For example, rich intent can characterize implicit user interests or intentions. By enriching intent, the subtlety and depth of user interactions can be captured. In some embodiments, the rich intent can be a comment, a search query, or other forms of interactive information.

[0056] This specification mainly uses the recommendation method as an example to exemplify the application scenario. For the application scenario of the network model training method (hereinafter referred to as the training method), please refer to the description of the application scenario of the recommendation method, and this specification will not repeat it.

[0057] For example, Figure 1 Schematic diagram of application scenarios of the recommended method of the embodiment of this specification. Figure 1 As shown, the application scenario 100 may include a target user 101 , a client 102 , a server 103 , and a network 104 .

[0058] The target user 101 may be a user who triggers the recommendation service. For example, the target user 101 may perform a target operation on the client 102 to trigger the recommendation service. In some embodiments, the target operation may be a search operation. For example, the target user 101 may perform a search on the client 102 to trigger a recommendation request through the search operation, and obtain a recommendation service based on the recommendation request. In other embodiments, the target operation may be a login operation to trigger a recommendation request through the login operation. For example, the target user 101 may trigger a recommendation request after logging into the client 102, and obtain a recommendation service based on the recommendation request.

[0059] The client 102 may be an electronic device that provides interactive functions to the target user 101. For example, the client 102 may provide an interactive interface to the target user 101, and the target user 101 may perform interactive operations in the interactive interface. In some embodiments, the client 102 executes the recommendation method described in this specification in response to detecting a search operation or a login operation triggered by the target user 101. At this time, the client 102 may store data or instructions for executing the recommendation method described in this specification, and may execute or be used to execute the data or instructions. In some embodiments, the client 102 may include a hardware device with a data information processing function and the necessary program required to drive the hardware device to work, so as to execute the recommendation method described in this specification.

[0060] In some embodiments, the client 102 may include a mobile device, a tablet computer, a laptop computer, a built-in device of a motor vehicle, or the like, or any combination thereof. In some embodiments, the mobile device may include a smart home device, a smart mobile device, a virtual reality device, an augmented reality device, or the like, or any combination thereof. In some embodiments, the smart home device may include a smart TV, a desktop computer, or the like, or any combination thereof. In some embodiments, the smart mobile device may include a smart phone, a personal digital assistant, a gaming device, a navigation device, or the like, or any combination thereof. In some embodiments, the built-in device in a motor vehicle may include an onboard computer, an onboard TV, or the like.

[0061] In some embodiments, the client 102 may be installed with one or more APPs. The APP can provide the target user 101 with the ability and interface to interact with the outside world through the network 104. The APP includes but is not limited to: web browser APP programs, search APP programs, chat APP programs, shopping APP programs, video APP programs, financial management APP programs, instant messaging tools, email clients, social platform software, etc.

[0062] like Figure 1 As shown, the client 102 can be connected to the server 103 for communication. The server 103 can be connected to one client 102 for communication, or can be connected to multiple clients 102 for communication. In some embodiments, the client 102 can interact with the server 103 through the network 104 to receive or send messages, etc. For example, the client 102 can interact with the server 103 through the network 104 to send a recommendation request to the server 103.

[0063] The server 103 may be a server that provides various services. For example, the server 103 may be a cloud server or a local server. The server 103 may be connected to a client 102 for communication and receive data sent by the client 102, or may be connected to multiple clients 102 for communication and receive data sent by each client 102.

[0064] In some embodiments, the recommendation method described in this specification may be executed on the server 103. In this case, the server 103 may store data or instructions for executing the recommendation method described in this specification, and may execute or be used to execute the data or instructions. The server 103 may include a hardware device with data information processing functions and necessary programs required to drive the hardware device to work.

[0065] The network 104 is a medium for providing a communication connection between the client 102 and the server 103. The network 104 can facilitate the exchange of information or data. Figure 1 As shown, the client 102 and the server 103 can be connected to the network 104 respectively, and transmit information or data to each other through the network 104.

[0066] In some embodiments, the network 104 may be any type of wired or wireless network, or a combination thereof. For example, the network 104 may include a cable network, a wired network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network TM, a short-range wireless network (ZigBee TM), a near field communication (NFC) network, or a similar network.

[0067] In some embodiments, the network 104 may include one or more network access points. For example, the network 104 may include a wired or wireless network access point, such as a base station or an Internet exchange point, through which one or more components of the client 102 and the server 103 can connect to the network 104 to exchange data or information.

[0068] It is worth mentioning that Figure 1The number of clients 102, servers 103, and networks 104 in the embodiment is only illustrative. Any number of clients 102, servers 103, and networks 104 may be provided as required. The recommended method provided in this specification may be fully executed on the client 102, fully executed on the server 103, or partially executed on the client 102 and partially executed on the server 103.

[0069] That is to say, Figure 1 and for Figure 1 The above description is only used to illustrate the application scenarios to which the recommended method of this specification may be applicable, and cannot be understood as a limitation on the application scenarios. In addition, for the application scenarios to which the training method of this specification may be applicable, please refer to the above description of the application scenarios to which the recommended method may be applicable, which will not be repeated here.

[0070] It is worth noting that the subject used to execute the training method provided in this specification may be a training system. The subject used to execute the recommendation method provided in this specification may be a recommendation system. The training system and the recommendation system may be the same system or different systems. If the training system and the recommendation system are different systems, there may be a communication link between the training system and the recommendation system.

[0071] In some embodiments, after the training system obtains the target network model based on the training method provided in this specification, the target network model can be transmitted to the recommendation system through a communication link. The recommendation system can store the target network model. In response to the recommendation request, the recommendation system can determine the recommendation result corresponding to the recommendation request based on the target network model stored therein.

[0072] In other embodiments, after obtaining the target network model based on the training method provided in this specification, the training system can deploy the target network model locally and provide a calling interface for calling the target network model. In response to the recommendation request, the recommendation system can call the target network model based on the calling interface to determine the recommendation result corresponding to the recommendation request based on the target network model.

[0073] This specification mainly takes the recommendation system as an example to exemplify the hardware structure of the recommendation system. For the hardware structure of the training system, please refer to the recommendation system, and this specification will not repeat it.

[0074] For example, Figure 2The hardware structure diagram of a recommendation system 200 provided according to an embodiment of the present specification is shown. The recommendation system 200 can execute the recommendation method described in the present specification. The recommendation method is introduced in other parts of the present specification. When the recommendation method is executed on the client 102, the recommendation system 200 can be the client 102. When the recommendation method is executed on the server 103, the recommendation system 200 can be the server 103. When the recommendation method can be partially executed on the client 102 and partially executed on the server 103, the recommendation system 200 can be a system including the client 102 and the server 103.

[0075] like Figure 2 As shown, the recommendation system 200 may include at least one storage medium 203 and at least one processor 202. In some embodiments, the recommendation system 200 may further include a communication port 204 and an internal communication bus 201. The recommendation system 200 may further include an I / O component 205.

[0076] The internal communication bus 201 can connect different system components. For example, the internal communication bus 201 can connect the storage medium 203, the processor 202, the communication port 204 and the I / O component 205.

[0077] The I / O component 205 supports input / output between the recommender system 200 and other components.

[0078] The communication port 204 is used for data communication between the recommendation system 200 and the outside world. For example, the communication port 204 can be used for data communication between the recommendation system 200 and the network 104. The communication port 204 can be a wired communication port or a wireless communication port.

[0079] The storage medium 203 may include a data storage device. The data storage device may be a non-temporary storage medium or a temporary storage medium. For example, the data storage device may include one or more of a disk 2031, a read-only storage medium (ROM) 2032, or a random access storage medium (RAM) 2033. The storage medium 203 also includes at least one instruction set stored in the data storage device. The instruction set includes computer program code, and the computer program code may include a program, a routine, an object, a component, a data structure, a process, a module, etc. for executing the recommended method provided in this specification.

[0080] At least one processor 202 may be communicatively connected to at least one storage medium 203. At least one processor 202 is used to execute the at least one instruction set mentioned above. When the recommendation system 200 is running, at least one processor 202 reads the at least one instruction set and executes the recommendation method provided in this specification according to the instructions of the at least one instruction set. The processor 202 may execute all the steps included in the recommendation method. The processor 202 may be in the form of one or more processors. In some embodiments, the processor 202 may include one or more hardware processors, such as a microcontroller, a microprocessor, a reduced instruction set computer (RISC), an application-specific integrated circuit (ASIC), an application-specific instruction set processor (ASIP), a central processing unit (CPU), a graphics processing unit (GPU), a physical processing unit (PPU), a microcontroller unit, a digital signal processor (DSP), a field programmable gate array (FPGA), an advanced RISC machine (ARM), a programmable logic device (PLD), any circuit or processor capable of performing one or more functions, etc., or any combination thereof.

[0081] For illustration purposes only, only one processor 202 is shown in the recommendation system 200 in the accompanying drawings. However, it should be noted that the recommendation system 200 in this specification may also include multiple processors, and therefore, the operations and / or method steps disclosed in this specification may be performed by one processor or jointly by multiple processors. For example, if the processor 202 of the recommendation system 200 is described in this specification as performing steps A and B, it should be understood that steps A and B may also be performed jointly or separately by two different processors 202 (for example, the first processor performs step A, the second processor performs step B, or the first and second processors perform steps A and B together).

[0082] See also Figure 3 , Figure 3 The flowchart of the training method provided in the embodiment of this specification is shown in FIG. The training method is applied to a preset network, and the preset network includes a large language model and a recommendation model. Figure 3 As shown, the method includes the following S301 to S304:

[0083] S301: Obtain a sample data set, wherein the sample data set includes sample behavior data of sample users.

[0084] Exemplarily, the sample data set can be understood as a data set used to train a preset network to obtain a target network model that can accurately predict the recommendation results of the user's true intention.

[0085] Combined with the above analysis, it can be seen that the training method provided in this specification can be applied to different application scenarios, so that the recommendation model trained in different recommendation scenarios can recommend different objects to users in different recommendation scenarios. For example, in the book recommendation scenario, the training method provided in this specification can be trained to obtain a target network model for recommending books to users; for another example, in the video recommendation scenario, the training method provided in this specification can be trained to obtain a target network model for recommending videos to users; and so on, which will not be listed here one by one.

[0086] Accordingly, for different application scenarios, the sample behavior data is data generated by the behavior of sample users in different application scenarios. For example, in a book recommendation scenario, the sample behavior data may include data generated by the browsing behavior of sample users browsing books. For another example, in a video recommendation scenario, the sample behavior data may include data generated by the viewing behavior of sample users watching videos.

[0087] This embodiment does not limit the manner in which the prediction system obtains the sample data set. For example, in one example, the prediction system can be connected to the data acquisition system and receive the sample data set sent by the data acquisition system. In another example, the training system can provide a tool for loading data, and the user can transfer the sample data set to the training system through the tool for loading data. Among them, the tool for loading data can be an interface for connecting to an external device, such as an interface for connecting to other storage devices, through which the sample data set transmitted by the external device is obtained; the tool for loading data can also be a display device, such as the training system can input an interface for the data loading function on the display device, and the user can import the sample data set into the training system through the interface.

[0088] S302: Obtain a first predicted recommendation result corresponding to the sample data set based on the large language model.

[0089] Exemplarily, the sample data set is used as input data of the large language model so that the large language model predicts the recommendation result based on the sample data set. For the sake of distinction, the recommendation result predicted based on the large language model is referred to as the first predicted recommendation result.

[0090] S303: Obtain a second predicted recommendation result corresponding to the sample data set based on the large language model and the recommendation model.

[0091] In some embodiments, combined with the above analysis, it can be seen that the training system can be divided into two stages to obtain the target network model, one stage is to train the large language model, and the other stage is to train the recommendation model based on the trained large language model. And the training data of the two stages are both sample data sets.

[0092] In some other embodiments, the training system may also adopt an end-to-end training method to train the target network model. For example, the training system uses a sample data set as training data and trains the large language model and the recommendation model as a whole. For example, the training system simultaneously trains the large language model and the recommendation model by gradient backpropagation.

[0093] Therefore, this step can be understood as that the training system can use the sample data set as input data, and predict the sample data set based on the large language model and the recommendation model to obtain the second predicted recommendation result.

[0094] In some embodiments, the sample data set includes: sample text modal data that characterizes the sample behavior data from a text dimension, and sample table modal data that characterizes the sample behavior data from a table dimension.

[0095] For example, Figure 4 As shown, the sample behavior data can be represented by different modalities, such as text modality and table modality. Among them, the sample behavior data of the text modality is sample text modality data, and the sample behavior data of the table modality is table text modality data.

[0096] Accordingly, the above S302 may include: obtaining sample text features corresponding to the sample text modal data based on the large language model, and determining a first predicted recommendation result according to the sample text features.

[0097] Exemplarily, the training system can use sample text modal data of the text modality as input data of the large language model to determine the sample text features based on the semantic analysis and generation capabilities of the large language model, and make predictions based on the sample text features to obtain a first predicted recommendation result.

[0098] Combined with the above analysis, it can be seen that based on the enriched intent, the training system can enable the preset network to learn the ability to determine the implicit recommendation intent. Therefore, the sample text feature can also be called the implicit recommendation intent feature, or hidden feature.

[0099] In some embodiments, the sample behavior data includes sample historical browsing information (user history) and sample current item information (current item); the above-mentioned "obtaining a first predicted recommendation result based on a large language model" may include: obtaining a prompt constructed based on a sample data set, wherein the prompt is used to indicate a recommendation result corresponding to a rich intent predicted based on the sample historical browsing information and the sample current item information, inputting the prompt into the large language model, and outputting a first predicted recommendation result.

[0100] Exemplarily, the sample text modality data includes sample history browsing information and sample current project information of the text modality. The sample history browsing information can be understood as the information browsed by the sample user in the past time period; the sample current project information can be understood as the information of the project currently browsed by the sample user.

[0101] Combined with the above analysis, taking the book recommendation scenario as an example: the sample historical browsing information can be understood as the relevant information of the sample user browsing books in the past time period; the sample current project information can be understood as the relevant information of the user currently browsing books.

[0102] This embodiment does not limit the way in which the training system obtains prompts. For example, the training system can construct prompts based on sample behavior data. For another example, the training system can output an interactive interface and obtain prompts input by the staff through the interactive interface.

[0103] Among them, the prompt may include three aspects: sample history browsing information, sample current project information, and enriched intent. The prompt is used to guide the large language model to guide the large language model to predict the recommendation results based on the enriched intent, so that the first predicted recommendation result can fit the deep-level interests, purposes and needs of the sample user.

[0104] For example, combined with Figure 4 It can be seen that the training system can obtain prompts constructed based on sample text modal data, and can use the prompts as input data for the large language model to obtain sample text features determined by the large language model, so as to finally obtain the first predicted recommendation result output by the large language model based on the sample text features.

[0105] Continue reading Figure 4 It can be seen that the large language model can output the history hidden vectors h based on the prompt history 、Item hidden vectors h item , generate hidden vectors h generate Based on the token reasoning ability (such as generation ability) of the large language model, it can be known that the large language model infers the next token based on the token inferred from the previous one. Therefore, in this embodiment, for example, the fifth token in the generated hidden vector is inferred based on the fourth token, and so on, which will not be listed here one by one. Of course, Figure 4 It only shows multiple tokens for demonstration purposes and should not be understood as a limitation on tokens.

[0106] In some embodiments, Figure 4As shown, the sample text feature may be the last token in the project hidden vector. The large language model outputting the first predicted recommendation result based on the sample text feature may be that the large language model generates a hidden vector based on the last token in the project hidden vector to determine the first predicted recommendation result.

[0107] In this embodiment, the training system obtains a prompt that is constructed by combining historical browsing information and current browsing information, and guides the large language model to use its generative ability to determine the first predicted recommendation result that can identify the deep-level interests of the sample user through the prompt. That is, the large language model fully combines the historical browsing situation and current browsing situation of the sample user to predict the first predicted recommendation result that enriches the intention of the sample user. Therefore, through the training method provided in this specification, the ability of the large language model to understand the behavior and preferences of the sample user more deeply and meticulously can be enhanced, thereby improving the accuracy and reliability of recommendations based on the large language model.

[0108] Accordingly, the above S303 may include: obtaining a second predicted recommendation result corresponding to the sample text feature and the sample table modal data based on the recommendation model.

[0109] For example, combining the above examples and Figure 4 It can be seen that the input of the recommendation model mainly consists of two parts: sample text features (such as Figure 4 The recommendation model is based on the last token in the item hidden vector described in Figure 4 The last token in the hidden vector of the item and the sample table modal data are used to predict the recommendation result to obtain the second predicted recommendation result.

[0110] In this embodiment, the training device obtains sample text features based on the large language model, uses the sample text features as part of the input of the recommendation model, and trains the large language model and the recommendation model as a whole at one time. That is, end-to-end training is achieved, thereby improving the effectiveness and reliability of the training. In addition, since the recommendation model uses the sample text features generated by the large language model, the recommendation model can be guided to understand the user's behavior and preferences more deeply and meticulously.

[0111] In some embodiments, the above “obtaining a second predicted recommendation result corresponding to the sample text feature and the sample table modal data based on the recommendation model” may include the following steps 1 and 2:

[0112] Step 1: Fuse the sample text features and sample table modal data to obtain fused features.

[0113] Exemplarily, the training system performs feature fusion on the sample behavior data of the text modality and the table modality to obtain features that fuse the sample behavior data of the two modalities (ie, fused features).

[0114] In some embodiments, step 1 may include the following sub-steps 1 and 2:

[0115] Sub-step 1: Encode the sample table modal data to obtain the sample table features.

[0116] This embodiment does not limit the encoding method. In some embodiments, the training system can encode the table modal data using a one-hot encoding method to obtain sample table features.

[0117] Sub-step 2: Concatenate the sample text features and the sample table features to obtain fused features.

[0118] In some embodiments, Figure 4 As shown, the training system can compress the sample text features based on the linear layer to obtain compressed features, and concatenate the compressed features with the sample table features to obtain concatenated features (ie, fused features).

[0119] Combined with the above analysis of sub-steps 1 and 2, it can be seen that in this embodiment, the training system encodes the data of the table modality (i.e., sample table modality data) into features (i.e., sample table features) and concatenates them with the sample text features to obtain fused features, which can significantly improve the flexibility and efficiency of data processing and provide a more comprehensive and richer information basis for subsequent modeling and prediction.

[0120] Step 2: Obtain the second predicted recommendation result corresponding to the fusion feature based on the recommendation model.

[0121] Exemplarily, the training system uses the fused features as input data of the recommendation model to obtain a second predicted recommendation result.

[0122] Combining the above analysis of step 1 and step 2, it can be seen that in this embodiment, the input data of the recommendation model can include features of two modes. This can improve the diversity and comprehensiveness of the input of the recommendation model, thereby improving the effectiveness of the prediction of the training recommendation model, and further improving the accuracy of the recommendation model in the application stage.

[0123] S304: Update the parameters of the preset network with the training objective of minimizing the first loss function between the first predicted recommendation result and the preset first label, and the second loss function between the second predicted recommendation result and the preset second label, to obtain a target network model, wherein the preset first label is determined based on the rich intent of the sample user, and the rich intent is used to characterize deeper needs and preferences than the explicit intent of the sample user.

[0124] Combining the above analysis, we can see that explicit intent and rich intent are relative concepts. Relatively speaking, explicit intent is direct, such as obtaining user intent directly from the behavior of sample users; rich intent is inferred and hidden, such as the behavior of sample users is not directly expressed, but by analyzing the behavior of sample users (such as context and interaction), the potential needs or purposes of sample users can be inferred.

[0125] In related technologies, the first preset tag is usually determined by explicit intent, such as whether the user clicks or not. In the technical solution provided in the specification, the first preset tag is determined by enriched intent, such as user interaction, so as to guide the large language model to recognize the user's deeper intent and interests through enriched intent.

[0126] The preset first label is a label corresponding to the recommendation result predicted by the large language model (ie, the first predicted recommendation result), and the preset second label is a label corresponding to the recommendation result predicted by the recommendation model (ie, the second predicted recommendation result).

[0127] The first loss function represents the difference between the first predicted recommendation result and the first label. Figure 4 It can be seen that the training system can construct a first loss function based on the item hidden vector (specifically, the last token in the item hidden vector) and the generated hidden vector.

[0128] In some embodiments, the first loss function It can be expressed by formula 1:

[0129]

[0130] Among them, ρ LLM is the generation probability of the large language model, w t is the tth word in the prompt, T is the total number of words in the prompt, h is the last token in the item hidden vector, and h generate To generate the hidden vector.

[0131] The second loss function represents the difference between the second predicted recommendation result and the second label. Figure 4 It can be seen that the training system can construct a second loss function based on the output of the recommendation model.

[0132] In some embodiments, the second loss function It can be expressed by formula 2:

[0133]

[0134] Where y is the second label, which can indicate whether the label is clicked. If the label is clicked, y is 1; if the label is not clicked, y is 0. ρ is the predicted probability of clicking the label.

[0135] Correspondingly, by minimizing the first loss function and the second loss function, the predicted recommendation results of the large language model and the recommendation model can be closer to the actual recommendation results, so that the target network model can have a more accurate prediction ability, thereby improving the accuracy and reliability of the target network model's recommendations based on recommendation requests.

[0136] Combined with the above analysis, it can be seen that the target network model can be obtained based on an end-to-end training method, and the target network model has the ability to obtain implicit recommendation intents. Therefore, the target network model can also be called an end-to-end generative framework for learning implicitRecommendation intents using LLMs (EGRec).

[0137] Based on the above analysis of S301-S304, it can be seen that in this embodiment, the training system uses rich intent to determine the first preset label for training, so as to train the large language model to identify the deeper user intent of the sample user, thereby training the recommendation model to determine the recommendation result based on the deeper user intent of the sample user. In addition, since the rich intent is used to determine the first preset label, rather than directly as a feature, the risk of feature leakage can be avoided. Moreover, under end-to-end learning, through gradient backpropagation, the large language model can be guided by the recommendation capability of the recommendation model, reducing the dependence on highly complex and resource-intensive large language models. Accordingly, in the application stage, when the recommendation system makes recommendations based on the target network model, the recommendation results can be made to be highly consistent with the user's deep needs and preferences, thereby improving the accuracy and reliability of the recommendation, and also improving the user experience.

[0138] In some embodiments, the training method may further include: constructing a contrastive learning loss function between sample text features and sample table features, wherein the contrastive learning loss function is used to represent the difference in representation of sample behavior data using sample text features and sample table features respectively.

[0139] Accordingly, the above S304 may include: updating the parameters of the preset network with minimizing the first loss function, the second loss function, and the contrastive learning loss function as the training goal to obtain the target network model.

[0140] For example, when different modalities are used to represent the same data, there may be differences between the features of different modal data. For example, the sample text modal data represents the knowledge of the sample behavior data in the semantic dimension (which can be called semantic knowledge), and the sample table modal data represents the knowledge of the sample behavior data in the identification (ID) type dimension (which can be called ID knowledge). Semantic knowledge and ID knowledge are different representation spaces, and there are differences between the two.

[0141] Corresponding to this embodiment, the training system can use text mode and table mode to characterize the sample behavior data. There may be differences between the sample text features corresponding to the text mode and the sample table features corresponding to the table mode when characterizing the sample behavior data.

[0142] That is, the sample behavior data is also characterized, but there may be differences between the sample text features and the sample table features. In order to eliminate this difference, in this embodiment, Figure 4 As shown, the system can be trained to construct a contrastive learning loss function to promote a unified representation space through the contrastive learning loss function, thereby narrowing the representation differences between the two different modalities.

[0143] Exemplarily, the training system constructs a contrastive learning loss function to bring positive samples closer and negative samples further away, wherein positive samples refer to the text modality representation and table modality representation corresponding to the same sample, and negative samples refer to the text modality representation and table modality representation corresponding to different samples in the same batch.

[0144] In some embodiments, the training system may regularize the sample text features and the sample table features to obtain respective corresponding regularized vectors (or regularized features), and determine the contrastive learning loss function based on the regularized vectors. It can be expressed by formula 3:

[0145]

[0146] Wherein, N is the number of samples in the sample data set. This embodiment does not limit the number of samples, which can be determined by the training system based on demand, historical records, experiments, etc. i is the number of rows of the matrix constructed based on the sample text features and sample table features, k is the number of columns of the matrix, τ is the preset coefficient, is the regularized sample text feature, is the regularized sample table feature.

[0147] In some embodiments, the training system may also pre-configure a penalty parameter to bring stability to the learning process based on the contrastive learning loss function through the size of the penalty parameter, thereby reducing the risk of overfitting and promoting robust training.

[0148] Based on the construction of the contrastive learning loss function, the training system can train the preset network with the goal of minimizing three functions (i.e., the first loss function, the second loss function, and the contrastive learning loss function), such as adjusting the parameters of the preset network until convergence to obtain the target network model.

[0149] Based on the above analysis, it can be seen that in this embodiment, the training system constructs a contrastive learning loss function to reduce the difference between the two modal representation sample behavior data of text and table through the contrastive learning loss function, thereby improving the recommendation accuracy and reliability of the trained target network model.

[0150] In some embodiments, the above-mentioned "updating the parameters of the preset network with minimizing the first loss function, the second loss function, and the comparative learning loss function as the training objective to obtain the target network model" may include: determining the loss function sum of the first loss function, the second loss function, and the comparative learning loss function, and updating the parameters of the preset network with minimizing the loss function sum as the training objective to obtain the target network model.

[0151] Exemplarily, the training system can add the first loss function, the second loss function, and the contrastive learning loss function to obtain the loss function sum, and perform end-to-end joint training with minimizing the loss function sum as the training goal to accelerate the full integration of the three loss functions.

[0152] Among them, the loss function and It can be expressed by formula 4:

[0153]

[0154] Among them, λ1 and λ2 are the control and Hyperparameters of weights.

[0155] In this embodiment, the training system constructs loss functions and performs training with the goal of minimizing the loss functions and, thereby improving the effectiveness and reliability of end-to-end training.

[0156] In order to verify the effectiveness of the above training method, the Amazon-Book dataset and the Industrial-Ads dataset can be used to conduct experiments on the Deep Neural Network (DNN), Wide and DeepModel (Wide&Deep), Deep&Cross Network Version 1 (DCNv1), Deep&Cross Network Version 2 (DCNv2), and eXtremeDeep Factorization Machine (xDeepFM). The area under the curve (ACU) and logarithmic loss (Logarithmic Loss) of different variants under different and same backbone networks are compared. The experimental results are shown in Table 1:

[0157] Table 1

[0158]

[0159] Combined with the above analysis, we can see that the training system can adopt end-to-end joint training, increase the generation capability, increase the contrastive learning loss function training, and adopt rich intent training. The EGRec with DNN as the CTR backbone can be used for ablation experiments to determine the effectiveness of the training including the above four dimensions. The experimental results shown in Table 2 are obtained:

[0160] Table 2

[0161]

[0162] Among them, Zero-shot in Table 2 represents end-to-end joint training, and the specific representation uses the last token in the item hidden vector, which is directly integrated into the recommendation model without fine-tuning. GRec in Table 2 represents a two-stage method, in which the large language model is first fine-tuned and then integrated into the recommendation model. EGRec in Table 2 labelRepresents the model obtained by replacing rich intents (such as comments, etc.) with explicit intents (such as whether the label is clicked) for training. Combined with Table 2, it can be seen that fine-tuning the large language model with rich intents shows that the explicit intent limits the generation ability of the large language model, while the rich intent brings out the generation ability of the large language model. The ERec in Table 2 represents the model that does not include the generation probability in the overall training. The w / o CL in Table 2 represents the model obtained by not including the contrastive learning loss function training. Based on Table 2, it can be seen that the contrastive learning loss function can effectively alleviate the representation difference between the text modality and the table modality.

[0163] In order to deeply analyze how the features generated by the large language model contribute to the final recommendation results, the Amazon-Book dataset can be used to construct the t-distributed stochastic neighbor embedding (t-SNE) graph and the maximum mean discrepancy (MMD) indicator as an example.

[0164] like Figures 5 to 8 As shown, red and blue represent instances with click labels of 1 and 0, respectively. Figure 5 This is a schematic diagram of the distribution obtained using the Zero-shot method; Figure 6 This is a schematic diagram of the distribution obtained using the GRec method; Figure 7 To adopt EGRec label Schematic diagram of the distribution obtained by the method; Figure 8 Schematic diagram of the distribution obtained using the EGRec method.

[0165] Among them, combined Figures 5 to 8 It can be seen that both Zero-shot and GRec methods are close to Gaussian distribution. label The distribution shapes of and EGRec deviate significantly from the Gaussian model. In particular, Figure 8 There is a clear red-blue separation on the right, indicating that EGRec has superior performance in this regard. label and EGRec, the MMD distance can be used for quantitative evaluation. The results show that the distance of EGRec is greater than that of EGRec. label , indicating that EGRec has more significant differences in distribution, thus characterizing its enhanced performance in identifying users’ deep intentions and interests.

[0166] The above example illustrates the target network model obtained through training from the training stage. Now, the reasoning application of the trained network model from the application stage is explained as follows.

[0167] Exemplarily, this specification provides a recommendation method. The recommendation method includes the following S901 and S902:

[0168] S901: In response to a recommendation request corresponding to a target user, obtain historical behavior data of the target user.

[0169] Based on the above analysis, it can be seen that the recommendation request may be triggered by the target user based on a search request, or may be triggered by the target user logging into the corresponding platform, etc.

[0170] Historical behavior data can be understood as data generated by the target user's browsing and other behaviors in the past period of time. The content of historical behavior data may be different for different application scenarios.

[0171] For example, taking the above book recommendation scenario as an example, the recommendation system is an e-book reading platform, or a system deployed in an e-book reading platform. Taking the recommendation system as an e-book reading platform as an example:

[0172] In some embodiments, when the target user logs in to the e-book reading platform, a recommendation request may be initiated to the e-book reading platform. The e-book reading platform obtains relevant information about the target user's historical browsing of books on the e-book reading platform to obtain historical behavior data.

[0173] In other embodiments, the electronic reading platform may output an interactive interface on the terminal device of the target user, and the interactive interface includes a search box. After the target user logs in to the electronic book reading platform, he may enter a search query in the search box, such as the name or type of the book he wants to search for, to initiate a recommendation request to the electronic reading platform. The electronic book reading platform obtains the information related to the books browsed by the target user on the electronic book reading platform and related to his search intention, so as to obtain historical behavior data.

[0174] S902: Inputting the historical behavior data into the target network model, and outputting the target recommendation result corresponding to the recommendation request, wherein the target network model is trained based on the training method described above.

[0175] Exemplarily, the recommendation system uses historical behavior data as input of a target network model, and predicts recommendation results based on the input by the target network model to obtain target recommendation results.

[0176] Combined with the above analysis, it can be seen that the target network model is trained based on rich intent. Therefore, the target network model can make recommendations for target users based on the deeper purposes and needs of target users. Therefore, the target recommendation results determined by the recommendation system based on the technical solutions of S901-S902 above have high accuracy and reliability, which can improve the experience of target users. And in the application stage, there is no need to use the rich intent of the target user as input to ensure the privacy security of the target user. That is, the information security of the target user can be ensured while improving the accuracy of the recommendation.

[0177] In some embodiments, the historical behavior data includes: target text modal data that characterizes the historical behavior data from a text dimension, target table modal data that characterizes the historical behavior data from a table dimension; the target network model includes a large language model and a recommendation model. S902 may include the following steps 1 and 2:

[0178] Step 1: Input the target text modal data into the large language model in the target network model, and output the target text features that represent the target user's intention.

[0179] Step 2: According to the recommendation model in the target network model, the target recommendation results corresponding to the target text features and the target table modal data are obtained.

[0180] In some embodiments, step 2 may include the following sub-steps 1 to 3:

[0181] Sub-step 1: Encode the target table modal data to obtain the target table features.

[0182] Sub-step 2: Concatenate the target text features and the target table features to obtain concatenated features.

[0183] Sub-step 3: Input the concatenated features into the recommendation model in the target network model and output the target recommendation result.

[0184] It is understandable that, in order to avoid tedious description, the same or similar technical features between the application stage and the training stage will not be described in detail in the application stage. Accordingly, the specific technical implementation of the above S902 can be found in the relevant description of the above training stage, which will not be described in detail here.

[0185] It is worth noting that the above examples are only used to illustrate the possible implementation methods of the training method and recommendation method of this specification, and cannot be understood as limiting the implementation methods of the training method and recommendation method of this specification. Exemplarily, based on the above technical concept, some of the above technical features can be combined to obtain a new embodiment; new technical features can be added on the basis of the above examples to obtain a new embodiment; some technical features can be reduced on the basis of the above examples to obtain a new embodiment; some technical features in the above examples can be replaced with other technical features; some technical features and the order of the above examples can be adjusted to obtain a new embodiment, etc., which will not be listed here one by one.

[0186] Based on the above technical concept, this specification also provides a computer-readable non-temporary storage medium, which stores at least one instruction set. When the at least one instruction set is executed by a processor, the steps of the training method or recommendation method described in this specification are implemented.

[0187] In some possible implementations, various aspects of this specification may also be implemented in the form of a program product, which includes a program code. When the program product is run on a training system, the program code is used to cause the training system to perform the steps of the training method described in this specification. The program product for implementing the above method may include a portable compact disk read-only memory (CD-ROM) including program code, and may be run on a training system. However, the program product of this specification is not limited to this, and in this specification, a readable storage medium may be any tangible medium containing or storing a program, which may be used by an instruction execution system or used in combination with it. The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media include: an electrical connection with one or more conductors, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. The computer-readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above. Program code for performing the operations of the present specification may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the training system, partially on the training system, as a stand-alone software package, partially on the training system and partially on a remote training system, or entirely on a remote training system.

[0188] In other embodiments, when the program product is run on a recommendation system, the program code is used to enable the recommendation system to perform the steps of the recommendation method described in this specification. For a specific description, please refer to the above description of the example of the training system, which will not be repeated here.

[0189] It should be noted that the collection, storage, use, processing, transmission, provision and disclosure of relevant user information (such as browsing information, etc.) involved in the technical solution of this specification are in compliance with the relevant laws and regulations and do not violate public order and good morals.

[0190] The above is a description of a specific embodiment of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require a specific order or a continuous order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0191] In summary, after reading this detailed disclosure, those skilled in the art will appreciate that the foregoing detailed disclosure may be presented only by way of example and may not be limiting. Although not explicitly stated herein, those skilled in the art will appreciate that this specification requires various reasonable changes, improvements and modifications to the embodiments. These changes, improvements and modifications are intended to be proposed by this specification and are within the spirit and scope of the exemplary embodiments of this specification.

[0192] In addition, certain terms in this specification have been used to describe embodiments of this specification. For example, "one embodiment", "an embodiment" and / or "some embodiments" mean that a particular feature, structure or characteristic described in conjunction with the embodiment may be included in at least one embodiment of this specification. Therefore, it can be emphasized and should be understood that two or more references to "an embodiment" or "one embodiment" or "an alternative embodiment" in various parts of this specification do not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics may be appropriately combined in one or more embodiments of this specification.

[0193] It should be understood that in the foregoing description of the embodiments of this specification, in order to help understand a feature and for the purpose of simplifying this specification, this specification combines various features in a single embodiment, figure or its description. However, this does not mean that the combination of these features is necessary. When reading this specification, it is entirely possible for a person skilled in the art to mark out some of the devices as separate embodiments. In other words, the embodiments in this specification can also be understood as the integration of multiple secondary embodiments. And the content of each secondary embodiment is also valid when it is less than all the features of a single aforementioned disclosed embodiment.

[0194] Each patent, patent application, publication of patent applications, and other materials, such as articles, books, specifications, publications, documents, articles, etc., cited herein, except to the extent that it is inconsistent or conflicting with this document or that has a limiting effect on the broadest scope of the claims, may be incorporated herein by reference and used for all purposes now or hereafter associated with this document. In addition, in the event of any inconsistency or conflict between the description, definition, and / or use of a term in any material and the description, definition, and / or use of a term in this document, the term in this document shall prevail.

[0195] Finally, it should be understood that the embodiments of the application disclosed herein are explanations of the principles of the embodiments of this specification. Other modified embodiments are also within the scope of this specification. Therefore, the embodiments disclosed in this specification are only used as examples and not as limitations. Those skilled in the art can adopt alternative configurations according to the embodiments in this specification to implement the applications in this specification. Therefore, the embodiments of this specification are not limited to the embodiments accurately described in the application.

Claims

1. A network model training method, the method is applied to a preset network, the preset network includes a large language model and a recommendation model; the method includes: Obtaining a sample data set, wherein the sample data set includes sample behavior data of a sample user; Obtaining a first predicted recommendation result corresponding to the sample data set based on the large language model; Obtaining a second predicted recommendation result corresponding to the sample data set based on the large language model and the recommendation model; and The parameters of the preset network are updated with the training objectives of minimizing a first loss function between the first predicted recommendation result and a preset first label, and a second loss function between the second predicted recommendation result and a preset second label to obtain a target network model, wherein the preset first label is determined based on the rich intent of the sample user, and the rich intent is used to characterize deeper needs and preferences than the explicit intention of the sample user.

2. The method according to claim 1, wherein: The sample data set includes: sample text modal data that characterizes the sample behavior data from a text dimension, and sample table modal data that characterizes the sample behavior data from a table dimension; The obtaining, based on the large language model, a first predicted recommendation result corresponding to the sample data set includes: obtaining, based on the large language model, sample text features corresponding to the sample text modal data, and determining the first predicted recommendation result according to the sample text features; The obtaining of the second predicted recommendation result corresponding to the sample data set based on the large language model and the recommendation model includes: obtaining the second predicted recommendation result corresponding to the sample text feature and the sample table modal data based on the recommendation model.

3. The method according to claim 2, wherein: The obtaining, based on the recommendation model, the second predicted recommendation result corresponding to the sample text feature and the sample table modal data includes: Performing feature fusion on the sample text features and the sample table modal data to obtain fused features; and The second predicted recommendation result corresponding to the fusion feature is obtained based on the recommendation model.

4. The method according to claim 3, wherein: The step of fusing the sample text features and the sample table modal data to obtain fused features includes: Encoding the sample table modal data to obtain sample table features; The sample text feature and the sample table feature are concatenated to obtain the fusion feature.

5. The method according to claim 4, wherein: The method further comprises: Constructing a contrastive learning loss function between the sample text feature and the sample table feature, wherein the contrastive learning loss function is used to represent the difference in representation of the sample behavior data using the sample text feature and the sample table feature respectively; And, updating the parameters of the preset network with the training objective of minimizing the first loss function between the first predicted recommendation result and the preset first label, and the second loss function between the second predicted recommendation result and the preset second label to obtain the target network model includes: updating the parameters of the preset network with the training objective of minimizing the first loss function, the second loss function, and the contrastive learning loss function to obtain the target network model.

6. The method according to claim 5, wherein: The updating of the parameters of the preset network with minimizing the first loss function, the second loss function, and the contrastive learning loss function as the training objective to obtain the target network model includes: Determine a loss function sum of the first loss function, the second loss function, and the contrastive learning loss function; and The target network model is obtained by minimizing the loss function and updating the parameters of the preset network for the training objective.

7. The method according to claim 1, wherein: The sample behavior data includes sample history browsing information and sample current project information; the first predicted recommendation result corresponding to the sample data set obtained based on the large language model includes: Obtaining a prompt constructed based on the sample data set, wherein the prompt is used to indicate a recommendation result corresponding to the enriched intent predicted based on the sample historical browsing information and the sample current project information; The prompt is input into the large language model, and the first predicted recommendation result is output.

8. The method according to claim 1, wherein: The enriched intent includes sample user comments and / or sample search queries.

9. A recommendation method comprising: In response to a recommendation request corresponding to a target user, obtaining historical behavior data of the target user; The historical behavior data is input into a target network model, and a target recommendation result corresponding to the recommendation request is output, wherein the target network model is trained based on the method described in any one of claims 1 to 8.

10. The method according to claim 9, wherein: The historical behavior data includes: target text modal data that characterizes the historical behavior data from a text dimension, and target table modal data that characterizes the historical behavior data from a table dimension; the target network model includes a large language model and a recommendation model; The step of inputting the historical behavior data into a target network model and outputting a target recommendation result corresponding to the recommendation request includes: Inputting the target text modality data into the large language model in the target network model, and outputting target text features representing the target user's intention; According to the recommendation model in the target network model, the target recommendation result corresponding to the target text feature and the target table modal data is obtained.

11. The method according to claim 10, wherein: Obtaining the target recommendation result corresponding to the target text feature and the target table modal data according to the recommendation model in the target network model, including: Encoding the target table modal data to obtain target table features; Splicing the target text feature and the target table feature to obtain a splicing feature; The concatenated features are input into the recommendation model in the target network model, and the target recommendation result is output.

12. A network model training system, wherein: The training system is applied to a preset network, wherein the preset network includes a large language model and a recommendation model; The training system comprises: At least one storage medium storing at least one instruction set for training the preset network; At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, the at least one instruction set is read, and the method as claimed in any one of claims 1 to 8 is executed according to the instructions of the at least one instruction set.

13. A recommendation system comprising: at least one storage medium storing at least one set of instructions for making recommendations; At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, the at least one instruction set is read, and the method as described in any one of claims 9 to 11 is executed according to the instructions of the at least one instruction set.

Images