Service processing method and apparatus, and computer device, storage medium and program product

Abstract

A service processing method executed by a computer device. The service processing method comprises: acquiring prompt information of a target item which is recalled for a service user, wherein the prompt information of the target item comprises information of the service user, features of at least one associated item which is associated with the service user, and features of the target item (S301); calling an item processing model under a model application inference framework, and on the basis of the prompt information of the target item, predicting the degree of association of the service user with the target item, wherein the item processing model is obtained by means of training under a model training inference framework, and is reused in the model application inference framework (S302), and the processing speed of the item processing model in the model application inference framework is higher than the processing speed of the item processing model in the model training inference framework; and performing service processing on the target item on the basis of the degree of association of the service user with the target item (S303).

Description

Business processing method, device, computer equipment, storage medium, and program product

[0001] Related applications

[0002] This application claims priority to Chinese patent application number 2023113635030, filed on October 20, 2023, entitled “Business processing method, device, computer equipment, storage medium, program product,” the entire text of which is hereby incorporated by reference. Technical Field

[0003] The present application relates to the field of computer technology, in particular to the field of artificial intelligence technology, and specifically to a business processing method, apparatus, computer equipment, storage medium, and program product. Background Art

[0004] Typically, when a business user accesses a service, the service provider can perform service processing. This processing can specifically refer to recommending items within the service that are highly relevant to the business user. In other words, the service provider can recommend items to the business user accessing the service. Currently, service providers can use artificial intelligence models (e.g., recommendation models) for service processing. However, these AI models typically have a large number of model parameters and slow inference speed, resulting in inefficient service processing.

[0005] Summary of the Invention

[0006] The embodiments of the present application provide a business processing method, apparatus, computer equipment, storage medium, and program product.

[0007] In one aspect, an embodiment of the present application provides a service processing method, which is executed by a computer device and includes:

[0008] Obtaining prompt information of a target item recalled for a business user, the prompt information of the target item including information of the business user, characteristics of at least one associated item associated with the business user, and characteristics of the target item;

[0009] Invoking an item processing model within a model application reasoning framework to predict the association between the business user and the target item based on the prompt information of the target item; the item processing model is trained within a model training reasoning framework and reused within the model application reasoning framework; the processing speed of the item processing model within the model application reasoning framework is higher than the processing speed of the item processing model within the model training reasoning framework; and

[0010] Based on the degree of association between the business user and the target item, business processing is performed on the target item.

[0011] Accordingly, an embodiment of the present application provides a service processing device, which includes:

[0012] an acquiring unit, configured to acquire prompt information of a target item recalled for a business user, wherein the prompt information of the target item includes information of the business user, characteristics of at least one associated item associated with the business user, and characteristics of the target item;

[0013] a processing unit configured to call an item processing model within a model application reasoning framework and predict the degree of association between the business user and the target item based on the prompt information of the target item; the item processing model is trained within a model training reasoning framework and reused within the model application reasoning framework; the processing speed of the item processing model within the model application reasoning framework is higher than the processing speed of the item processing model within the model training reasoning framework; and

[0014] The processing unit is further configured to perform business processing on the target item based on the degree of association between the business user and the target item.

[0015] Accordingly, an embodiment of the present application provides a computer device, comprising:

[0016] a processor suitable for implementing a computer program;

[0017] Computer-readable storage medium, the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the above-mentioned business processing method.

[0018] Accordingly, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is read and executed by a processor of a computer device, the computer device executes the above-mentioned business processing method.

[0019] Accordingly, an embodiment of the present application provides a computer program product, which includes a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the above-mentioned service processing method.

[0020] The details of one or more embodiments of the present application are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the present application will become apparent from the description, drawings, and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the conventional technology, the following briefly introduces the drawings required for use in the embodiments or the conventional technology descriptions. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the disclosed drawings without any creative work.

[0022] FIG1 is a schematic diagram of the architecture of a business processing system provided in an embodiment of the present application;

[0023] FIG2 is a schematic diagram of the training and testing processes of an item processing model provided in an embodiment of the present application;

[0024] FIG3 is a flow chart of a business processing method provided in an embodiment of the present application;

[0025] FIG4 is a schematic diagram of a training process of an item processing model provided in an embodiment of the present application;

[0026] FIG5 is a schematic diagram of the structure of an item processing model before transformation under a model application reasoning framework provided in an embodiment of the present application;

[0027] FIG6 is a schematic diagram of the structure of a modified item processing model under a model application reasoning framework provided in an embodiment of the present application;

[0028] FIG7 is a schematic flow chart of another article processing method provided in an embodiment of the present application;

[0029] FIG8 is a schematic diagram of a concurrency mechanism provided in an embodiment of the present application;

[0030] FIG9 is a schematic diagram of another concurrency mechanism provided in an embodiment of the present application;

[0031] FIG10 is a schematic diagram of another concurrency mechanism provided in an embodiment of the present application;

[0032] FIG11 is a schematic structural diagram of a service processing device provided in an embodiment of the present application;

[0033] FIG12 is a schematic structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0035] The embodiments of the present application relate to a service, which may specifically refer to an item recommendation service. The service processing logic of the item recommendation service is to recommend to the service user, when the service user makes a service call, a candidate item with a high degree of relevance to the service user from among multiple candidate items recalled for the service user. The items mentioned in the embodiments of the present application may include, but are not limited to, any of the following: electronic goods (e.g., smart retail goods and e-commerce goods) and media information (e.g., video, audio, image, article, and website), etc.

[0036] Business service providers can use artificial intelligence models to perform business processing. Specifically, business service providers can use artificial intelligence models to predict the association between business users and candidate items recalled for business users, and obtain the degree of association between business users and candidate items. Therefore, based on the degree of association between business users and candidate items, business service providers can recommend candidate items with a high degree of association with business users to business users.

[0037] At present, in the field of artificial intelligence, with the popularity of ChatGPT (Chat Generative Pre-trained Transformer), more and more LLMs (Large Language Models) have emerged. The embodiment of the present application does not limit the type of large-scale generative language models. A large-scale generative language model is a generative language model whose model framework is an encoder-decoder framework, an encoder-only framework, or a decoder-only framework. Among them, the encoder-decoder framework is a very common model framework in deep learning. One of its most notable features is that it is an end-to-end learning algorithm. Such a model is also called sequence to sequence learning. The so-called encoding refers to converting the input sequence into a fixed-length vector, and the so-called decoding refers to converting the previously generated fixed vector into an output sequence. The encoder-only framework means that the model has the encoding function; the decoder-only framework means that the model has the decoding function. The embodiment of the present application is to illustrate with the example that the large-scale generative language model is a generative language model whose model framework is an Encoder-Decoder framework, for example, the large-scale generative language model can be an MT5 model whose model framework is an Encoder-Decoder framework. The MT5 model is a multilingual natural language processing model that adopts the Transformer model architecture and is pre-trained on a large-scale multilingual corpus. The Transformer model is a natural language processing model that can learn the contextual semantics of each word based on the relationship between each word (token) in the input sequence. The Transformer model can achieve good semantic learning effects based on the Encoder-Decoder framework and the Attention (attention) mechanism. The biggest advantage of the Transformer model is that it can be efficiently parallelized. Words (tokens) are the constituent elements of the input sequence, and input sequences of types such as sentences, paragraphs, and articles can be decomposed into data structures with words (tokens) as units and input into large-scale generative language models.

[0038] How to apply large-scale generative language models as item processing models to the item recommendation business has become a current research hotspot. However, large-scale generative language models currently cannot meet the item recommendation needs of the item recommendation business. Specifically: ① Large-scale generative language models have a large number of model parameters, slow model inference speed, and high model inference latency. The model inference performance limits the implementation of large-scale generative language models in the real-time item recommendation business and cannot meet the real-time recommendation and parallel inference requirements of the item recommendation business; ② The tokens (words) generated by large-scale generative language models cannot meet the recommendation scoring and ranking requirements of the item recommendation business. Based on ① and ②, it can be seen that large-scale generative language models cannot currently be applied as item processing models to the item recommendation business.

[0039] To address the poor inference performance of large-scale generative language models (i.e., their inability to meet the real-time recommendation requirements of item recommendation services), the industry has proposed the following three solutions: The first solution uses model compression or model distillation to reduce the model parameters of large-scale generative language models, thereby improving model inference speed at the expense of model accuracy. The second solution involves fine-tuning the pre-trained model, exiting inference early based on a set threshold, and reducing the number of model network layers. The third solution involves caching the intermediate results of the decoding module in large-scale generative language models to improve model inference speed. However, in the practice of item recommendation services, it has been found that the above three solutions are not desirable. The first and second solutions sacrifice result accuracy to improve inference speed, which is not ideal for real-time item recommendation applications. This sacrifice of result accuracy cannot guarantee the effectiveness of large-scale generative language models in real-time item recommendation scenarios, resulting in a significant decrease in the estimated Area Under the Column (AUC). The estimated AUC is a model evaluation metric that can be used to evaluate binary classification models. A higher estimated AUC value indicates higher classification accuracy, while a lower estimated AUC value indicates lower classification accuracy. For the third solution, it is necessary to increase the memory cache. Under the condition of limited resources, this solution cannot improve the model inference speed of large-scale generative language models.

[0040] Based on this, an embodiment of the present application proposes a business processing method. This business processing method addresses the problem of poor reasoning performance of large-scale generative language models (that is, large-scale generative language models cannot meet the real-time recommendation requirements of item recommendation services) and proposes a solution for separating the training framework and reasoning framework of large-scale generative language models. The processing speed of large-scale generative language models under the reasoning framework is higher than the processing speed of large-scale generative language models under the training framework. In the model reasoning stage of large-scale generative language models, an reasoning acceleration framework that can improve the model processing speed is used without sacrificing the model accuracy of the large-scale generative language model. In this way, the model reasoning speed of large-scale generative language models can be improved while ensuring the model effect, thereby improving business processing efficiency.

[0041] This business processing method addresses the problem that large-scale generative language models cannot meet the recommendation scoring and ranking requirements of the item recommendation business. It proposes a solution for optimizing and transforming the decoder of the large-scale generative language model. The optimized and transformed large-scale generative language model is able to score candidate items. In other words, the optimized and transformed large-scale generative language model can output the correlation between business users and candidate items, use the correlation between business users and candidate items as the score of the candidate items, and rank and recommend the candidate items. This makes the large-scale generative language model adaptable to the item recommendation business and successfully applied as an item processing model in the item recommendation business.

[0042] The embodiment of the present application does not limit the type of training framework used by the large-scale generative language model (i.e., the item processing model). For example, the training framework can be a pytorch framework. Similarly, the embodiment of the present application does not limit the type of reasoning framework used by the large-scale generative language model (i.e., the item processing model). The reasoning framework can be any reasoning acceleration framework that can improve the processing speed of the large-scale generative language model (i.e., the item processing model); for example, the reasoning framework can be the reasoning acceleration framework of Faster-Transformer, which is a performance optimization solution proposed for Transformer model reasoning. Faster-Transformer is written in C++ and CUDA and relies on the highly optimized cuBLAS; wherein C++ (C plus plus) is a computer high-level programming language, CUDA (Compute Unified Device Architecture) is a general-purpose parallel computing architecture, and cuBLAS (CUDA Basic Linear Algebra Subroutine Library) is CUDA's basic linear algebra subroutine library. That is to say, the embodiment of the present application can adapt the item recommendation business to fine-tune the large-scale generative language model (i.e., the item processing model) under the PyTorch framework. In order to improve the model inference speed in the model inference stage, the Faster-Transformer-based inference acceleration framework can be adopted, and the fine-tuned large-scale generative language model (i.e., the item processing model) can be reused. The decoder of the large-scale generative language model (i.e., the item processing model) can be optimized and scored under the inference acceleration framework.

[0043] In addition, the business processing method proposed in the embodiments of the present application can also be combined with cloud technology. For example, cloud computing in cloud technology can further improve the model reasoning speed of the large-scale generative language model (i.e., the item processing model) in the business processing method through a distributed computing model. Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and network within a wide area network or local area network to realize the calculation, storage, processing, and sharing of data. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, application technology, etc. applied in the cloud computing business model. It can form a resource pool that is used on demand and is flexible and convenient. Cloud computing is a computing model that distributes computing tasks on a resource pool composed of a large number of computers, enabling various application systems to obtain computing power, storage space, and information services as needed. The network that provides resources is called a "cloud". The resources in the "cloud" appear to users to be infinitely scalable and can be obtained at any time, used on demand, expanded at any time, and paid for on a per-use basis.

[0044] The following describes the business processing system provided in the embodiment of the present application with reference to the accompanying drawings. The business processing system may be suitable for implementing the business processing method provided in the embodiment of the present application.

[0045] As shown in Figure 1, the business processing system may include a recommendation request device 101 and an item recommendation device 102. The embodiment of the present application does not limit the connection method between the recommendation request device 101 and the item recommendation device 102. A direct communication connection can be established between the recommendation request device 101 and the item recommendation device 102 through wired communication, or an indirect communication connection can be established between the recommendation request device 101 and the item recommendation device 102 through wireless communication.

[0046] in:

[0047] (1) Recommendation request device 101:

[0048] The recommendation request device 101 may be a device used by a business user. A business product corresponding to the business may be running on the recommendation request device 101. The business user may access the business through the business product running on the recommendation request device 101. Accessing the business can be understood as the business user logging into the business, browsing business content, or performing a business operation. The present embodiment does not limit the type of business product; business products may include business applications, business applets, business browsers, and business websites. The recommendation request device 101 may be used to initiate an item recommendation request to the item recommendation device 102. The item recommendation request may be used to request that items highly relevant to the business user be recommended to the business user. The item recommendation request may be generated when the business user accesses the business; for example, when the business user logs into the business; when the business user browses business content; or when the business user performs a business operation (e.g., an item recommendation operation or an operation to access an item within the business). The recommendation request device 101 may also display relevant information about the recommended items (e.g., item name, item title, item image, item video, and item price, etc.) to the business user.

[0049] (2) Item recommendation device 102:

[0050] The item recommendation device 102 can be used to recommend items with a high degree of relevance to business users. Specifically: S101, the item recommendation device 102 can respond to an item recommendation request sent by the recommendation request device 101 and recall multiple candidate items for the business user. S102, for any candidate item among the multiple candidate items (which can be represented as a target item), the item recommendation device 102 can invoke an item processing model within a model application reasoning framework (the model application reasoning framework refers to the reasoning framework mentioned above) to predict the relevance between the business user and the candidate item. S103, after obtaining the relevance between the business user and each candidate item, the item recommendation device 102 can select an item with a high degree of relevance to the business user from the multiple candidate items based on the relevance between the business user and each candidate item, and recommend the item whose relevance to the business user meets the recommendation criteria to the business user. S104, the item recommendation information is returned to the recommendation request device 101. The relevance meeting the recommendation criteria can be a relevance greater than or equal to a preset threshold, or a ranking number less than or equal to a preset ranking number after sorting in descending order of relevance.

[0051] For the item recommendation device 102, before the item processing model officially recommends items online, the item processing model needs to go through a model training phase and a model testing phase. After the item processing model passes the model test, it can officially recommend items online.

[0052] As shown in Figure 2, in the model training stage: S201a, training samples for training the item processing model can be prepared; S202a, in order to adapt to the item recommendation business, the initial item processing model can be modified so that the initial item processing model has the ability to score items; S203a, the initial item processing model can be trained based on the training samples under the model training inference framework (the model training inference framework refers to the training framework mentioned above) to obtain the item processing model.

[0053] As shown in Figure 2, in the model testing phase: S201b, test samples for testing the item processing model can be prepared; S202b, the item processing model can be deployed and adapted with the model application reasoning framework. Deployment adaptation means that in order to adapt to the item recommendation business, similar to the model training phase, the item processing model under the model application reasoning framework is modified so that the item processing model under the model application reasoning framework has the ability to score items; S203b, the item processing model under the model application reasoning framework is deployed online. Specifically, after the item processing model is deployed and adapted with the model application reasoning framework, the framework C++ code can be compiled to obtain the lib file (a static data connection library). Online deployment means deploying the online model reasoning service based on the compiled lib file in combination with the model reasoning code. During deployment, it is necessary to pay attention to the consistency of the compilation environment versions such as cuda version and gcc (a compiler); S204b, the item processing model is tested. Perform model calibration. Since the frameworks used in the model training stage and the model inference stage in the embodiment of the present application are different, model calibration is required. Model calibration refers to evaluating the item processing model based on the same batch of test samples under the model training inference framework (for example, the pytorch framework) used in the model training stage and the model application inference framework (for example, the Faster-Transformer framework) used in the model inference stage. If the difference between the evaluation results of the item processing model under the two frameworks (the evaluation result can specifically refer to the estimated AUC value) is less than the threshold thres (the threshold thres can be set according to the empirical value, for example, the threshold thres can usually be 0.00001), it can be determined that the item processing model test has passed; S205b, after determining that the item processing model test has passed, the item processing model can be put online, connected to the real online traffic, and the item processing model can be called under the model application inference framework to perform item recommendations.

[0054] It should be noted that the recommendation request device 101 can be a terminal, and the item recommendation device 102 can be a terminal or a server; the terminal mentioned in the embodiment of the present application can include but is not limited to any of the following: smart phones, tablet computers, laptops, desktop computers, smart watches, smart home appliances, intelligent voice interaction devices, vehicle terminals and aircraft, but is not limited to these; the server mentioned in the embodiment of the present application can be a separate physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), as well as basic cloud computing services such as big data and artificial intelligence platforms. The embodiment of the present application does not limit this. In addition, the business processing system shown in Figure 1 is illustrated by taking the example of the recommendation request device 101 and the item recommendation device 102 being integrated into different devices. The recommendation request device 101 and the item recommendation device 102 can also be integrated into the same device. For example, the recommendation request device 101 and the item recommendation device 102 can be integrated into the same terminal. In the same terminal, the recommendation request device 101 can be understood as the front end of the terminal, and the item recommendation device 102 can be understood as the backend service of the terminal.

[0055] It is understandable that the business processing system of the embodiment shown in FIG1 is intended to more clearly illustrate the technical solution of the embodiment of the present application, and does not constitute a limitation on the technical solution provided by the embodiment of the present application. A person of ordinary skill in the art will know that with the evolution of the system architecture and the emergence of new business scenarios, the technical solution provided by the embodiment of the present application is equally applicable to similar technical problems. In addition, the relevant data collection and processing (for example, collecting information of business users) in the embodiment of the present application should be strictly in accordance with the requirements of relevant laws and regulations when applied in the example, and the informed consent or separate consent of the subject of personal information should be obtained, and subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the subject of personal information.

[0056] Based on the above introduction to the business processing method and the business processing system, the specific application of the business processing method provided in the embodiment of the present application is introduced below.

[0057] Based on the types of items in the business, the business processing method provided by the embodiment of the present application can be applied to businesses such as smart retail, e-commerce recommendation, and information flow recommendation. Specifically, in the smart retail business, the items recommended to business users may be, for example, smart retail products. The correlation between the business user and each smart retail product recalled for the business user can be analyzed. Based on the relationship between the business user and each smart retail product, smart retail products with a higher correlation with the business user are recommended to the business user. In the e-commerce recommendation scenario, the items recommended to business users may be, for example, e-commerce products. The correlation between the business user and each e-commerce product recalled for the business user can be analyzed. Based on the relationship between the business user and each e-commerce product, e-commerce products with a higher correlation with the business user are recommended to the business user. In the information flow recommendation business, the items recommended to business users may be, for example, media information such as videos, audios, images, and articles. The correlation between the business user and each media information recalled for the business user can be analyzed. Based on the relationship between the business user and each media information, media information with a higher correlation with the business user is recommended to the business user.

[0058] Based on the complexity of the business, the business processing method provided by the embodiment of the present application can be applied to simple businesses with relatively sparse data (for example, a small number of items in the business, less relevant information about items in the business, etc.) and poor recommendation effects of traditional ranking recommendation models. The business processing method provided by the embodiment of the present application can also be applied to complex businesses involving multiple business scenarios or multiple interfaces. Multiple business scenarios can be understood as the simultaneous existence of multiple business scenarios in a business, and item recommendations can be made to business users in multiple business scenarios. For example, a business can include an e-commerce recommendation business scenario and an information flow recommendation business scenario. E-commerce product recommendations can be made to business users in the e-commerce recommendation business scenario, and information flow recommendations can be made to business users in the information flow recommendation business scenario, and multiple business scenarios can make item recommendations in parallel. Multiple interfaces can be understood as a business that accesses multiple sub-businesses through different interfaces, and can make item recommendations to business users in each sub-business, and the sub-business services of each sub-business can make item recommendations in parallel; or, multiple interfaces can be understood as a business that accesses multiple business users through different interfaces, and can make item recommendations for each business user, and each business user can make item recommendations in parallel. For example, at least 5 business users can be accessed in the business at the same time. On the basis of maintaining the same model effect, the reasoning acceleration framework using the embodiment of the present application can reduce the reasoning time by 5 times compared with the native framework.

[0059] It can be seen that the embodiments of the present application can utilize the rich language knowledge in the natural language model to fine-tune the item recommendation business, thereby improving the learning ability of the large language model for the item recommendation scenario. In addition, the embodiments of the present application support high-concurrency reasoning, can access multiple business scenarios, multiple sub-businesses or multiple business users, can solve performance bottlenecks, and promote the implementation of large language models in the industry.

[0060] The following is a detailed introduction to the service processing method provided in the embodiment of the present application with reference to the accompanying drawings.

[0061] This embodiment of the present application provides a business processing method that introduces the modification of an item processing model during the model training phase and the model inference phase to adapt to business (i.e., item recommendation business). The business processing method can be executed by a computer device, which can be, for example, the item recommendation device 102 in the business processing system shown in FIG1 , or the computer device can be, for example, an integrated device of the recommendation request device 101 and the item recommendation device 102. As shown in FIG3 , the business processing method can include, but is not limited to, the following steps S301 to S303:

[0062] S301 : Acquire prompt information of a target item recalled for a business user. The prompt information of the target item includes information of the business user, features of at least one associated item associated with the business user, and features of the target item.

[0063] In a service provided by a business provider (i.e., an item recommendation service), when a business user accesses the service, the service provider can recall multiple candidate items for the business user. This service provider needs to analyze the association between the business user and each candidate item through an item processing model. Then, based on the association between the business user and each candidate item, the service provider can select a recommended item from the multiple candidate items and make item recommendations to the business user based on the recommended item. To facilitate understanding of the analysis process of the association between a business user and each candidate item, this article uses any one of the multiple candidate items (any candidate item can be represented as a target item) as an example to illustrate the analysis process of the association between the business user and the target item.

[0064] Specifically, the association between the business user and the target item is analyzed based on the prompt information of the target item. The prompt information of the target item is used as input information of the item processing model, and the prompt information of the target item can be obtained. The prompt information of the target item may include information of the business user, characteristics of at least one associated item associated with the business user, and characteristics of the target item; wherein, at least one associated item associated with the business user refers to a historical item that the business user has historically visited in the business, and the characteristics of an item (including an associated item, a target item or a candidate item) refers to information used to characterize the characteristics of the item.

[0065] Unlike the input information of traditional recommendation models, the item processing model is a generative language model. Its input information (i.e., prompt information) consists of natural language. The prompt information for a target item can be used to inquire whether a business user will visit the target item given their historical access to related items. The item processing model's analysis principle is to understand the semantics of the target item's prompt information, analyze the similarity between the target item's features and the features of at least one related item, and analyze the association between the target item's features and the business user's information. It then generates a predicted association label (answer) between the business user and the target item and the degree of association under the predicted association label. The predicted association label represents the business user's willingness to visit the target item (e.g., yes (the business user will visit the target item) or no (the business user will not visit the target item)). The degree of association represents the probability of the business user visiting the target item given the corresponding predicted association label and serves as a basis for ranking target items.

[0066] The construction method for prompt information for each candidate item recalled for a business user is the same. This section describes the construction method for prompt information for the target item. First, the item ID sequence (item1, item2, …, itemn) of the associated items the business user has historically accessed, the item ID of the target item (now_click_item), and the business user's information can be obtained from the data center. The item ID can be used to uniquely identify the corresponding item. The business user's access operations on the associated items in the business can include at least click-to-browse operations and purchase operations. Therefore, the item ID sequence of the associated items the business user has historically accessed can include a click ID sequence (click_item1, click_item2, …, click_itemn) of the associated items the business user has historically clicked on, and a purchase ID sequence (buy_item1, buy_item2, …, buy_itemn) of the associated items the business user has historically purchased. The business user's information can include at least the business user's ID and business user characteristics. The business user's ID can be used to uniquely identify the business user, and the business user's characteristics refer to information that characterizes the business user. Then, based on the item ID sequence (item1, item2, ..., itemn) of the associated items the business user has historically accessed, the features of the associated items the business user has historically accessed can be obtained from the material library. Furthermore, based on the item ID (now_click_item) of the target item, the features of the target item can be obtained from the material library. For example, if the item feature is the item title, the features of the associated items the business user has historically accessed can include the click title sequence (title1, title2, title3, ..., titlen) of the associated items the business user has historically clicked, and the purchase title sequence (buy_title1, buy_title2, ..., buy_titlen) of the associated items the business user has historically purchased. For example, if the item feature is the item title, the feature of the target item is the title of the target item (now_click_title). Based on this, the constructed prompt information for the target item can include any of the following:

[0067] (1) Basic prompt information: Basic prompt information may include the business user's ID, the characteristics of the associated items that the business user has previously visited, and the characteristics of the target item. Basic prompt information can be used to inquire whether the business user will visit the target item if the business user has previously visited the associated items. For example, basic prompt information may be expressed as: If the business user has clicked on the associated items: [title1], [title2], [title3], …, [titlen], will the business user click on the target item [now_click_title]?

[0068] (2) Prompt information containing time-sensitive features: Prompt information containing time-sensitive features may include the business user's ID, the features of the associated items that the business user has accessed in different historical periods, and the features of the target item. Prompt information containing time-sensitive features can be used to inquire whether the business user will access the target item if the business user has accessed the corresponding associated items in different historical periods. For example, prompt information containing time-sensitive features can be expressed as follows: If the business user has clicked on the associated items: [title1], [title2], [title3], …, [titlen], and has recently clicked on the associated items: [title1], [title2] …, will the business user click on the target item [now_click_title]?

[0069] (3) Prompt information containing time-sensitive features + business user characteristics: Prompt information containing time-sensitive features + business user characteristics may include the characteristics of the business user, the characteristics of the associated items accessed by the business user in different historical periods, and the characteristics of the target item. Prompt information containing time-sensitive features + business user characteristics can be used to inquire whether the business user with the characteristics will access the target item if the business user accesses the corresponding associated items in different historical periods. For example, prompt information containing time-sensitive features + business user characteristics can be expressed as: If a business user with [feature 1], [feature 2], … has clicked on the associated items: [title1], [title2], [title3], …, [titlen], and has recently clicked on the associated items [title1], [title2] …, will the business user click on the target item [now_click_title]?

[0070] (4) Prompt information containing timeliness characteristics + business user characteristics + access characteristics: The prompt information containing timeliness characteristics + business user characteristics + access characteristics may include the characteristics of the business user, the characteristics of the associated items accessed by the business user in different historical periods and different access operations, and the characteristics of the target item. The prompt information containing timeliness characteristics + business user characteristics + access characteristics can be used to inquire whether the business user will access the target item if the business user with the characteristics of the business user accesses the corresponding associated items in different historical periods and different access operations. For example, the prompt information containing timeliness characteristics + business user characteristics + access characteristics can be expressed as: If a business user with [business user characteristics 1], [business user characteristics 2], ... has clicked on the associated items: [title1], [title2], [title3], ..., [titlen], and has recently purchased the associated items [title1], [title2], ..., then will the business user click on the target item [now_click_title].

[0071] It should be noted that, in the actual item recommendation process, candidate items can adopt prompt information in any of the above four forms; or, in the actual item recommendation process, candidate items can adopt prompt information in the target form. The target form is a form selected from the above four forms in the model calibration link that can enable the model to have better reasoning effect (for example, higher reasoning accuracy). This is conducive to improving the accuracy of item recommendations and improving the accuracy of business processing.

[0072] S302, calling the item processing model in the model application reasoning framework, and predicting the association between the business user and the target item based on the prompt information of the target item; the item processing model is trained in the model training reasoning framework and reused in the model application reasoning framework; the processing speed of the item processing model in the model application reasoning framework is higher than the processing speed of the item processing model in the model training reasoning framework.

[0073] The model application inference framework is the inference framework used by the item processing model during the application phase. This framework is a technical framework used to make predictions based on a trained algorithm model. The model training inference framework is the inference framework used during the training phase of the item processing model. The degree of association between a business user and a target item represents the degree of association between the business user and the target item.

[0074] After obtaining the target item's prompt information, the item processing model can be invoked within the model application reasoning framework. Based on the target item's prompt information, the item processing model is predicted to determine the association between the business user and the target item, thereby determining the degree of association between the business user and the target item. The item processing model is trained within the model training reasoning framework and reused within the model application reasoning framework. This means that the item processing model is trained within the model training reasoning framework, while the item processing model is reasoned within the model application reasoning framework. The item processing model's processing speed within the model application reasoning framework is higher than that within the model training reasoning framework.

[0075] The item processing model is a large-scale generative language model based on the Transformer Encoder-Decoder architecture, while the model required for the item recommendation business is a classification model, and the item recommendation business requires the item processing model to output the probability of whether the business user visits the candidate item (that is, it is necessary to output the correlation between the business user and the candidate item). Therefore, the embodiment of the present application transforms the item processing model and transforms the original large-scale generative language model into a classification model adapted to the item recommendation business. Because the training framework used in the training process of the item processing model in the embodiment of the present application is different from the reasoning framework used in the reasoning process of the item processing model, the embodiment of the present application adaptively transforms the item processing model in the training framework and the reasoning framework respectively. The following introduces the adaptive transformation of the item processing model in the training process and the adaptive transformation of the item processing model in the reasoning process:

[0076] (1) Training process of item processing model:

[0077] The item processing model can be a model obtained by pre-training on a large-scale multilingual corpus. The training of the item processing model mentioned in the embodiment of the present application can be understood as fine-tuning the item processing model. As shown in Figure 4, during the training process of the item processing model, the item processing model may include an encoder and a decoder. The encoder can be used to encode the input information of the item processing model, and the decoder can be used to decode the encoding result of the encoder to generate a predicted associated label. The adaptive transformation of the item processing model during the training process may include: the embodiment of the present application transforms the decoder part of the item processing model, transforms the original generative language model into a classification model adapted to the item recommendation business, and the original decoder outputs the token of the current round (i.e., the predicted associated label). The embodiment of the present application transforms the decoder, adds a probability prediction module (softmax), and outputs the probability value of the estimated token (i.e., the degree of association under the predicted associated label) through the probability prediction module (softmax).

[0078] After adaptive transformation, the training process of the item processing model may include: obtaining an initial processing model (i.e., the initial item processing model mentioned above) and sample data. The sample data may include a training set for training the initial processing model. The training set may include multiple training samples. Each training sample may include prompt information (prompt) of the training item and a marked associated label corresponding to the training item (original sample label or answer). The initial processing model can be trained based on the training set under the model training inference framework to obtain an item processing model.

[0079] In more detail, the training process of the item processing model is iterative. The iteration can be understood as using the first training sample to perform the first training on the initial processing model, using the second training sample to perform the second training on the item processing model obtained by the first training, and so on, until the training termination condition is reached. The item processing model that meets the training termination condition is used as the final trained item processing model. Any training process in the iterative training process may include: inputting the prompt information of the training item in the training sample (which may be called the sample prompt) into the item processing model, and dividing the prompt information of the training item into multiple words (word 1-word n ​​as shown in Figure 4) through the word segmentation module (tokenizer); outputting the contextual semantic information (embedding) of each word in the prompt information of the training item through the encoder (Encoder) of the item processing model, and then generating an estimated token (i.e., a predicted associated label) and a token probability (i.e., the degree of association under the predicted associated label) through the decoder (Decoder); then, the loss information (loss) can be calculated based on the estimated token (i.e., the predicted associated label), the labeled associated label of the training item, and the token probability (i.e., the degree of association under the predicted associated label). The embodiment of the present application does not limit the type of loss function used to calculate the loss information. For example, the loss function used to calculate the loss information can be a cross-entropy loss function; then, the model parameters of the encoder (Encoder) and decoder (Decoder) in the item processing model can be adjusted through model back propagation based on the loss information (loss).

[0080] (2) Reasoning process of the item handling model:

[0081] The item processing model uses different frameworks during training and inference. After training, the item processing model is deployed in the inference framework (i.e., the model application inference framework). Similar to the training process, the inference process of the item processing model requires the transformation of the original large-scale generative language model into a classification model suitable for the item recommendation business. As shown in Figure 5, during the inference process of the item processing model, the structure of the item processing model before transformation within the model application inference framework can include:

[0082] ① n1 encoders: Each encoder has the same structure and can encode the encoder inputs. Furthermore, each decoder can include a self-attention module and a feedforward neural network module. The self-attention module can learn the associations between words in the encoded input, and the feedforward neural network module can perform nonlinear mapping on the output of the self-attention module, further learning the output of the self-attention module.

[0083] ②Decoder: The decoder can include n2 decoding modules (decoders) and label prediction modules (beam search). The n2 decoding modules (decoders) have the same structure. The decoding modules can be used to combine the decoding input information (decoding inputs) and the encoding results (encoder outputs) of n1 encoders for decoding processing. The label prediction module (beam search) can be used to predict associated labels based on the decoding results of the n2 decoding modules (decoders) to generate predicted associated labels. Furthermore, any decoding module (decoder) may include a self-attention mechanism module (Self-Attention), a cross-attention mechanism module (encoder-decoder Cross-Attention) and a feedforward neural network module (Feed Forward Network); wherein, the self-attention mechanism module (Self-Attention) can be used to learn the correlation between each word in the decoding input information, the cross-attention mechanism module (encoder-decoder Cross-Attention) can learn the correlation between the encoding input information and the decoding input information, and the feedforward neural network module can be used to perform nonlinear mapping on the output result of the cross-attention mechanism module, and further learn the output result of the cross-attention mechanism module.

[0084] As shown in Figure 6, the adaptive transformation of the item processing model during inference is indicated by a dotted box. This adaptive transformation may include: For the item processing model, in the first architecture, the item processing model decoder (Decoding) only outputs the generated token (i.e., the predicted associated label) and does not support classification models or output token probabilities (i.e., the degree of association under the predicted associated label). To align with the transformation of the item processing model during training, the label prediction module (Beam Search) can be modified using a probability prediction module (Softmax). After this transformation, the probability prediction module (Softmax) and the label prediction module (Beam Search) are combined into a prediction module. The softmax value of the original token can be obtained as the token probability and output, i.e., the output is [output id, output prob], where output id is the predicted associated label and output prob is the degree of association under the predicted associated label.

[0085] After adaptive transformation, the reasoning process of the item processing model may include: taking the target item as an example, the encoded input information of the item processing model is the prompt information of the target item. After the prompt information of the target item is input into the item processing model, it is divided into multiple words, and the representation information of each word in the prompt information of the target item can be obtained. The representation information of any word refers to a vector representation that can be used to uniquely represent the word; in the model application reasoning framework, the encoder is called to perform semantic analysis on each word based on the correlation between the representation information of each word to obtain the contextual semantic information of each word; in the model application reasoning framework, the decoding module (specifically, the cross-attention mechanism module and the feedforward neural network module of the decoding module) is called to generate a decoded representation of the decoding input information based on the correlation between the contextual semantic information of each word and the contextual semantic information of the decoding input information (the contextual semantic information of the decoding input information can be obtained through the self-attention mechanism module of the decoding module); the prediction module can be called in the model application reasoning framework to perform association prediction between the business user and the target item based on the decoded representation of the decoded input information to obtain the degree of association between the business user and the target item.

[0086] In more detail, the process of calling the prediction module in the model application reasoning framework to predict the association between the business user and the target item based on the decoded representation of the decoded input information and obtaining the degree of association between the business user and the target item may include: calling the label prediction module in the prediction module in the model application reasoning framework to predict the association label between the business user and the target item based on the decoded representation of the decoded input information and obtaining the predicted association label between the business user and the target item; calling the probability prediction module in the prediction module in the model application reasoning framework to predict the association probability between the business user and the target item based on the decoded representation of the decoded input information and obtaining multiple candidate probabilities under the predicted association label; selecting the degree of association between the business user and the target item from the multiple candidate probabilities, specifically selecting the largest candidate probability from the multiple candidate probabilities as the degree of association between the business user and the target item.

[0087] In addition to modifying the structure of the item processing model during its inference process, the present embodiment also optimizes the memory cache by caching the information of business users, the features of associated items, and the features of candidate items that have appeared in the prompt information. Representation caching refers to caching the representation information of words in the information of business users, the features of associated items, and the features of candidate items that have appeared. When repeated words appear, the cached representation information can be directly used without performing representation analysis again, reducing the encoding time of the encoder in the item processing model. In other words, the present embodiment adds a cache module before the encoder of the item processing model. For the item recommendation business, the material library is prioritized. The business user information, the features of associated users, and the features of candidate items in multiple requests from different business users may be repeated. Therefore, the representation information of the item features and business user information can be calculated in advance, reducing the calculation time of real-time representation analysis.

[0088] The process of mapping the memory cache transformation to each word included in the prompt information of the target item and obtaining the representation information of each word in the prompt information of the target item may include: obtaining a representation information library (i.e., a cache module), which may include representation information of multiple words (i.e., words that have appeared); searching the representation information of each word in the prompt information of the target item in the representation information library; if a word matching the first word in the prompt information of the target item is found in the representation information library, the representation information of the matching word may be determined as the representation information of the first word; and performing a representation analysis on the second word in the prompt information of the target item other than the first word to obtain the representation information of the second word. That is, among the words included in the prompt information of the target item, for words that have been subjected to representation analysis, their representation information is cached in the cache module and can be directly obtained in the cache module. For words that have not been subjected to representation analysis, their representation information does not exist in the cache module and needs to be subjected to representation analysis.

[0089] As described above, before the item processing model can officially recommend items online, it must undergo a model training phase and a model testing phase. After the item processing model passes the model testing phase, it can officially recommend items online. The following details the model testing process for the item processing model: The sample data may also include a test set for testing the trained item processing model. The test set may include multiple test samples, each of which may include prompt information for the test item and a test association label corresponding to the test item. The model performance of the item processing model can be evaluated based on the test set within a model training inference framework to obtain first evaluation information. The model performance of the item processing model can be evaluated based on the test set within a model application inference framework to obtain second evaluation information. If the difference between the first evaluation information and the second evaluation information is less than a difference threshold, the item processing model test can be determined to have passed. The item processing model can then be invoked within the model application inference framework to predict associations between business users and each candidate item recalled for the business user. If the difference between the first evaluation information and the second evaluation information is greater than or equal to the difference threshold, the item processing model test can be determined to have failed, and further adjustments to the item processing model are required.

[0090] Among them, the evaluation information (i.e., the first evaluation information and the second evaluation information) may specifically refer to the estimated AUC value. The process of evaluating the model performance of the item processing model based on the test set under the model training and reasoning framework is similar to the process of evaluating the model performance of the item processing model based on the test set under the model application and reasoning framework. Here, the process of evaluating the model performance of the item processing model based on the test set under the model training and reasoning framework is introduced, which may specifically include: the test samples in the test set may include positive test samples and negative test samples, the positive test samples refer to the test items that the test user is willing to access, and the negative test samples refer to the test items that the test user refuses to access; the item processing model can be called under the model training and reasoning framework. An association prediction is performed on the test samples in the test set and the corresponding test users to obtain a predicted association label; the proportion of first samples that are correctly judged as positive in the positive test samples can be counted, and the first sample proportion can specifically be the ratio between the positive test samples that are correctly judged as positive in the positive test samples and all the positive test samples; the proportion of second samples that are incorrectly judged as positive in the negative test samples can be counted, and the second sample proportion can specifically be the ratio between the negative test samples that are incorrectly judged as positive in the negative test samples and all the negative test samples; then, first evaluation information can be determined based on the ratio between the first sample proportion and the second sample proportion; wherein, positive means that the predicted association label indicates that the test user is willing to access the corresponding test item.

[0091] S303: Perform business processing on the target item based on the correlation between the business user and the target item.

[0092] In this embodiment, target items can be categorized based on the degree of association between the business user and the target item, allowing for different business processing to be performed on target items categorized into different categories. Business processing is specific business processing performed on the target item by the business user. Specifically, it can include recommending the target item, deciding not to recommend the target item, binding the target item to the business user, or unbinding the target item from the business user.

[0093] In a business (i.e., an item recommendation business), when a business user accesses the business, the business service provider can recall multiple candidate items for the business user, and can call the item processing model under the model application reasoning framework to predict the association between the business user and each candidate item, thereby obtaining a predicted association label and the degree of association under the predicted association label between the business user and each candidate item; wherein, the predicted association label of any candidate item can be used to indicate that the business user is willing to access the corresponding candidate item, or that the business user refuses to access the corresponding candidate item. Furthermore, among the candidate items recalled for the business user, a predicted association label can be selected from multiple candidate items to indicate the intended candidate item that the business user is willing to access, and then the intended candidate items can be sorted based on the degree of association of the intended candidate items, and the intended candidate items that are arranged before the target position among the intended candidate items can be used as recommended items for the business user, so that the recommended items can be recommended to the business user.

[0094] That is to say, for the target item, based on the association between the business user and the target item, business processing of the target item may include: if the predicted association tag corresponding to the target item indicates that the business user is willing to access the target item, then the arrangement position of the target item in the intended candidate items may be checked, and if the arrangement position of the target item in the intended candidate items is before the target position, then the target item may be recommended to the business user as a recommended item; if the predicted association tag corresponding to the target item indicates that the business user refuses to access the target item, then the target item may be refused to be recommended to the business user; if the arrangement position of the target item in the intended candidate items is the target position or after the target position, then the target item may be refused to be recommended to the business user.

[0095] In the embodiments of the present application, by adaptively modifying the item processing model, which was originally a large-scale generative language model, the item processing model can be transformed into a classification model suitable for the item recommendation business, making it possible to implement large-scale generative language models in the item recommendation business. Furthermore, by separating the inference framework of the item processing model's training phase from the inference framework of the application phase, an inference acceleration framework that can increase the processing speed of the item processing model is used during the inference phase of the item processing model, thereby improving the efficiency of item recommendation and thus improving the efficiency of business processing. In addition, during the inference process of the item processing model, by caching the representation information of words that have appeared in the prompt information, the representation information can be directly obtained when the word appears again, without consuming time-consuming representation analysis, further improving the model inference speed, improving the efficiency of item recommendation, and thus improving the efficiency of business processing.

[0096] This embodiment of the present application provides a business processing method that introduces a method for accelerating an item processing model using a model application reasoning framework, as well as a concurrent mechanism for the item processing model. This business processing method can be executed by a computer device, such as the item recommendation device 102 in the business processing system shown in FIG1 , or the integrated device of the recommendation request device 101 and the item recommendation device 102 . As shown in FIG7 , this business processing method may include, but is not limited to, the following steps S701 to S704:

[0097] S701 : Acquire prompt information of a target item recalled for a business user. The prompt information of the target item includes information of the business user, features of at least one associated item associated with the business user, and features of the target item.

[0098] The execution process of step S701 in the embodiment of the present application is the same as the execution process of step S301 in the embodiment shown in FIG3 above. For the specific execution process, please refer to the description of step S301 in the embodiment shown in FIG3 above, which will not be repeated here.

[0099] S702: Optimize the item processing model according to the optimization method defined by the model application reasoning framework.

[0100] S703 , calling the optimized item processing model to predict the association between the business user and the target item based on the prompt information of the target item, and obtaining the association degree between the business user and the target item.

[0101] In steps S702-S703, as described above, the processing speed of the item processing model in the model application reasoning framework is higher than the processing speed of the item processing model in the model training reasoning framework. This is because the model application reasoning framework defines the optimization method for the item processing model to perform association prediction. The item processing model can be optimized according to the optimization method defined by the model application reasoning framework. Then, the optimized item processing model is called to perform association prediction between the business user and the target item based on the prompt information of the target item, and the association degree between the business user and the target item is obtained. This can improve the efficiency of the item processing model in performing association prediction. The optimization method of the model application reasoning framework is mainly reflected in either or both of operator fusion and matrix calculation optimization. The following introduces the optimization methods of these two model application reasoning frameworks respectively:

[0102] (1) Operator fusion:

[0103] The association prediction logic of the item processing model can be implemented by the cooperation between operators at each layer in the item processing model. Each layer of operators needs to call the kernel of the computer device CPU (Central Processing Unit) to perform operations. If the number of layers of operators in the item processing model is too large, calling the kernel and the interaction between the kernels will consume a lot of time. Therefore, the operators at each layer in the item processing model can be fused. Specifically, the optimization method of the model application reasoning framework can include the fusion method between operators at each layer in the item processing model. In this case, optimizing the item processing model according to the optimization method can include: fusing the operators at each layer of the item processing model according to the fusion method; calling the optimized item processing model based on the prompt information of the target item to predict the association between the business user and the target item, and obtaining the association degree between the business user and the target item. This can include: calling the fused item processing model based on the prompt information of the target item to predict the association between the business user and the target item, and obtaining the association degree between the business user and the target item.

[0104] Furthermore, the operators at each level of the item processing model may include both matrix and non-matrix operators. Matrix operations are the core operations for performing association prediction in the item processing model, and therefore require significant time to perform. In this case, fusing the operators at each level of the item processing model in a fusion manner may include retaining the matrix operators unchanged and fusing the non-matrix operators.

[0105] That is to say, the embodiment of the present application can fuse some operators with shorter computing time in the item processing model, reduce the time consumed by these operators to call the kernel separately and the time consumed by the kernels to interact with each other, thereby improving the association prediction efficiency of the item processing model, which can improve the item recommendation efficiency and thus improve the business processing efficiency.

[0106] (2) Matrix calculation optimization:

[0107] As described above, matrix operation is the core operation of the item processing model for association prediction. Matrix operation takes a long time in the association prediction process of the item processing model. Therefore, optimizing the matrix operation algorithm and improving the matrix operation speed can improve the efficiency of the association prediction of the item processing model. The optimization method of the model application reasoning framework may include a matrix operation optimization method of the matrix operation operator in the item processing model. In this case, optimizing the item processing model according to the optimization method may include: optimizing the matrix operation algorithm of the matrix operation operator according to the matrix operation optimization method. In other words, the embodiment of the present application can optimize the matrix operation algorithm of the matrix operation operator in the item processing model, which can improve the matrix operation speed of the matrix operation operator, improve the item recommendation efficiency, and thus improve the business processing efficiency.

[0108] S704: Perform business processing on the target item based on the correlation between the business user and the target item.

[0109] The execution process of step S704 in the embodiment of the present application is the same as the execution process of step S303 in the embodiment shown in FIG3 above. For the specific execution process, please refer to the description of step S303 in the embodiment shown in FIG3 above, which will not be repeated here.

[0110] On the basis of the above steps S701-S704, in order to further improve the efficiency of item recommendation, the embodiment of the present application can support a concurrent mechanism. The concurrent mechanism supported by the embodiment of the present application can include any one of the first concurrent mechanism, the second concurrent mechanism, and the third concurrent mechanism; wherein the first concurrent mechanism refers to the concurrent association prediction of a single business user with different candidate items, the second concurrent mechanism refers to the concurrent association prediction of a single business user with candidate items in different business scenarios, and the third concurrent mechanism refers to the concurrent association prediction of multiple business users. The following introduces the different concurrent mechanisms supported by the embodiment of the present application:

[0111] (1) The first concurrent mechanism:

[0112] As shown in Figure 8, the target item is any one of multiple candidate items recalled for a business user. These multiple candidate items can be divided into N candidate item sets (as shown in Figure 8: candidate item set 1 - candidate item set N), where N is an integer greater than 1. In this case, the item processing model can be invoked within the model application reasoning framework to perform association prediction for each of the N candidate item sets in sequence. For each candidate item set, the item processing model can be invoked within the model application reasoning framework to concurrently perform association prediction for the business user and each candidate item in the candidate item set, obtaining the degree of association between the business user and each candidate item in the candidate item set. In other words, the association prediction process for each candidate item in each candidate item set is concurrent. Furthermore, the business user and the candidate items in each candidate item set can each form a request packet, resulting in a total of N "business user-candidate item set" request packets. Each request packet concurrently requests the item processing model, which then processes multiple candidate items in the request packet in parallel.

[0113] Through the first concurrent mechanism, association predictions can be made for a single business user and different candidate items concurrently, rather than sequentially. This can improve the efficiency of item recommendations for a single business user, thereby improving the business processing efficiency of a single business user.

[0114] (2) Second concurrent mechanism:

[0115] As shown in Figure 9, the target item is any one of multiple candidate items recalled for a business user. The multiple candidate items belong to N different business scenarios of the same business (as shown in Figure 9, Business Scenario 1 - Business Scenario N). The number of candidate items included in each of the N business scenarios can be the same or different, with N being an integer greater than 1. N item processing models are deployed within the model application reasoning framework, each corresponding to each of the N business scenarios. In this case, the N item processing models can be invoked within the model application reasoning framework to predict the association between the business user and each candidate item in the corresponding business scenario based on the prompt information for each candidate item in the corresponding business scenario, thereby obtaining the degree of association between the business user and each candidate item in the corresponding business scenario. The association prediction processes of the N item processing models are performed in parallel. In other words, the business user and the candidate items in each business scenario can each form a request packet, resulting in a total of N "business user - candidate item in the business scenario" request packets. These N item processing models are then requested in parallel, and the N item processing models process the N request packets in parallel.

[0116] It's important to note that N item processing models can be deployed on the same device or on N different devices. When N item processing models are deployed on the same device, it can be understood that each item processing model acts as an item processing service on the device, and the N item processing services process N request packets in parallel. When N item processing models are deployed on N different devices, it can be understood that the computer device is a distributed system consisting of N devices, and the N devices process N request packets in a distributed and parallel manner.

[0117] Through the second concurrent mechanism, association predictions can be made for a single business user and candidate items in different business scenarios concurrently, rather than making association predictions for a single business user and candidate items in each business scenario in turn. This can improve the efficiency of recommending items to a single business user in multiple business scenarios at the same time, thereby improving the business processing efficiency of a single business user.

[0118] (3) The third concurrent mechanism:

[0119] As shown in Figure 10, the business service provider has retrieved candidate items for N business users (Business User 1 - Business User N, as shown in Figure 10), where N is an integer greater than 1. N item processing models are deployed within the model application reasoning framework, each corresponding to one of the N business users. In this scenario, the N item processing models can be invoked within the model application reasoning framework to perform association predictions for their corresponding business users. The association prediction processes for the N item processing models are performed in parallel. In other words, each of the N business users and the candidate items corresponding to each business user forms a request packet, resulting in a total of N "business user-candidate item" request packets. These requests are then made to the N item processing models in parallel, and the N item processing models process the N request packets in parallel.

[0120] It's important to note that N item processing models can be deployed on the same device or on N different devices. When N item processing models are deployed on the same device, it can be understood that each item processing model acts as an item processing service on the device, and the N item processing services process N request packets in parallel. When N item processing models are deployed on N different devices, it can be understood that the computer device is a distributed system consisting of N devices, and the N devices process N request packets in a distributed and parallel manner.

[0121] Through the third concurrent mechanism, item recommendations can be made to multiple business users concurrently, rather than to each business user in turn. This can support more business users to access item recommendations at the same time, and improve the efficiency of the business service provider in recommending items to multiple business users.

[0122] In the embodiment of the present application, the model application reasoning framework defines an optimization method for the item processing model to perform association prediction (including operator fusion and matrix calculation optimization). After the item processing model is optimized according to the optimization method defined by the model application reasoning framework, the optimized item processing model is called to perform association prediction, which can improve the efficiency of the item processing model in performing association prediction, improve the efficiency of item recommendation, and thus improve the efficiency of business processing. In addition, based on the concurrency mechanism, the embodiment of the present application can concurrently perform association predictions for a single business user and different candidate items, thereby improving the efficiency of item recommendation for a single business user; based on the concurrency mechanism, the embodiment of the present application can also concurrently perform association predictions for a single business user and candidate items in different business scenarios, thereby improving the efficiency of item recommendation for a single business user in multiple business scenarios at the same time; based on the concurrency mechanism, the embodiment of the present application can also concurrently perform item recommendation for multiple business users, thereby improving the efficiency of the business service provider in recommending items to multiple business users at the same time.

[0123] The above describes in detail the method of the embodiment of the present application. In order to facilitate better implementation of the above scheme of the embodiment of the present application, the device of the embodiment of the present application is provided below accordingly.

[0124] Please refer to Figure 11, which is a structural diagram of a business processing device provided in an embodiment of the present application. The business processing device can be set in the computer device provided in the embodiment of the present application. The computer device can be, for example, the item recommendation device 102 in the business processing system shown in Figure 1, or it can be an integrated device of the recommendation request device 101 and the item recommendation device 102 in the business processing system shown in Figure 1. The business processing device shown in Figure 11 can be a computer program running on a computer device. The business processing device can be used to execute some or all of the steps in the method embodiment shown in Figure 3 or Figure 7. Please refer to Figure 11. The business processing device may include the following units:

[0125] An acquiring unit 1101 is configured to acquire prompt information of a target item recalled by a business user, where the prompt information of the target item includes information of the business user, characteristics of at least one associated item associated with the business user, and characteristics of the target item;

[0126] Processing unit 1102 is configured to invoke the item processing model within the model application reasoning framework and predict the relevance between the business user and the target item based on the target item's prompt information. The item processing model is trained within the model training reasoning framework and reused within the model application reasoning framework. The item processing model is processed faster within the model application reasoning framework than within the model training reasoning framework.

[0127] The processing unit 1102 is further configured to perform business processing on the target item based on the association between the business user and the target item.

[0128] In one implementation, a model application reasoning framework defines an optimization method for performing association prediction on an item processing model. Processing unit 1102 is configured to invoke the item processing model within the model application reasoning framework to perform association prediction between a business user and a target item based on prompt information of the target item. When the degree of association between the business user and the target item is determined, the processing unit 1102 is configured to execute the following steps:

[0129] Optimize the item processing model according to the optimization method;

[0130] The optimized item processing model is called to predict the association between the business user and the target item based on the prompt information of the target item, and the association degree between the business user and the target item is obtained.

[0131] In one implementation, the optimization method includes a fusion method between operators at various layers in the item processing model;

[0132] The processing unit 1102 is configured to optimize the item processing model according to the optimization method, and is specifically configured to perform the following steps:

[0133] Fuse the operators of each layer of the item processing model according to the fusion method;

[0134] Processing unit 1102 is used to call the optimized item processing model to predict the association between the business user and the target item based on the prompt information of the target item, and obtain the correlation between the business user and the target item. It is specifically used to perform the following steps: call the fused item processing model to predict the association between the business user and the target item based on the prompt information of the target item, and obtain the correlation between the business user and the target item.

[0135] In one implementation, the optimization method includes a matrix operation optimization method of a matrix operation operator in the item processing model;

[0136] The processing unit 1102 is configured to optimize the item processing model according to the optimization method, and is specifically configured to perform the following steps:

[0137] According to the matrix operation optimization method, the matrix operation algorithm of the matrix operation operator is optimized.

[0138] In one implementation, the item processing model includes an encoder and a decoder, and the decoder includes a decoding module and a prediction module. After inputting the prompt information of the target item into the item processing model, the prompt information is divided into multiple words. The processing unit 1102 is configured to call the item processing model within the model application reasoning framework and predict the association between the business user and the target item based on the prompt information of the target item, specifically performing the following steps:

[0139] Obtain the representation information of each word in the prompt information of the target item; call the encoder under the model application reasoning framework to perform semantic analysis on each word based on the correlation between the representation information of each word, and obtain the contextual semantic information of each word; call the decoding module under the model application reasoning framework to generate a decoded representation of the decoded input information based on the correlation between the contextual semantic information of each word and the contextual semantic information of the decoded input information; call the prediction module under the model application reasoning framework to perform association prediction between the business user and the target item based on the decoded representation of the decoded input information, and obtain the degree of association between the business user and the target item.

[0140] In one implementation, the prediction module is obtained by modifying the label prediction module using a probability prediction module. The processing unit 1102 is configured to call the prediction module under the model application reasoning framework to perform association prediction between the business user and the target item based on the decoded representation of the decoded input information. When the association between the business user and the target item is obtained, the processing unit 1102 is configured to perform the following steps:

[0141] In the model application reasoning framework, the label prediction module is called to predict the associated label between the business user and the target item based on the decoded representation of the decoded input information, and the predicted associated label between the business user and the target item is obtained. In the model application reasoning framework, the probability prediction module is called to predict the associated probability between the business user and the target item based on the decoded representation of the decoded input information, and multiple candidate probabilities under the predicted associated label are obtained.

[0142] Select the relevance between the business user and the target item from multiple candidate probabilities.

[0143] In one implementation, the processing unit 1102 is configured to obtain representation information of each word in the prompt information of the target item by specifically executing the following steps: obtaining a representation information library, the representation information library including representation information of multiple words; searching the representation information library for representation information of each word in the prompt information of the target item; if a word matching the first word in the prompt information of the target item is found in the representation information library, determining the representation information of the matching word as the representation information of the first word; and performing representation analysis on a second word other than the first word in the prompt information of the target item to obtain representation information of the second word.

[0144] In one implementation, a target item is any one of multiple candidate items recalled for a business user; the multiple candidate items belong to N different business scenarios of the same business, where N is an integer greater than 1; N item processing models are deployed under the model application reasoning framework, and the N item processing models correspond one-to-one to the N business scenarios;

[0145] The processing unit 1102 is further configured to execute the following steps: calling N item processing models within a model application reasoning framework to predict the association between the business user and each candidate item in the corresponding business scenario based on the prompt information of each candidate item in the corresponding business scenario, thereby obtaining the degree of association between the business user and each candidate item in the corresponding business scenario; wherein the association prediction processes of the N item processing models are performed in parallel.

[0146] In one implementation, the processing unit 1102 is further used to perform the following steps: obtaining an initial processing model and sample data, the sample data including a training set and a test set; training the initial processing model based on the training set under a model training reasoning framework to obtain an item processing model; evaluating the model performance of the item processing model based on the test set under the model training reasoning framework to obtain first evaluation information; evaluating the model performance of the item processing model based on the test set under the model application reasoning framework to obtain second evaluation information; if the difference between the first evaluation information and the second evaluation information is less than a difference threshold, calling the item processing model under the model application reasoning framework to perform association prediction between the business user and each candidate item recalled for the business user, where the target item is any candidate item recalled for the business user.

[0147] In one implementation, the test samples in the test set include positive test samples and negative test samples; positive test samples refer to test items that the test user is willing to access, and negative test samples refer to test items that the test user refuses to access; processing unit 1102 is configured to evaluate the model performance of the item processing model based on the test set in the model training and inference framework, and upon obtaining first evaluation information, specifically to perform the following steps:

[0148] In the model training and inference framework, the item processing model is called to predict the association between the test samples in the test set and the corresponding test users to obtain the predicted association labels; the proportion of first samples that are correctly judged as positive in the positive test samples is counted; the proportion of second samples that are incorrectly judged as positive in the negative test samples is counted; and the first evaluation information is determined based on the ratio between the first sample proportion and the second sample proportion; wherein, positive means that the predicted association label indicates that the test user is willing to access the corresponding test item.

[0149] According to another embodiment of the present application, the various units in the business processing device shown in Figure 11 can be individually or all combined into one or several other units to form, or one (or some) of the units can be further divided into multiple functionally smaller units to form, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions. In actual applications, the functions of one unit can also be implemented by multiple units, or the functions of multiple units can be implemented by one unit. In other embodiments of the present application, the business processing device may also include other units. In actual applications, these functions can also be implemented with the assistance of other units, and can be implemented by the collaboration of multiple units.

[0150] According to another embodiment of the present application, a business processing device as shown in FIG11 can be constructed, and the business processing method of the embodiment of the present application can be implemented by running a computer program capable of executing the steps involved in some or all of the methods shown in FIG3 or FIG7 on a general-purpose computing device such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), a random access memory medium (RAM), and a read-only memory medium (ROM). The computer program can be recorded on, for example, a computer-readable storage medium, and loaded into the above-mentioned computing device via the computer-readable storage medium and run therein.

[0151] In an embodiment of the present application, a target item is recalled by a business user. After obtaining prompt information about the target item being recalled by the business user, the item processing model can be invoked within the model application reasoning framework to perform association detection between the business user and the target item based on the prompt information of the target item, thereby obtaining a correlation between the business user and the target item. Based on this correlation, business processing can be performed on the target item. The item processing model can be trained within the model training reasoning framework and reused within the model application reasoning framework. The processing speed of the item processing model within the model application reasoning framework is higher than that within the model training reasoning framework. It is not difficult to see that the item processing model can be used for business processing within a business. The model application reasoning framework can be understood as the reasoning framework used by the item processing model during the model reasoning process, and the model training reasoning framework can be understood as the training framework used by the item processing model during the model training process. In an embodiment of the present application, the training framework and the reasoning framework of the item processing model can be separated. During the reasoning phase of the item processing model, a reasoning acceleration framework that can increase the processing speed of the item processing model can be used, thereby improving business processing efficiency.

[0152] Based on the above method and device embodiments, an embodiment of the present application provides a computer device. Please refer to Figure 12, which is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. The computer device shown in Figure 12 includes at least a processor 1201, an input interface 1202, an output interface 1203, and a computer-readable storage medium 1204. The processor 1201, input interface 1202, output interface 1203, and computer-readable storage medium 1204 may be connected via a bus or other means.

[0153] Computer-readable storage medium 1204 may be stored in a memory of a computer device. Computer-readable storage medium 1204 is used to store a computer program, which includes computer instructions. Processor 1201 is used to execute the computer program stored in computer-readable storage medium 1204. Processor 1201 (or CPU (Central Processing Unit)) is the computing and control core of the computer device and is suitable for implementing computer programs, specifically loading and executing computer programs to implement corresponding method flows or corresponding functions.

[0154] The embodiment of the present application also provides a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space that stores the operating system of the computer device. In addition, a computer program suitable for being loaded and executed by the processor is also stored in the storage space. It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory (Non-Volatile Memory), such as at least one disk storage; optionally, it can also be at least one computer-readable storage medium located away from the aforementioned processor.

[0155] The computer device may be, for example, the item recommendation device 102 in the business processing system shown in FIG1 , or may be an integrated device of the recommendation request device 101 and the item recommendation device 102 in the business processing system shown in FIG1 . In a specific implementation, the computer program stored in the computer-readable storage medium 1204 may be loaded and executed by the processor 1201 to implement the corresponding steps in the business processing method shown in FIG3 or FIG7 . In a specific implementation, the computer program in the computer-readable storage medium 1204 is loaded by the processor 1201 and executes the following steps:

[0156] Obtaining prompt information of a target item recalled for a business user, the prompt information of the target item including information of the business user, characteristics of at least one associated item associated with the business user, and characteristics of the target item;

[0157] The item processing model is called within the model application reasoning framework to predict the association between the business user and the target item based on the target item's prompt information. The item processing model is trained within the model training reasoning framework and reused within the model application reasoning framework. The item processing model is processed faster within the model application reasoning framework than within the model training reasoning framework.

[0158] Based on the correlation between the business user and the target item, business processing is performed on the target item.

[0159] In one implementation, the model application reasoning framework defines an optimization method for the item processing model to perform association prediction; the computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201 to call the item processing model under the model application reasoning framework, and predict the association between the business user and the target item based on the prompt information of the target item, specifically for performing the following steps: optimizing the item processing model according to the optimization method; calling the optimized item processing model to perform association prediction between the business user and the target item based on the prompt information of the target item, and obtaining the association between the business user and the target item.

[0160] In one implementation, the optimization method includes a fusion method between operators at various layers in the item processing model;

[0161] When the computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201 to optimize the item processing model according to the optimization method, the computer program is specifically configured to perform the following steps: fusing the operators of each layer of the item processing model according to the fusion method;

[0162] The computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201 to call the optimized item processing model to predict the association between the business user and the target item based on the prompt information of the target item, and to obtain the association between the business user and the target item. Specifically, the computer program is used to perform the following steps:

[0163] The fused item processing model is called to predict the association between the business user and the target item based on the prompt information of the target item, and the association degree between the business user and the target item is obtained.

[0164] In one implementation, the optimization method includes a matrix operation optimization method of a matrix operation operator in the item processing model;

[0165] When the computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201 to optimize the item processing model according to the optimization method, it is specifically used to perform the following steps:

[0166] According to the matrix operation optimization method, the matrix operation algorithm of the matrix operation operator is optimized.

[0167] In one implementation, the item processing model includes an encoder and a decoder, and the decoder includes a decoding module and a prediction module. After inputting the prompt information of the target item into the item processing model, the prompt information is divided into multiple words. The computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201. The computer program invokes the item processing model under the model application reasoning framework and predicts the association between the business user and the target item based on the prompt information of the target item, specifically performing the following steps:

[0168] Obtain the representation information of each word in the prompt information of the target item; call the encoder under the model application reasoning framework to perform semantic analysis on each word based on the correlation between the representation information of each word, and obtain the contextual semantic information of each word; call the decoding module under the model application reasoning framework to generate a decoded representation of the decoded input information based on the correlation between the contextual semantic information of each word and the contextual semantic information of the decoded input information; call the prediction module under the model application reasoning framework to perform association prediction between the business user and the target item based on the decoded representation of the decoded input information, and obtain the degree of association between the business user and the target item.

[0169] In one implementation, the prediction module is obtained by modifying the tag prediction module using a probability prediction module. The computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201. The prediction module is called under the model application reasoning framework to predict the association between the business user and the target item based on the decoded representation of the decoded input information. When the association between the business user and the target item is obtained, the prediction module is specifically configured to perform the following steps:

[0170] In the model application reasoning framework, the label prediction module is called to predict the associated label between the business user and the target item based on the decoded representation of the decoded input information, and the predicted associated label between the business user and the target item is obtained; in the model application reasoning framework, the probability prediction module is called to predict the associated probability between the business user and the target item based on the decoded representation of the decoded input information, and multiple candidate probabilities under the predicted associated label are obtained; the degree of association between the business user and the target item is selected from the multiple candidate probabilities.

[0171] In one implementation, when the computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201 to obtain the representation information of each word in the prompt information of the target item, the computer program is specifically configured to perform the following steps:

[0172] A representation information library is obtained, wherein the representation information library includes representation information of multiple words; representation information of each word in the prompt information of the target item is searched in the representation information library; if a word matching the first word in the prompt information of the target item is found in the representation information library, the representation information of the matching word is determined as the representation information of the first word; and a representation analysis is performed on a second word in the prompt information of the target item other than the first word to obtain representation information of the second word.

[0173] In one implementation, a target item is any one of multiple candidate items recalled for a business user; the multiple candidate items belong to N different business scenarios of the same business, where N is an integer greater than 1; N item processing models are deployed under the model application reasoning framework, and the N item processing models correspond one-to-one to the N business scenarios;

[0174] The computer program in the computer-readable storage medium 1204 is loaded by the processor 1201 and is further configured to perform the following steps:

[0175] Under the model application reasoning framework, N item processing models are called to perform association prediction between the business user and each candidate item in the corresponding business scenario based on the prompt information of each candidate item in the corresponding business scenario, and the association degree between the business user and each candidate item in the corresponding business scenario is obtained; among which, the association prediction process of the N item processing models is parallel.

[0176] In one implementation, the computer program in the computer-readable storage medium 1204 is loaded by the processor 1201 and is also used to perform the following steps: obtaining an initial processing model and sample data, the sample data including a training set and a test set; training the initial processing model based on the training set under a model training reasoning framework to obtain an item processing model; evaluating the model performance of the item processing model based on the test set under the model training reasoning framework to obtain first evaluation information; evaluating the model performance of the item processing model based on the test set under the model application reasoning framework to obtain second evaluation information; if the difference between the first evaluation information and the second evaluation information is less than a difference threshold, calling the item processing model under the model application reasoning framework to perform association prediction between the business user and each candidate item recalled for the business user, where the target item is any candidate item recalled for the business user.

[0177] In one implementation, the test samples in the test set include positive test samples and negative test samples; positive test samples refer to test items that the test user is willing to access, and negative test samples refer to test items that the test user refuses to access; the computer program in the computer-readable storage medium 1204 is loaded and executed by the processor 1201 to evaluate the model performance of the item processing model based on the test set under the model training and reasoning framework. When the first evaluation information is obtained, it is specifically used to perform the following steps: calling the item processing model under the model training and reasoning framework to predict the association between the test samples in the test set and the corresponding test users to obtain predicted association labels; counting the proportion of first samples in the positive test samples that are correctly judged as positive; counting the proportion of second samples in the negative test samples that are incorrectly judged as positive; determining the first evaluation information based on the ratio between the first sample ratio and the second sample ratio; wherein, positive means that the predicted association label indicates that the test user is willing to access the corresponding test item.

[0178] In an embodiment of the present application, a target item is recalled by a business user. After obtaining prompt information about the target item being recalled by the business user, an item processing model can be called within a model application reasoning framework. Based on the prompt information about the target item, the association between the business user and the target item is predicted. Business processing can then be performed on the target item based on the association between the business user and the target item. The item processing model can be trained within a model training reasoning framework and reused within the model application reasoning framework. The processing speed of the item processing model within the model application reasoning framework is higher than the processing speed of the item processing model within the model training reasoning framework. It is not difficult to see that the item processing model can be used for business processing within a business. The model application reasoning framework can be understood as the reasoning framework used by the item processing model during the model reasoning process, and the model training reasoning framework can be understood as the training framework used by the item processing model during the model training process. In an embodiment of the present application, the training framework and reasoning framework of the item processing model can be separated. During the reasoning phase of the item processing model, a reasoning acceleration framework that can increase the processing speed of the item processing model can be used, thereby improving business processing efficiency.

[0179] According to one aspect of the present application, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the service processing methods provided in the various optional embodiments described above.

[0180] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0181] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A business processing method, executed by a computer device, comprising: Acquire prompt information of a target item recalled for a business user, wherein the prompt information of the target item includes information of the business user, characteristics of at least one associated item associated with the business user, and characteristics of the target item; Invoking an item processing model in a model application reasoning framework to predict the association between the business user and the target item based on the prompt information of the target item; the item processing model is trained in a model training reasoning framework and reused in the model application reasoning framework; the processing speed of the item processing model in the model application reasoning framework is higher than the processing speed of the item processing model in the model training reasoning framework; and Based on the degree of association between the business user and the target item, business processing is performed on the target item.

2. The method of claim 1, wherein the model application reasoning framework defines an optimization method for the item processing model to perform association prediction; The calling of the item processing model under the model application reasoning framework and predicting the association between the business user and the target item based on the prompt information of the target item include: Optimizing the item processing model according to the optimization method; The optimized item processing model is called to predict the association between the business user and the target item based on the prompt information of the target item, so as to obtain the degree of association between the business user and the target item.

3. The method according to claim 2, wherein the optimization method comprises a fusion method between operators at each layer in the item processing model; The optimizing the item processing model according to the optimization method includes: Performing fusion processing on operators at each layer of the item processing model according to the fusion method; The calling of the optimized item processing model predicts the association between the business user and the target item based on the prompt information of the target item to obtain the association between the business user and the target item, including: The fused item processing model is called to predict the association between the business user and the target item based on the prompt information of the target item, so as to obtain the degree of association between the business user and the target item.

4. The method according to claim 2 or 3, wherein the optimization method comprises a matrix operation optimization method of a matrix operation operator in the item processing model; The optimizing the item processing model according to the optimization method includes: According to the matrix operation optimization method, the matrix operation algorithm of the matrix operation operator is optimized.

5. The method according to any one of claims 1 to 4, wherein the item processing model comprises an encoder and a decoder, and the decoder comprises a decoding module and a prediction module; the prompt information of the target item is divided into a plurality of words after being input into the item processing model; The calling of the item processing model under the model application reasoning framework and predicting the association between the business user and the target item based on the prompt information of the target item include: Obtaining representation information of each word in the prompt information of the target object; Calling the encoder under the model application reasoning framework to perform semantic analysis on each of the words based on the correlation between the representation information of each of the words to obtain contextual semantic information of each of the words; Calling the decoding module under the model application reasoning framework to generate a decoding representation of the decoding input information based on the association between the contextual semantic information of each word and the contextual semantic information of the decoding input information; The prediction module is called under the model application reasoning framework to predict the association between the business user and the target item based on the decoded representation of the decoded input information, so as to obtain the degree of association between the business user and the target item.

6. The method according to claim 5, wherein the prediction module is obtained by transforming the label prediction module using a probability prediction module; the calling of the prediction module under the model application reasoning framework to predict the association between the business user and the target item based on the decoded representation of the decoded input information to obtain the degree of association between the business user and the target item comprises: Call the decoding table of the label prediction module based on the decoding input information under the model application reasoning framework , predicting the associated label between the business user and the target item to obtain the predicted associated label between the business user and the target item; Calling the probability prediction module under the model application reasoning framework to predict the association probability between the business user and the target item based on the decoded representation of the decoded input information, and obtaining a plurality of candidate probabilities under the predicted association label; The degree of association between the business user and the target item is selected from the plurality of candidate probabilities.

7. The method according to claim 5 or 6, wherein obtaining the representation information of each word in the prompt information of the target object comprises: Acquire a representation information library, wherein the representation information library includes representation information of a plurality of words; Searching the representation information library for the representation information of each word in the prompt information of the target object; If a word matching the first word in the prompt information of the target item is found in the representation information library, the representation information of the matching word is determined as the representation information of the first word; A representation analysis is performed on a second word other than the first word in the prompt information of the target object to obtain representation information of the second word.

8. The method according to any one of claims 1 to 7, wherein the target item is any one of a plurality of candidate items recalled for the business user; the plurality of candidate items belong to N different business scenarios of the same business, where N is an integer greater than 1; N item processing models are deployed under the model application reasoning framework, and the N item processing models correspond one-to-one to the N business scenarios; The method further comprises: Calling the N item processing models under the model application reasoning framework to predict the association between the business user and each of the candidate items in the corresponding business scenario based on the prompt information of each of the candidate items in the corresponding business scenario, and obtaining the degree of association between the business user and each of the candidate items in the corresponding business scenario; The association prediction processes of the N item processing models are performed in parallel.

9. The method according to any one of claims 1 to 8, further comprising: Acquire an initial processing model and sample data, wherein the sample data includes a training set and a test set; Training the initial processing model based on the training set under the model training inference framework to obtain the item processing model; Evaluating the model performance of the item processing model based on the test set under the model training inference framework to obtain first evaluation information; Evaluating the model performance of the item processing model based on the test set under the model application reasoning framework to obtain second evaluation information; If the difference between the first evaluation information and the second evaluation information is less than a difference threshold, the item processing model is called under the model application reasoning framework to perform association prediction between the business user and each candidate item recalled for the business user, and the target item is any candidate item recalled for the business user.

10. The method according to claim 9, wherein the test samples in the test set include positive test samples and negative test samples; the positive test samples refer to test items that the test user is willing to access, and the negative test samples refer to test items that the test user refuses to access; The step of evaluating the model performance of the item processing model based on the test set under the model training reasoning framework to obtain first evaluation information includes: Calling the item processing model under the model training and reasoning framework to predict the association between the test samples in the test set and the corresponding test users to obtain predicted association labels; Counting the proportion of first samples correctly judged as positive in the positive test samples; Counting the proportion of second samples in the negative test samples that are wrongly judged as positive; determining the first evaluation information according to a ratio between the first sample ratio and the second sample ratio; The positive refers to that the predicted associated tag indicates that the test user is willing to access the corresponding test item.

11. A service processing device, comprising: The acquisition unit is used to acquire the prompt information of the target item recalled by the business user, wherein the prompt information of the target item The information includes information of the business user, characteristics of at least one associated item associated with the business user, and characteristics of the target item; A processing unit, configured to call an item processing model in a model application reasoning framework, and predict the association between the business user and the target item based on the prompt information of the target item; the item processing model is trained in a model training reasoning framework and reused in the model application reasoning framework; the processing speed of the item processing model in the model application reasoning framework is higher than the processing speed of the item processing model in the model training reasoning framework; and The processing unit is further configured to perform business processing on the target item based on the degree of association between the business user and the target item.

12. A computer device, comprising: a processor suitable for implementing a computer program; A computer-readable storage medium storing a computer program, wherein the computer-readable storage medium stores a computer program, wherein the computer program is suitable for being loaded by the processor and executing the business processing method according to any one of claims 1 to 10.

13. A computer-readable storage medium storing a computer program, wherein the computer program is suitable for being loaded by a processor and executing the service processing method according to any one of claims 1 to 10.

14. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the service processing method according to any one of claims 1 to 10 is implemented.

Images