Material rough discharge model training method, material rough discharge method, material rough discharge device and material rough discharge equipment

Abstract

The invention provides training of a material rough arrangement model, and a material rough arrangement method, device and equipment. The training method of the material rough arrangement model comprises the steps of obtaining a training sample set based on a material recommendation system; inputting the training sample into a material rough arrangement model for feature extraction and embedding processing to obtain a material embedding vector of a sample candidate material, a user embedding vector and an interest sequence embedding vector of a sample user, and a cross embedding vector between the sample user and the sample candidate material; on the basis of the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector, performing prediction processing on the interest degree of the user on the material to obtain a rough arrangement estimation score corresponding to the sample candidate material output by the material rough arrangement model; and performing parameter adjustment on the material rough arrangement model based on the difference between the rough arrangement pre-estimated score corresponding to the sample candidate material and the fine arrangement pre-estimated score obtained in the material fine arrangement model until the trained material rough arrangement model is obtained.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a training method, device, equipment and material rough arrangement method for a material rough arrangement model. Background Art

[0002] The material recommendation process includes: recall, rough sorting, fine sorting and re-sorting. The rough sorting stage is executed between the recall stage and the fine sorting stage. It is necessary to select hundreds of candidate materials that meet the post-link goals from tens of thousands of candidate material sets and send the target materials to the fine sorting stage.

[0003] The current material rough sorting model has poor expressiveness and real-time performance, and it is difficult to respond promptly to rapid changes in data distribution, resulting in a long recommendation process. Summary of the invention

[0004] The present application provides a material rough ranking model training method, device, equipment and material rough ranking method, which can improve the accuracy of material recommendation.

[0005] The technical solution of this application is implemented as follows:

[0006] The present application provides a training method for a material rough ranking model, the method comprising: obtaining a training sample set based on a material recommendation system; each training sample in the training sample set includes material information of a sample candidate material, user information and historical behavior information of a sample user; using the training sample set to iteratively train the initial material rough ranking model, inputting the training sample into the material rough ranking model for feature extraction and embedding processing, obtaining a material embedding vector of the sample candidate material, a user embedding vector and an interest sequence embedding vector of the sample user, and a cross embedding vector between the sample user and the sample candidate material; and, based on the material embedding vector, the user embedding vector, the interest sequence embedding vector, and the cross embedding vector, performing prediction processing on the user's interest in the material, obtaining a rough ranking estimated score corresponding to the sample candidate material output by the material rough ranking model; inputting the training sample into a material fine ranking model used by the material recommendation system, obtaining a fine ranking estimated score corresponding to the sample candidate material output by the material fine ranking model; adjusting parameters of the material rough ranking model based on the difference between the rough ranking estimated score corresponding to the sample candidate material and the fine ranking estimated score, until the set convergence condition is reached, thereby obtaining a trained material rough ranking model.

[0007] In some possible implementations, the material rough ranking model includes a preprocessing layer, a four-tower layer and a first neural network, wherein: the preprocessing layer is used to perform feature extraction processing on the material information of the input sample candidate materials, the user information and historical behavior information of the sample users, and obtain the material characteristics of the sample candidate materials, the user characteristics and interest sequences of the sample users, and the cross-statistical characteristics between the sample users and the sample candidate materials; the interest sequence characteristics are used to characterize the interest preferences of the sample users, and the cross-statistical characteristics are used to characterize the relevant behavioral statistical information of the sample users on the sample candidate materials; the four-tower layer is used to embed the material characteristics, user characteristics, interest sequence characteristics and cross-statistical characteristics respectively, and obtain the corresponding material embedding vector, user embedding vector, interest sequence embedding vector and cross-embedding vector; the first neural network is used to predict the user's interest in the material based on the material embedding vector, user embedding vector, interest sequence embedding vector and cross-embedding vector, and obtain the rough ranking estimated score corresponding to the sample candidate material.

[0008] In some possible implementations, the four-tower layer includes a material tower, a user tower, an interest tower and a cross tower, wherein: the material tower is used to embed the input material features to obtain the corresponding material embedding vector; the user tower is used to embed the input user features to obtain the corresponding user embedding vector; the interest tower is used to embed the input interest sequence features to obtain the corresponding interest sequence embedding vector; the cross tower is used to embed the input cross-statistical features to obtain the corresponding cross-embedding vector.

[0009] In some possible implementations, the input interest sequence features are embedded to obtain a corresponding interest sequence embedding vector, including: vectorizing the interest sequence features to obtain an interest sequence vector; matrixing the interest sequence vector to obtain a weight vector; the elements in the weight vector correspond one-to-one to the elements in the interest sequence vector; performing vector inner product processing on the interest sequence vector and the weight vector to obtain a vector inner product result; and pooling the vector inner product result to obtain an interest sequence embedding vector.

[0010] In some possible implementations, based on the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector, a prediction process of the user's interest level in the material is performed to obtain a rough-ranked estimated score corresponding to the sample material, including: fusing the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector to obtain a fused vector; predicting the user's interest level in the material on the fused vector to obtain a rough-ranked estimated score corresponding to each sample candidate material.

[0011] In some possible implementations, after obtaining the material embedding vector, the user embedding vector, and the interest sequence embedding vector, the method further includes: saving the material embedding vector, the user embedding vector, and the interest sequence embedding vector to a database of a material recommendation system.

[0012] In some possible implementations, based on the difference between the rough ranking estimated score and the fine ranking estimated score corresponding to the sample candidate material, the parameters of the material rough ranking model are adjusted until the set convergence condition is reached to obtain a trained material rough ranking model, including: obtaining the labeling information of each sample candidate material; the labeling information is used to indicate whether the sample user clicks on the sample candidate material; based on the labeling information, the fine ranking estimated score, the rough ranking estimated score and a preset loss function corresponding to each sample candidate material, the loss value of the material rough ranking model is calculated; the loss function is used to characterize the difference between the rough ranking estimated score and the labeling information, the difference between the fine ranking estimated score and the labeling information, and the difference between the rough ranking estimated score and the fine ranking estimated score; based on the loss value of the material rough ranking model, the parameters of the material rough ranking model are adjusted until the set convergence condition is reached to obtain a trained material rough ranking model.

[0013] The present application provides a material rough ranking method, which includes: in response to a material recommendation request of a target user, obtaining user information and historical behavior information of the target user, as well as material information of multiple candidate materials; based on the user information and historical behavior information of the target user, as well as the material information of multiple candidate materials, obtaining a user embedding vector and an interest sequence embedding vector of the target user, as well as a material embedding vector corresponding to each candidate material in a database of a material recommendation system; inputting the user information and historical behavior information of the target user, as well as the material information of each candidate material, into a trained material rough ranking model for feature extraction and embedding processing, to obtain a cross-embedding vector between the target user and each candidate material; and, based on the user embedding vector and interest sequence embedding vector of the target user, as well as the material embedding vector and cross-embedding vector corresponding to each candidate material, performing prediction processing on the user's interest level in the material, to obtain a rough ranking estimated score corresponding to each candidate material output by the material rough ranking model; the material rough ranking model is trained using the above-mentioned material rough ranking model training method; and based on the rough ranking estimated score corresponding to each candidate material, multiple candidate materials are sorted.

[0014] In some possible implementations, the user information and historical behavior information of the target user and the material information of each candidate material are input into the trained material rough-sorting model for feature extraction and vectorization processing to obtain a cross-embedding vector between the target user and each candidate material, which may include: preprocessing the user information and historical behavior information of the target user through a preprocessing layer of the material rough-sorting model to obtain cross-statistical features; vectorizing the cross-statistical features through a cross-tower of the material rough-sorting model to obtain a cross-embedding vector.

[0015] In some possible implementations, based on the user embedding vector and interest sequence embedding vector of the target user, the material embedding vector and cross embedding vector corresponding to each candidate material, a prediction process is performed on the user's interest level in the material to obtain a rough ranking estimated score corresponding to each candidate material output by the material rough ranking model, including: fusing the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector to obtain a fused vector; and predicting the user's interest level in the material on the fused vector to obtain a rough ranking estimated score corresponding to each sample candidate material.

[0016] The present application provides a training device for a material rough ranking model, the device comprising: a sample acquisition module, used to obtain a training sample set based on a material recommendation system; each training sample in the training sample set comprises material information of a sample candidate material, user information and historical behavior information of a sample user; a training module, used to iteratively train an initial material rough ranking model using the training sample set, input the training sample into the material rough ranking model for feature extraction and embedding processing, and obtain a material embedding vector of the sample candidate material, a user embedding vector and an interest sequence embedding vector of the sample user, and a cross embedding vector between the sample user and the sample candidate material; and, Based on the material embedding vector, user embedding vector, interest sequence embedding vector, and cross embedding vector, the user's interest level in the material is predicted, and the rough ranking estimated score corresponding to the sample candidate material output by the material rough ranking model is obtained; the knowledge distillation module is used to input the training samples into the material fine ranking model used by the material recommendation system, and obtain the fine ranking estimated score corresponding to the sample candidate material output by the material fine ranking model; the parameter adjustment module is used to adjust the parameters of the material rough ranking model based on the difference between the rough ranking estimated score and the fine ranking estimated score corresponding to the sample candidate material, until the set convergence condition is reached, and a trained material rough ranking model is obtained.

[0017] In some possible implementations, the material rough ranking model includes a preprocessing layer, a four-tower layer and a first neural network, wherein: the preprocessing layer is used to perform feature extraction processing on the material information of the input sample candidate materials, the user information and historical behavior information of the sample users, and obtain the material characteristics of the sample candidate materials, the user characteristics and interest sequences of the sample users, and the cross-statistical characteristics between the sample users and the sample candidate materials; the interest sequence characteristics are used to characterize the interest preferences of the sample users, and the cross-statistical characteristics are used to characterize the relevant behavioral statistical information of the sample users on the sample candidate materials; the four-tower layer is used to embed the material characteristics, user characteristics, interest sequence characteristics and cross-statistical characteristics respectively, and obtain the corresponding material embedding vector, user embedding vector, interest sequence embedding vector and cross-embedding vector; the first neural network is used to predict the user's interest in the material based on the material embedding vector, user embedding vector, interest sequence embedding vector and cross-embedding vector, and obtain the rough ranking estimated score corresponding to the sample candidate material.

[0018] In some possible implementations, the four-tower layer includes a material tower, a user tower, an interest tower and a cross tower, wherein: the material tower is used to embed the input material features to obtain the corresponding material embedding vector; the user tower is used to embed the input user features to obtain the corresponding user embedding vector; the interest tower is used to embed the input interest sequence features to obtain the corresponding interest sequence embedding vector; the cross tower is used to embed the input cross-statistical features to obtain the corresponding cross-embedding vector.

[0019] In some possible implementations, the interest tower is used to vectorize the interest sequence features to obtain the interest sequence vector; matrix the interest sequence vector to obtain the weight vector; the elements in the weight vector correspond one-to-one to the elements in the interest sequence vector; perform vector inner product processing on the interest sequence vector and the weight vector to obtain the vector inner product result; and perform pooling processing on the vector inner product result to obtain the interest sequence embedding vector.

[0020] In some possible implementations, the training module is used to fuse the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector to obtain a fused vector; and to predict the user's interest level in the material on the fused vector to obtain a rough estimated score corresponding to each sample candidate material.

[0021] In some possible implementations, the above-mentioned device also includes a data storage module, which is used to save the material embedding vector, the user embedding vector and the interest sequence embedding vector to a database of the material recommendation system.

[0022] In some possible implementations, a parameter adjustment module is used to obtain the labeling information of each sample candidate material; the labeling information is used to indicate whether the sample user clicks on the sample candidate material; based on the labeling information, the fine ranking estimated score, the rough ranking estimated score and a preset loss function corresponding to each sample candidate material, the loss value of the material rough ranking model is calculated; the loss function is used to characterize the difference between the rough ranking estimated score and the labeling information, the difference between the fine ranking estimated score and the labeling information, and the difference between the rough ranking estimated score and the fine ranking estimated score; based on the loss value of the material rough ranking model, the parameters of the material rough ranking model are adjusted until the set convergence condition is reached to obtain a trained material rough ranking model.

[0023] The present application provides a rough material ranking device, which includes: a data acquisition module, which is used to respond to a material recommendation request of a target user and obtain user information and historical behavior information of the target user, as well as material information of multiple candidate materials; a vector acquisition module, which is used to obtain a user embedding vector and an interest sequence embedding vector of the target user, as well as a material embedding vector corresponding to each candidate material in a database of a material recommendation system based on the user information and historical behavior information of the target user, as well as the material information of multiple candidate materials; a data processing module, which is used to input the user information and historical behavior information of the target user, as well as the material information of each candidate material, into a trained rough material ranking model for feature extraction and embedding processing, so as to obtain a cross embedding vector between the target user and each candidate material; and, based on the user embedding vector and interest sequence embedding vector of the target user, the material embedding vector and the cross embedding vector corresponding to each candidate material, predict the user's interest in the material, and obtain a rough ranking estimated score corresponding to each candidate material output by the rough material ranking model; a sorting module, which is used to train the rough material ranking model using the above-mentioned rough material ranking model training method; and sort multiple candidate materials based on the rough ranking estimated score corresponding to each candidate material.

[0024] In some possible implementations, the data processing module is used to preprocess the user information and historical behavior information of the target user through the preprocessing layer of the material rough sorting model to obtain cross-statistical features; and vectorize the cross-statistical features through the cross tower of the material rough sorting model to obtain a cross-embedded vector.

[0025] In some possible implementations, the data processing module is used to fuse the user embedding vector and the interest sequence embedding vector, the material embedding vector and the cross embedding vector corresponding to each candidate material through the fusion layer of the material rough ranking model to obtain a fusion vector; and predict the user's interest level in the material through the first neural network of the material rough ranking model to obtain a rough ranking estimated score corresponding to each candidate material.

[0026] The present application provides an electronic device, comprising: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided by the present application.

[0027] The present application provides a computer storage medium storing executable instructions for implementing the method provided by the present application when the executable instructions are executed by a processor.

[0028] The present application provides a computer program product, including a computer program or instructions, for implementing the method provided by the present application when the computer program or instructions are executed by a processor.

[0029] This application has the following beneficial effects:

[0030] In this application, a material embedding vector, a user embedding vector, an interest sequence embedding vector, and a cross-embedding vector are obtained based on training samples. The material rough ranking model outputs the rough ranking estimated scores corresponding to the sample candidate materials based on the above four embedding vectors. Based on the difference between the rough ranking estimated scores output by the material rough ranking model and the fine ranking estimated scores output by the material fine ranking model, the parameters of the material rough ranking model are adjusted until the set convergence conditions are reached to obtain a trained material rough ranking model. The expression ability of the material rough ranking model can be improved by introducing cross-embedding vectors. By introducing the material fine ranking model to guide the learning of the material rough ranking module, the generalization ability of the material rough ranking model and the scoring consistency with the material fine ranking model can be improved, and the prediction ability of the material rough ranking model can be improved, thereby improving the accuracy of material recommendation.

[0031] Furthermore, in the material rough ranking method, the user embedding vector, interest sequence embedding vector and material embedding vector are calculated offline. When a recommendation request is received from a target user, the user embedding vector, interest sequence embedding vector and material embedding vector can be directly called in the database, and a cross-embedding vector is generated in response to the recommendation request; the material rough ranking model realizes decoupled modeling in four dimensions of user embedding vector, interest sequence embedding vector, material embedding vector and cross-embedding vector, and then in the material recommendation process, while ensuring the recommendation accuracy, the recommendation time is greatly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the architecture of the recommendation system provided in the embodiment of the present application;

[0033] Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0034] Figure 3 It is an optional flow chart of the training method of the material rough sorting model provided in the embodiment of the present application;

[0035] Figure 4 It is a structural schematic diagram of a material rough arrangement model and a material fine arrangement model provided in an embodiment of the present application;

[0036] Figure 5 It is an optional flow chart of the material rough sorting method provided in the embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0038] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0039] If similar descriptions of "first / second" appear in the application documents, the following explanation is added. In the following description, the terms "first / second / third" involved are merely to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first / second / third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0041] Cascade sorting architecture has been widely used in material recommendation scenarios that require large-scale sorting, such as search, recommendation, and advertising. Taking the material recommendation system as an example, it includes modules such as recall, rough sorting, fine sorting, and re-sorting. Rough sorting is between recall and fine sorting. Generally, it is necessary to select hundreds of target materials that meet the post-link goals from a set of tens of thousands of candidate materials and send them to the subsequent fine sorting module. Rough sorting has very strict time requirements. Generally, scoring needs to be completed within 10 to 20 milliseconds (ms), which requires balancing computing power, real time (RT), and the final scoring effect.

[0042] Before the birth of deep learning, coarse-grained ranking was mainly divided into the following two stages: 1) The first generation of coarse-grained ranking models used static quality scores, that is, statistically analyzing the historical average click-through rate (CTR) of materials. This process only used information on the material side, which resulted in limited expression of the algorithm. The advantage was that it could be updated very quickly; 2) The second generation of coarse-grained ranking was an early machine learning model represented by logistic regression (LR). The model structure was relatively simple, with certain personalized expression capabilities, and could be updated and served online to meet some application requirements. With the advent of the information age, traditional coarse-grained ranking models can no longer adapt to the increase in the number of users, the changes in user consumption behavior, and the increase in the number of materials in actual applications.

[0043] The most widely used rough ranking model currently has two optimization routes: 1) One route is a dual-tower structure represented by DSSM (deep structured semantic models), where user features and material features are input on both sides respectively. After deep network calculation, user vectors and material vectors are output respectively. After obtaining their respective embedding vectors, the ranking scores are calculated through inner product and other operations. The advantage of the dual-tower structure is that it can ensure fast online service when facing massive candidate sets; 2) The other route is a lightweight degenerate version of fine ranking, which achieves the interactive model effect on rough ranking through feature selection, knowledge distillation, network structure compression and other methods, such as FSCD (interactive features and structural representation features), COLD (computing power cost-aware online and lightweight deep pre-ranking system), AutoFAS (automatic feature and architecture selection for pre-ranking system), etc.

[0044] The main idea of ​​knowledge distillation is to guide the training of the student model through the teacher model, and to "distill" the feature representation learned by the complex and strong learning ability teacher model, and pass it to the student model with small parameters and weak learning ability, so as to obtain a fast and expressive student model.

[0045] In recent years, many effective coarse-ranking system architectures and methods have been proposed. Although these methods differ in network structure and strategy, they are basically improved around the above two routes. By adding the above two methods, the model's ability to sort materials can be effectively improved in the coarse-ranking stage of the recommendation system. However, for such a framework, there are mainly the following problems in practical applications: 1) The model's expression ability is still limited. Although the vector inner product greatly improves the calculation speed and saves computing power, it also causes the model to be unable to use cross-features, and its ability is greatly limited; 2) There is a cold start problem, which is not friendly to newly released materials and new users; 3) The version synchronization of user vectors and material vectors affects the iteration efficiency; 4) The real-time performance of the model is poor, because user vectors and material vectors generally need to be calculated in advance, and this advance calculation time will slow down the update speed of the entire system, making it difficult for the system to respond to rapid changes in data distribution in a timely manner. In order to solve the above problems, in the actual application of the coarse-grained ranking model, each time a new version of the model is iterated, the corresponding material vector and user vector must be output respectively. This makes the iteration process very long and the iteration efficiency is low. At the same time, a lot of resources are needed to update the model to ensure the effectiveness of the material vector and user vector, which also increases the time consumption of the recommendation process.

[0046] In response to the above problems, an embodiment of the present application provides a method for training a rough material ranking model to improve the accuracy of material recommendations.

[0047] The following describes an exemplary application of the electronic device provided by the embodiment of the present application. The electronic device provided by the embodiment of the present application can be implemented as various types of user terminals such as a laptop computer, a tablet computer, a desktop computer, a mobile device (for example, a mobile phone, a wearable smart watch, a dedicated messaging device), and can also be implemented as a server. The following describes an exemplary application when the electronic device is implemented as a server.

[0048] See also Figure 1 , Figure 1 It is a schematic diagram of the architecture of the recommendation system provided in an embodiment of the present application. Electronic devices (electronic device 40-1 and electronic device 40-2 are shown as examples) are connected to server 200 via network 300. Network 300 can be a wide area network or a local area network, or a combination of the two.

[0049] In some possible implementations, user A may publish material information A through terminal 40-1, and user B may publish material information B through terminal 40-2. Material information A and material information B are uploaded to server 200 through network 300. Server 200 may store material information A and material information B in database 500. In order to accurately push materials according to the interests of users, server 200 may push materials that have the highest degree of matching with the interests of users to users after sorting, and the materials pushed to users may be displayed on the graphical interface of electronic device 400 (graphical interface 41-1 and graphical interface 41-2 are shown as examples).

[0050] In some embodiments, the server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and big data and artificial intelligence platforms. The terminal may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart watch, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present invention.

[0051] See also Figure 2 , Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, Figure 2 The electronic device 400 shown may be the above-mentioned terminal and / or server 200. The electronic device 400 includes: at least one processor 410, a memory 450, at least one network interface 420 and a user interface 430. The various components in the electronic device 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 440 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0052] Processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0053] The user interface 430 includes one or more output devices 431 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0054] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0055] The memory 450 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0056] In some embodiments, memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0057] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0058] A network communication module 452, for reaching other computing devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 include: Bluetooth, wireless compatibility certification (WiFi), and universal serial bus (USB), etc.;

[0059] a presentation module 453 for enabling presentation of information via one or more output devices 431 (e.g., display screen, speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripherals and displaying content and information);

[0060] The input processing module 454 is used to detect one or more user inputs or interactions from one of the one or more input devices 432 and translate the detected inputs or interactions.

[0061] In some embodiments, the training device for the material rough arrangement model provided in the embodiments of the present application can be implemented in a software manner. Figure 2 A training device 455 for a material rough-sorting model stored in a memory 450 is shown, which may be software in the form of a program and a plug-in, etc. The training device 455 for a material rough-sorting model includes the following software modules: a sample collection module 4551, a training module 4552, a knowledge distillation module 4553 and a parameter adjustment module 4554.

[0062] In some embodiments, the material rough discharge device provided in the embodiments of the present application can be implemented in a software manner. Figure 2 A material rough sorting device 456 stored in a memory 450 is shown, which may be software in the form of a program and a plug-in, etc. The material rough sorting device 456 includes the following software modules: a data acquisition module 4561, a vector acquisition module 4562, a data processing module 4563 and a sorting module 4564.

[0063] These modules are logical, so they can be arbitrarily combined or further divided according to the functions to be implemented. The functions of each module will be explained below.

[0064] In other embodiments, the material rough sorting device provided in the embodiments of the present application can be implemented in hardware. As an example, the device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the material rough sorting method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application specific integrated circuits (application specific integrated circuit, ASIC,), DSP, programmable logic device (programmable logic device, PLD,), complex programmable logic device (complex programmable logic device, CPLD), field programmable gate array (field-programmable gate array, FPGA) or other electronic components.

[0065] In some embodiments, the recommendation system provided in the embodiments of the present application may be a material recommendation system. The training method of the material rough sorting model provided in the embodiments of the present application is described below using the material recommendation system as an example.

[0066] The following will describe the training method of the material rough sorting model provided in the embodiment of the present application in combination with the exemplary application and implementation of the electronic device provided in the embodiment of the present application.

[0067] In the embodiments of the present application, for ease of explanation, the training device of the material rough arrangement model is referred to as the training device.

[0068] See also Figure 3 , Figure 3 This is an optional flow chart of the training method of the material rough sorting model provided in the embodiment of the present application. Figure 3 The steps shown illustrate the method for training a rough material sorting model.

[0069] S301, obtaining a training sample set based on a material recommendation system, wherein each training sample in the training sample set includes material information of a sample candidate material, user information and historical behavior information of a sample user.

[0070] It can be understood that the training device generates a training sample set based on the information provided by the material recommendation system. The training sample set may include multiple training samples. Each training sample includes material information of the sample candidate material, user information and historical behavior information of the sample user. User information may include information such as the user's gender, age, occupation, preferred areas of posting, preferred areas of interactive materials, and activity. Historical behavior information may include historical materials that the user has clicked on. Material information may include information such as the length, content, posting time, and unique identifier (identity document, ID) of the material.

[0071] S302, use the training sample set to iteratively train the initial material rough-sorting model, input the training samples into the material rough-sorting model for feature extraction and embedding processing, obtain the material embedding vector of the sample candidate material, the user embedding vector and interest sequence embedding vector of the sample user, and the cross-embedding vector between the sample user and the sample candidate material; and, based on the material embedding vector, the user embedding vector, the interest sequence embedding vector, and the cross-embedding vector, perform prediction processing on the user's interest level in the material, and obtain the rough-sorting estimated score corresponding to the sample candidate material output by the material rough-sorting model.

[0072] It can be understood that the training device inputs the training sample set into the initial material rough ranking model, and the material rough ranking model performs prediction processing based on the training samples in the training sample set, and outputs the rough ranking estimated scores of the sample candidate materials.

[0073] In some possible implementations, the material rough ranking model includes a preprocessing layer, a four-tower layer and a first neural network, wherein: the preprocessing layer is used to perform feature extraction processing on the material information of the input sample candidate materials, the user information and historical behavior information of the sample users, and obtain the material characteristics of the sample candidate materials, the user characteristics and interest sequence characteristics of the sample users, and the cross-statistical characteristics between the sample users and the sample candidate materials; the interest sequence characteristics are used to characterize the interest preferences of the sample users, and the cross-statistical characteristics are used to characterize the relevant behavioral statistical information of the sample users on the sample candidate materials; the four-tower layer is used to embed the material characteristics, user characteristics, interest sequence characteristics, and cross-statistical characteristics respectively, and obtain the corresponding material embedding vector, user embedding vector, interest sequence embedding vector, and cross-embedding vector; the first neural network is used to predict the user's interest in the material based on the material embedding vector, user embedding vector, interest sequence embedding vector and cross-embedding vector, and obtain the rough ranking estimated score corresponding to the sample candidate material.

[0074] In some embodiments, user features are used to characterize a user's personal portrait, and material features are used to characterize material information of a material.

[0075] In some embodiments, the preprocessing layer in the material rough sorting model can perform feature extraction processing on the user information and historical behavior information of the input sample users to obtain user features and interest sequence features; perform feature extraction processing on the material information of the sample candidate materials to obtain material features; perform feature combination processing on the material features and user features to capture the interactive information and nonlinear relationship between the user features and the material features to obtain cross-statistical features.

[0076] In some embodiments, the preprocessing layer can extract the material ID in the historical behavior information to obtain the ID sequence of multiple historical materials clicked by the user. After performing feature extraction processing on the ID sequence of multiple historical materials, the interest sequence feature can be obtained.

[0077] In some embodiments, the interest sequence feature can be used to represent the materials that the user has clicked and interacted with in the past. For example, interest sequence feature A can represent that user A has clicked on materials A, B, C, and D, and liked material B and forwarded material C in the past seven days (i.e., the preset first time period). The preset first time period can be set according to actual conditions, and the embodiments of the present application do not specifically limit this.

[0078] In some embodiments, the feature combination processing methods for material features and user features may include but are not limited to any of the following: feature multiplication, feature division, hash cross processing, feature cross processing through one-hot encoding, etc.

[0079] In some embodiments, the cross-statistical features may be used to characterize the user's interaction with the material, as well as the statistical values ​​of the user's click rate and interaction rate with the material.

[0080] In one example, the cross-statistical feature A may indicate that the material recommendation system recommends material A, material B, material C, and material D to user A. After obtaining the recommended materials, user A clicks on material A, material B, and material C, and forwards material C. The cross-statistical feature A may also indicate that user A's click rate on the recommended materials is 75%, and the interaction rate on the recommended materials is 25%.

[0081] In some embodiments, the user's interactive behaviors on the material include but are not limited to: forwarding, liking, collecting, viewing time, etc.

[0082] In some embodiments, the cross-statistical feature is used to represent the user's interaction with the material, and the statistical feature is used to represent the statistical value of the user's click rate and interaction rate with the material.

[0083] In some embodiments, the embedding processing performed by the four-tower layer on material features, user features, interest sequence features, and cross-statistical features is embedding processing; the obtained material embedding vector, user embedding vector, interest sequence embedding vector, and cross embedding vector are embedding vectors.

[0084] In some possible implementations, the four-tower layer includes a material tower, a user tower, an interest tower and a cross tower, wherein: the material tower is used to embed the input material features to obtain the corresponding material embedding vector; the user tower is used to embed the input user features to obtain the corresponding user embedding vector; the interest tower is used to embed the input interest sequence features to obtain the corresponding interest sequence embedding vector; the cross tower is used to embed the input cross-statistical features to obtain the corresponding cross-embedding vector.

[0085] It can be understood that the four-tower layer includes four independent processing modules, namely the material tower, user tower, interest tower and cross tower. The material tower is used to process material features, the user tower is used to process user features, the interest tower is used to process interest sequence features, and the cross tower is used to process cross statistical features.

[0086] In the embodiment of the present application, a cross tower is introduced into the rough material arrangement model to solve the information loss caused by the lack of information cross-talk between the user side and the material side.

[0087] In some possible implementations, the input interest sequence features are embedded to obtain a corresponding interest sequence embedding vector, including: vectorizing the interest sequence features to obtain an interest sequence vector; matrixing the interest sequence vector to obtain a weight vector; the elements in the weight vector correspond one-to-one to the elements in the interest sequence vector; performing vector inner product processing on the interest sequence vector and the weight vector to obtain a vector inner product result; and pooling the vector inner product result to obtain an interest sequence embedding vector.

[0088] It can be understood that the interest tower embeds the input interest sequence features to obtain an interest sequence embedding vector. Each element in the interest sequence embedding vector represents a historical behavior of the user (such as clicks and interactions, etc.). The interest sequence embedding vector is a multidimensional vector. By matrixing each element in the interest sequence embedding vector, the weight corresponding to each element in the interest sequence embedding vector can be generated. All weights are combined together to form a weight vector.

[0089] In some embodiments, taking the material ID sequence {V1, V2, V3, ..., Vn} in the historical behavior information as an example, the training device extracts features from the material ID sequence to obtain the interest sequence features, and vectorizes the interest sequence features to obtain the interest sequence vector [E1 E2 E3 ... En]; the weight vector is The expression (1) for performing vector inner product processing on the interest sequence vector and the weight vector and performing pooling processing on the obtained vector inner product result is as follows:

[0090]

[0091] In the formula, E interest represents the embedding vector of the interest sequence, and mean{} represents the pooling process.

[0092] Each element in the interest sequence vector corresponds to an element in the weight vector, such as E1 corresponds to W1, E2 corresponds to W2, E3 corresponds to W2, and En corresponds to Wn. The larger the weight, the higher the user's interest in the material, which means that the user is more likely to click on the material.

[0093] In the embodiment of the present application, the interest sequence feature can enhance the distinction between different samples, that is, the type of material that the user is interested in can be more accurately inferred based on the interest sequence features of different users. Introducing the interest tower into the material rough sorting model to further process the interest sequence feature can enable the material rough sorting model to better understand the user, capture the dynamic preference of the user's interest, and increase the interaction effect.

[0094] In some possible implementations, based on the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector, the user's interest level in the material is predicted to obtain a rough-ranked estimated score corresponding to the sample material, which may include: fusing the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector to obtain a fused vector; predicting the user's interest level in the material on the fused vector to obtain a rough-ranked estimated score corresponding to each sample candidate material.

[0095] It can be understood that the first neural network in the rough material ranking model can perform fusion processing (such as concat processing) and cross processing on the material embedding vector, user embedding vector, interest sequence embedding vector and cross embedding vector to obtain a fusion vector; then, the fusion vector is processed to predict the user's interest in the material to estimate the user's interest in each sample candidate material, and obtain the rough ranking estimated score corresponding to each sample candidate material. The higher the rough ranking estimated score, the higher the user's interest in the material corresponding to the rough ranking estimated score, and the higher the probability of click and interaction.

[0096] In some embodiments, the first neural network may be an enhanced two-stream multi-layer perceptron model (final multi-layer perceptron, Final MLP), or other types of neural networks may be selected according to actual needs, which is not specifically limited in the embodiments of the present application.

[0097] In some possible implementations, after obtaining the material embedding vector, the user embedding vector, and the interest sequence embedding vector, the above method may further include: saving the material embedding vector, the user embedding vector, and the interest sequence embedding vector to a database of the material recommendation system.

[0098] It can be understood that the material embedding vector, user embedding vector and interest sequence embedding vector are saved in the database. When the material rough sorting model is put into online use (i.e., the material rough sorting method provided in the embodiment of the present application), when the user sends a recommendation request, the above three embedding vectors can be directly obtained in the database.

[0099] S303, input the training samples into the material ranking model used by the material recommendation system, and obtain the estimated scores of the sample candidate materials output by the material ranking model.

[0100] It can be understood that in order to enhance the consistency between rough ranking and fine ranking, a knowledge distillation architecture can be introduced to train the material rough ranking model. The specific training method includes: using the fine ranking estimated score output by the material fine ranking model to calculate the loss value of the material rough ranking model.

[0101] S304, based on the difference between the rough ranking estimated score and the fine ranking estimated score corresponding to the sample candidate materials, adjust the parameters of the material rough ranking model until the set convergence condition is reached, and obtain the trained material rough ranking model.

[0102] It can be understood that based on the idea of ​​knowledge distillation, the difference between the more accurate fine ranking estimated score and the rough ranking estimated score is calculated through a preset function expression, and the parameters of the material rough ranking model are adjusted based on the difference between the fine ranking estimated score and the rough ranking estimated score.

[0103] In some possible implementations, the above S304 may include: obtaining the marking information of each sample candidate material; the marking information is used to indicate whether the sample user clicks on the sample candidate material; based on the marking information corresponding to each sample candidate material, the fine ranking estimated score, the rough ranking estimated score and a preset loss function, the loss value of the material rough ranking model is calculated; the loss function is used to characterize the difference between the rough ranking estimated score and the marking information, the difference between the fine ranking estimated score and the marking information, and the difference between the rough ranking estimated score and the fine ranking estimated score; based on the loss value of the material rough ranking model, the parameters of the material rough ranking model are adjusted until the set convergence condition is reached to obtain a trained material rough ranking model.

[0104] It can be understood that the tag information of the sample candidate material can represent the actual behavior of the sample user towards the sample candidate material. For example, the database of the material recommendation system can record the interactive behavior of all exposed materials, that is, whether the user clicks on the material; and set tag information for the exposed materials based on the click information. If the user clicks on the material, the tag value corresponding to the tag information of the material is assigned to 1; if the user does not click on the material, the tag value corresponding to the tag information of the material is assigned to 0.

[0105] In some embodiments, the expression (2) for calculating the loss value of the material rough arrangement model is as follows:

[0106]

[0107] Where, L is the loss value of the material rough arrangement model, F s (X s ; W s ) represents the rough estimation score output by the rough material sorting model, X s Represents the input data of the material rough-cut model, W s It represents the operation process of the material rough arrangement model on the input data, F t (X t ,X*;W t) represents the estimated score of the material fine sorting model output, Xs and X* represent the input data of the material fine sorting model, where X* represents the data input into the material fine sorting model but not into the material rough sorting model; W t Represents the calculation process of the material sorting model on the input data, y s Indicates the true label value of the sample candidate material of the input material rough sorting model, y t L represents the true label value of the sample candidate material input into the material ranking model, and λ represents the preset one-dimensional hyperparameter. s , L t and L d is the preset loss function (loss), where L s is the loss function associated with the rough material sorting model, L t is the loss function associated with the material sorting model, L d It is the knowledge distillation loss function, which is used to minimize the gap between the coarse ranking and the fine ranking scores.

[0108] It should be noted that L s represents the loss of the coarse-grained model, L t Represents the loss of the refined ranking model, L s represents the knowledge distillation loss. s The value of y is 0 or 1, 0 means that the sample user did not click on the sample candidate material; 1 means that the sample user clicked on the sample candidate material; similarly, y t The value of is 0 or 1, where 0 means that the sample user did not click on the sample candidate material; 1 means that the sample user clicked on the sample candidate material.

[0109] In some embodiments, the value of λ will gradually increase with the iteration of the training steps of the material rough sorting model to enhance L s As a result, during the iterative training process, the rough estimation scores output by the rough material sorting model gradually align with the fine estimation scores output by the fine material sorting model.

[0110] In some embodiments, the estimated scores of the refined sorting model of the materials can be stored in the log of the refined sorting model of the materials. The rough sorting model of the materials can directly obtain the estimated scores of the refined sorting from the log.

[0111] In the embodiment of the present application, by using knowledge distillation to train and optimize the rough material ranking model, the optimization objectives of the rough ranking and fine ranking stages can be kept consistent with the final optimization goal of the recommendation task, further improving the full-link consistency of the material recommendation system, and further increasing the adaptability of the rough material ranking model to the environment and its ability to capture changes in user interests.

[0112] An embodiment of the present application provides a material rough ranking method, which includes: in response to a material recommendation request of a target user, obtaining user information and historical behavior information of the target user, and material information of multiple candidate materials; based on the user information and historical behavior information of the target user, and the material information of multiple candidate materials, obtaining a user embedding vector and an interest sequence embedding vector of the target user, and a material embedding vector corresponding to each candidate material in a database of a material recommendation system; inputting the user information and historical behavior information of the target user, and the material information of each candidate material into a trained material rough ranking model for feature extraction and embedding processing, and obtaining a cross-embedding vector between the target user and each candidate material; and, based on the user embedding vector and interest sequence embedding vector of the target user, and the material embedding vector and cross-embedding vector corresponding to each candidate material, performing prediction processing on the user's interest in the material, and obtaining a rough ranking estimated score corresponding to each candidate material output by the material rough ranking model; the material rough ranking model is trained using the above-mentioned material rough ranking model training method; and multiple candidate materials are sorted based on the rough ranking estimated score corresponding to each candidate material.

[0113] It can be understood that when the trained material rough ranking model is put into use online, it receives material recommendation requests from target users; in response to the material recommendation requests, the rough ranking estimated score of each candidate material is obtained through the material rough ranking model; then the candidate materials are sorted, and based on the sorting results, the materials that enter the fine ranking stage (i.e., the target materials) are screened out from multiple candidate materials.

[0114] In some embodiments, the rough sorting device obtains the rough sorting estimated scores of material A, material B, material C, material D and material E, and sorts them from high to low. The sorting result is: material D, material C, material A, material E, material B. The preset screening condition is: the top 40% of the materials. Then, based on the sorting result and the screening condition, material D and material C are the target materials and can enter the next stage of fine sorting.

[0115] In some embodiments, during the online use of the rough-ranking model of materials, it is not necessary to calculate the user embedding vector, the interest sequence embedding vector and the material embedding vector every time the rough-ranking estimated score is calculated. The user embedding vector, the interest sequence embedding vector and the material embedding vector are calculated offline and stored in the database of the material recommendation system. The rough-ranking device of materials can directly obtain the user embedding vector, the interest sequence embedding vector and the material embedding vector in the database in response to the user's material recommendation request.

[0116] In some embodiments, during the online use of the rough material ranking model, the user embedding vector, interest sequence embedding vector and material embedding vector are calculated offline and updated in real time. After each update, the updated user embedding vector, interest sequence embedding vector and material embedding vector are used to replace the embedding vector before the update. For example, the material embedding feature is updated every time the material is exposed; the user sequence embedding feature is updated every time the user clicks.

[0117] In some possible implementations, the above-mentioned inputting the user information and historical behavior information of the target user and the material information of each candidate material into the trained material rough-sorting model for feature extraction and vectorization processing to obtain the cross-embedding vector between the target user and each candidate material may include: preprocessing the user information and historical behavior information of the target user through the preprocessing layer of the material rough-sorting model to obtain cross-statistical features; vectorizing the cross-statistical features through the cross-tower of the material rough-sorting model to obtain a cross-embedding vector.

[0118] In some embodiments, during the online use of the material rough sorting model, the cross-statistical features are generated only when the user issues a material recommendation request. The material rough sorting model responds to the user's recommendation request, generates cross-statistical features, and generates cross-embedding vectors based on the cross-statistical features. The user embedding vector, interest sequence embedding vector, and material embedding vector are directly obtained from the database.

[0119] In some possible implementations, the above-mentioned prediction processing of the user's interest in the material based on the user embedding vector and the interest sequence embedding vector of the target user, the material embedding vector and the cross-embedding vector corresponding to each candidate material, and obtaining the rough ranking estimated score corresponding to each candidate material output by the material rough ranking model may include: fusing the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross-embedding vector to obtain a fused vector; predicting the user's interest in the material on the fused vector to obtain a rough ranking estimated score corresponding to each sample candidate material.

[0120] It can be understood that in the online process, the first neural network in the material rough ranking model can perform fusion processing (such as concat processing) and cross processing on the material embedding vector, user embedding vector, interest sequence embedding vector and cross embedding vector to obtain a fusion vector; then, the fusion vector is processed to predict the user's interest in the material to estimate the user's interest in each candidate material and obtain the rough ranking estimated score corresponding to each candidate material. The higher the rough ranking estimated score, the higher the user's interest in the material corresponding to the rough ranking estimated score, and the higher the probability of click and interaction.

[0121] Figure 4 It is a structural schematic diagram of a material rough arrangement model and a material fine arrangement model provided in an embodiment of the present application.

[0122] Figure 5 This is an optional flow chart of the material rough sorting method provided in the embodiment of the present application. Figure 4 The model structure and Figure 5 The steps shown illustrate the rough material sorting method.

[0123] S501, a material rough sorting device responds to a material recommendation request from a target user and obtains user information, historical behavior information, and material information of a plurality of candidate materials from the target user.

[0124] S502, the material rough sorting device obtains a user embedding vector, an interest sequence embedding vector and a material embedding vector from a database of a material recommendation system based on user information, historical behavior information and material information.

[0125] S503, the material rough sorting device preprocesses the user information and historical behavior information through the preprocessing layer of the material rough sorting model 40 to obtain cross-statistical features.

[0126] S504, the material rough sorting device inputs the cross statistical features into the cross tower 41 to obtain a cross embedding vector.

[0127] S505, the material rough ranking device fuses the user embedding vector, the interest sequence embedding vector, the material embedding vector and the cross embedding vector through the first neural network 42 to obtain a fused vector; and performs linear transformation on the fused vector to obtain a rough ranking estimated score corresponding to each candidate material.

[0128] S506, the material rough sorting device sorts the rough sorting estimated score corresponding to each candidate material, and selects the target material to enter the fine sorting stage from multiple candidate materials based on the sorting result.

[0129] Among them, in the material ranking model 50, the input features include material features, reader features, blogger features, user interaction features, context features, etc. Reader features are used to characterize the personal portrait of the user who receives the recommended material, publisher features are used to represent the personal portrait of the user who publishes the material, user interaction features can characterize the interaction between users (such as attention, comments, etc.), and context features are used to characterize information such as the reading volume and interaction of the material.

[0130] At this point, the material rough discharge device has completed the above-mentioned material rough discharge method.

[0131] In an embodiment of the present application, a material embedding vector, a user embedding vector, an interest sequence embedding vector, and a cross-embedding vector are obtained based on the training sample. The material rough ranking model outputs the rough ranking estimated scores corresponding to the sample candidate materials based on the above four embedding vectors. Based on the difference between the rough ranking estimated scores output by the material rough ranking model and the fine ranking estimated scores output by the material fine ranking model, the parameters of the material rough ranking model are adjusted until the set convergence conditions are reached to obtain a trained material rough ranking model. The expression ability of the material rough ranking model can be improved by introducing the cross-embedding vector. By introducing the material fine ranking model to guide the learning of the material rough ranking module, the generalization ability of the material rough ranking model and the scoring consistency with the material fine ranking model can be improved, and the prediction ability of the material rough ranking model can be improved, thereby improving the accuracy of material recommendation.

[0132] Furthermore, in the material rough ranking method, the user embedding vector, interest sequence embedding vector and material embedding vector are calculated offline. When a recommendation request is received from a target user, the user embedding vector, interest sequence embedding vector and material embedding vector can be directly called in the database, and a cross-embedding vector is generated in response to the recommendation request; the material rough ranking model realizes decoupled modeling in four dimensions of user embedding vector, interest sequence embedding vector, material embedding vector and cross-embedding vector, and then in the material recommendation process, while ensuring the recommendation accuracy, the recommendation time is greatly reduced.

[0133] Furthermore, the material fine ranking model is used as the teacher model and the material rough ranking model is used as the student model. The advantages of the material fine ranking model are complex cross-features, more powerful model structure, stronger fitting ability and generalization ability, which is conducive to improving the prediction ability of the material rough ranking model. The knowledge distillation of the material rough ranking model is carried out through the fine ranking estimated score of the material fine ranking model, which is conducive to the consistency of the material rough ranking model and the material fine ranking model, and improves the expression ability of the material rough ranking model.

[0134] Furthermore, by introducing interest towers into the rough material sorting model, we can use a weighted aggregation method to capture the different interests of users based on the different contributions of each user behavior in the historical behavior information.

[0135] Furthermore, by introducing cross towers into the rough material arrangement model, the information loss caused by the late crossover of information between the user tower and the material tower can be reduced, further enhancing the model performance.

[0136] In an embodiment of the present application, when using a real-time updated rough material ranking model for online AB testing, the model is used in the rough material ranking stage, and the final optimization effect is: the user's click rate on recommended materials increased by 1.36%, the average time per person increased by 1.45%, and the interaction rate increased by 1.8%.

[0137] The following continues to describe the exemplary structure of the training device 455 of the material rough arrangement model provided in the embodiment of the present application implemented as a software module. In some embodiments, such as Figure 2 As shown, the software modules in the training device 455 of the material rough ranking model stored in the memory 450 may include: a sample acquisition module 4551, which is used to obtain a training sample set based on the material recommendation system; each training sample in the training sample set includes material information of the sample candidate material, user information and historical behavior information of the sample user; a training module 4552, which is used to iteratively train the initial material rough ranking model using the training sample set, input the training sample into the material rough ranking model for feature extraction and embedding processing, and obtain the material embedding vector of the sample candidate material, the user embedding vector and the interest sequence embedding vector of the sample user, and the cross embedding between the sample user and the sample candidate material. vector; and, based on the material embedding vector, the user embedding vector, the interest sequence embedding vector, and the cross embedding vector, predict the user's interest in the material, and obtain the rough ranking estimated score corresponding to the sample candidate material output by the material rough ranking model; the knowledge distillation module 4553 is used to input the training samples into the material fine ranking model used by the material recommendation system, and obtain the fine ranking estimated score corresponding to the sample candidate material output by the material fine ranking model; the parameter adjustment module 4554 is used to adjust the parameters of the material rough ranking model based on the difference between the rough ranking estimated score corresponding to the sample candidate material and the fine ranking estimated score, until the set convergence condition is reached, and a trained material rough ranking model is obtained.

[0138] In some possible implementations, the material rough ranking model includes a preprocessing layer, a four-tower layer and a first neural network, wherein: the preprocessing layer is used to perform feature extraction processing on the material information of the input sample candidate materials, the user information and historical behavior information of the sample users, and obtain the material characteristics of the sample candidate materials, the user characteristics and interest sequences of the sample users, and the cross-statistical characteristics between the sample users and the sample candidate materials; the interest sequence characteristics are used to characterize the interest preferences of the sample users, and the cross-statistical characteristics are used to characterize the relevant behavioral statistical information of the sample users on the sample candidate materials; the four-tower layer is used to embed the material characteristics, user characteristics, interest sequence characteristics and cross-statistical characteristics respectively, and obtain the corresponding material embedding vector, user embedding vector, interest sequence embedding vector and cross-embedding vector; the first neural network is used to predict the user's interest in the material based on the material embedding vector, user embedding vector, interest sequence embedding vector and cross-embedding vector, and obtain the rough ranking estimated score corresponding to the sample candidate material.

[0139] In some possible implementations, the four-tower layer includes a material tower, a user tower, an interest tower and a cross tower, wherein: the material tower is used to embed the input material features to obtain the corresponding material embedding vector; the user tower is used to embed the input user features to obtain the corresponding user embedding vector; the interest tower is used to embed the input interest sequence features to obtain the corresponding interest sequence embedding vector; the cross tower is used to embed the input cross-statistical features to obtain the corresponding cross-embedding vector.

[0140] In some possible implementations, the interest tower is used to vectorize the interest sequence features to obtain the interest sequence vector; matrix the interest sequence vector to obtain the weight vector; the elements in the weight vector correspond one-to-one to the elements in the interest sequence vector; perform vector inner product processing on the interest sequence vector and the weight vector to obtain the vector inner product result; and perform pooling processing on the vector inner product result to obtain the interest sequence embedding vector.

[0141] In some possible implementations, the training module 4552 is used to fuse the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector to obtain a fused vector; and to predict the user's interest level in the material on the fused vector to obtain a rough estimated score corresponding to each sample candidate material.

[0142] In some possible implementations, the above-mentioned device also includes a data storage module, which is used to save the material embedding vector, the user embedding vector and the interest sequence embedding vector to a database of the material recommendation system.

[0143] In some possible implementations, the parameter adjustment module 4554 is used to obtain the marking information of each sample candidate material; the marking information is used to indicate whether the sample user clicks on the sample candidate material; based on the marking information, the fine ranking estimated score, the rough ranking estimated score and the preset loss function corresponding to each sample candidate material, the loss value of the material rough ranking model is calculated; the loss function is used to characterize the difference between the rough ranking estimated score and the marking information, the difference between the fine ranking estimated score and the marking information, and the difference between the rough ranking estimated score and the fine ranking estimated score; based on the loss value of the material rough ranking model, the parameters of the material rough ranking model are adjusted until the set convergence condition is reached to obtain a trained material rough ranking model.

[0144] The embodiment of the present application provides a rough material sorting device 456, which includes: a data acquisition module 4561, which is used to respond to the material recommendation request of the target user, obtain the user information and historical behavior information of the target user, and the material information of multiple candidate materials; a vector acquisition module 4562, which is used to obtain the user embedding vector and interest sequence embedding vector of the target user, and the material embedding vector corresponding to each candidate material in the database of the material recommendation system based on the user information and historical behavior information of the target user, and the material information of multiple candidate materials; a data processing module 4563, which is used to convert the user information and historical behavior information of the target user, and the material information of multiple candidate materials into a data processing module 4564; The material information of each candidate material is input into the trained material rough ranking model for feature extraction and embedding processing to obtain the cross-embedding vector between the target user and each candidate material; and based on the user embedding vector and interest sequence embedding vector of the target user, the material embedding vector and cross-embedding vector corresponding to each candidate material, the user's interest in the material is predicted to obtain a rough ranking estimated score corresponding to each candidate material output by the material rough ranking model; a sorting module 4564 is used to train the material rough ranking model using the above-mentioned material rough ranking model training method; and multiple candidate materials are sorted based on the rough ranking estimated score corresponding to each candidate material.

[0145] In some possible implementations, the data processing module 4563 is used to preprocess the user information and historical behavior information of the target user through the preprocessing layer of the material rough sorting model to obtain cross-statistical features; and vectorize the cross-statistical features through the cross tower of the material rough sorting model to obtain a cross-embedded vector.

[0146] In some possible implementations, the data processing module 4563 is used to fuse the user embedding vector and the interest sequence embedding vector, the material embedding vector and the cross embedding vector corresponding to each candidate material through the fusion layer of the material rough ranking model to obtain a fusion vector; and predict the user's interest level in the material through the first neural network of the material rough ranking model to obtain a rough ranking estimated score corresponding to each candidate material.

[0147] The embodiment of the present application provides a computer program product or a computer program, which includes computer instructions, and the computer instructions are stored in a computer storage medium. The processor of the computer device reads the computer instructions from the computer storage medium, and the processor executes the computer instructions, so that the computer device executes the training method of the material rough arrangement model and the material rough arrangement method described in the embodiment of the present application.

[0148] The present application embodiment provides a computer storage medium storing executable instructions, wherein executable instructions are stored. When the executable instructions are executed by a processor, the processor will execute the material rough arrangement model training method and material rough arrangement method provided in the present application embodiment, for example Figure 3 The training method and Figure 5 The rough material sorting method is shown.

[0149] In some embodiments, the computer storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0150] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0151] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a hypertext markup language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0152] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0153] The above are only embodiments of the present application and are not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A training method for a rough material sorting model, characterized in that: The method comprises: Acquire a training sample set based on the material recommendation system; each training sample in the training sample set includes material information of the sample candidate material, user information and historical behavior information of the sample user; Iteratively training the initial rough material ranking model using the training sample set, inputting the training sample into the rough material ranking model for feature extraction and embedding processing, and obtaining the material embedding vector of the sample candidate material, the user embedding vector and interest sequence embedding vector of the sample user, and the cross embedding vector between the sample user and the sample candidate material; and, Based on the material embedding vector, the user embedding vector, the interest sequence embedding vector, and the cross embedding vector, a prediction process is performed on the user's interest level in the material, and a rough ranking estimated score corresponding to the sample candidate material output by the material rough ranking model is obtained; Inputting the training sample into the material ranking model used by the material recommendation system to obtain the estimated ranking score corresponding to the sample candidate material output by the material ranking model; Based on the difference between the rough ranking estimated score and the fine ranking estimated score corresponding to the sample candidate material, the parameters of the material rough ranking model are adjusted until the set convergence condition is reached, thereby obtaining a trained material rough ranking model.

2. The method according to claim 1, characterized in that The material rough sorting model includes a pre-processing layer, a four-tower layer and a first neural network, wherein: The preprocessing layer is used to perform feature extraction processing on the input material information of the sample candidate material, the user information and historical behavior information of the sample user, to obtain the material features of the sample candidate material, the user features and interest sequence features of the sample user, and the cross-statistical features between the sample user and the sample candidate material; the interest sequence features are used to characterize the interest preferences of the sample user, and the cross-statistical features are used to characterize the relevant behavioral statistical information of the sample user on the sample candidate material; The four tower layers are used to embed the material features, the user features, the interest sequence features, and the cross-statistical features, respectively, to obtain corresponding material embedding vectors, user embedding vectors, interest sequence embedding vectors, and cross-embedding vectors; The first neural network is used to predict the user's interest in a material based on the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector, and obtain a rough estimated score corresponding to the sample candidate material.

3. The method according to claim 2, characterized in that The four tower layers include a material tower, a user tower, an interest tower and a cross tower, wherein: The material tower is used to embed the input material features to obtain the corresponding material embedding vector; The user tower is used to embed the input user features to obtain the corresponding user embedding vector; The interest tower is used to embed the input interest sequence features to obtain the corresponding interest sequence embedding vector; The cross tower is used to embed the input cross statistical features to obtain the corresponding cross embedding vector.

4. The method according to claim 3, characterized in that The embedding process of the inputted interest sequence feature to obtain the corresponding interest sequence embedding vector includes: Performing vectorization processing on the interest sequence feature to obtain an interest sequence vector; Matrix processing is performed on the interest sequence vector to obtain a weight vector; the elements in the weight vector correspond one-to-one to the elements in the interest sequence vector; Performing vector inner product processing on the interest sequence vector and the weight vector to obtain a vector inner product result; Pooling is performed on the vector inner product results to obtain the embedding vector of the interest sequence.

5. The method according to claim 2, characterized in that: The predicting process of the user's interest in the material based on the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector to obtain a rough estimated score corresponding to the sample material includes: Fusing the material embedding vector, the user embedding vector, the interest sequence embedding vector and the cross embedding vector to obtain a fused vector; The fusion vector is processed to predict the user's interest in the material, and a rough estimated score corresponding to each sample candidate material is obtained.

6. The method according to claim 2, characterized in that After obtaining the material embedding vector, the user embedding vector and the interest sequence embedding vector, the method further includes: The material embedding vector, the user embedding vector and the interest sequence embedding vector are saved in a database of the material recommendation system.

7. The method according to claim 1, characterized in that The method of adjusting parameters of the material rough ranking model based on the difference between the rough ranking estimated score and the fine ranking estimated score corresponding to the sample candidate material until the set convergence condition is reached to obtain a trained material rough ranking model includes: Obtaining tag information of each sample candidate material; the tag information is used to indicate whether the sample user clicks on the sample candidate material; Based on the marking information, the refined ranking estimated score, the rough ranking estimated score and the preset loss function corresponding to each sample candidate material, the loss value of the rough ranking model of the material is calculated; the loss function is used to characterize the difference between the rough ranking estimated score and the marking information, the difference between the refined ranking estimated score and the marking information, and the difference between the rough ranking estimated score and the refined ranking estimated score; Based on the loss value of the material rough arrangement model, the parameters of the material rough arrangement model are adjusted until the set convergence condition is reached, so as to obtain a trained material rough arrangement model.

8. A material rough sorting method, characterized in that: The method comprises: In response to a material recommendation request from a target user, obtaining user information and historical behavior information of the target user, and material information of a plurality of candidate materials; Based on the user information and historical behavior information of the target user and the material information of the multiple candidate materials, a user embedding vector and an interest sequence embedding vector of the target user and a material embedding vector corresponding to each candidate material are obtained in a database of a material recommendation system; Input the user information and historical behavior information of the target user, and the material information of each candidate material into the trained material rough ranking model for feature extraction and embedding processing to obtain the cross-embedding vector between the target user and each candidate material; and based on the user embedding vector and interest sequence embedding vector of the target user, the material embedding vector and cross-embedding vector corresponding to each candidate material, perform prediction processing on the user's interest in the material to obtain the rough ranking estimated score corresponding to each candidate material output by the material rough ranking model; the material rough ranking model is trained by the method described in any one of claims 1 to 7; The multiple candidate materials are sorted based on the rough estimated score corresponding to each candidate material.

9. A rough material discharge device, characterized in that: The device comprises: A data collection module, configured to obtain user information and historical behavior information of the target user and material information of multiple candidate materials in response to a material recommendation request of the target user; A vector acquisition module, configured to obtain a user embedding vector and an interest sequence embedding vector of the target user and a material embedding vector corresponding to each candidate material in a database of a material recommendation system based on the user information and historical behavior information of the target user and the material information of the multiple candidate materials; A data processing module, for inputting the user information and historical behavior information of the target user and the material information of each candidate material into the trained material rough ranking model for feature extraction and embedding processing, so as to obtain a cross-embedding vector between the target user and each candidate material; and, based on the user embedding vector and interest sequence embedding vector of the target user, the material embedding vector and cross-embedding vector corresponding to each candidate material, performing prediction processing on the user's interest in the material, so as to obtain a rough ranking estimated score corresponding to each candidate material output by the material rough ranking model; the material rough ranking model is trained by the method described in any one of claims 1 to 7; A sorting module is used to sort the multiple candidate materials based on the rough estimated score corresponding to each candidate material.

10. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; A processor, configured to implement the method according to any one of claims 1 to 8 when executing the executable instructions or computer programs stored in the memory.

Images