Information recommendation method and device, electronic equipment, storage medium and program product

Abstract

The application provides an information recommendation method and device, electronic equipment, a storage medium and a program product. The method comprises the following steps: fusing a first historical behavior sequence of a target object in a target domain and a second historical behavior sequence of the target object in a source domain to obtain a fused behavior sequence; determining a first sequence feature of the first historical behavior sequence and a fused sequence feature of the fused behavior sequence; determining a second sequence feature of the second historical behavior sequence based on the first historical behavior sequence, the second historical behavior sequence and the fused behavior sequence; predicting an object feature of the target object in the target domain based on the first sequence feature, the second sequence feature and the fused sequence feature; and performing information recommendation for the target object in the target domain by combining the object feature and information features of each to-be-recommended information in the target domain. Through the application, the accuracy of information recommendation can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an information recommendation method, apparatus, electronic device, storage medium, and program product. Background Technology

[0002] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0003] In related technologies, the focus is usually on the migration of target objects from source domain preferences to target domain preferences or from global preferences to single domain preferences. However, the modeling of the mixed behavior sequence of target objects and the relationship between source domain preferences, target domain preferences and mixed domain preferences is neglected. As a result, in the process of cross-domain recommendation, the lack of modeling of the relationship between target domain preferences and mixed domain preferences leads to poor recommendation accuracy. Summary of the Invention

[0004] This application provides an information recommendation method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can effectively improve the accuracy of information recommendation.

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

[0006] This application provides an information recommendation method, including:

[0007] The first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain are fused to obtain a fused behavior sequence.

[0008] Determine the first sequence features of the first historical behavior sequence, and determine the fusion sequence features of the fused behavior sequence;

[0009] Based on the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence, the second sequence feature of the second historical behavior sequence is determined;

[0010] Based on the first sequence features, the second sequence features, and the fused sequence features, predict the object features of the target object in the target domain;

[0011] By combining the object features and the information features of each piece of information to be recommended in the target domain, information recommendation is performed on the target object within the target domain.

[0012] This application provides an information recommendation device, including:

[0013] The fusion module is used to fuse the first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain to obtain a fused behavior sequence.

[0014] The first determining module is used to determine the first sequence feature of the first historical behavior sequence and to determine the fusion sequence feature of the fusion behavior sequence;

[0015] The second determining module is used to determine the second sequence feature of the second historical behavior sequence based on the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence;

[0016] The prediction module is used to predict the object features of the target object in the target domain based on the first sequence features, the second sequence features, and the fused sequence features;

[0017] The information recommendation module is used to combine the object features and the information features of each piece of information to be recommended in the target domain to recommend information about the target object within the target domain.

[0018] In some embodiments, the fusion module is further configured to obtain each first sequence element in the first historical behavior sequence and the timestamp of each first sequence element; obtain each second sequence element in the second historical behavior sequence and the timestamp of each second sequence element; sort each first sequence element and each second sequence element according to the order of the timestamps; and determine the sorted first sequence elements and second sequence elements as the fused behavior sequence.

[0019] In some embodiments, the first determining module is further configured to obtain a first encoding feature of the first historical behavior sequence and a fusion encoding feature of the fused behavior sequence, wherein the first encoding feature includes a plurality of first sub-features, and the first sub-features correspond one-to-one with the first sequence elements in the first historical behavior sequence; based on the first encoding feature and the fusion encoding feature, determine the weight of each first sub-feature in the first encoding feature; and perform a weighted summation of each first sub-feature according to the weight of each first sub-feature to obtain the first sequence feature.

[0020] In some embodiments, the first determining module is further configured to perform the following processing on each of the first sub-features in the first coding feature: obtaining a target first sub-feature in the first coding feature and a target fusion sub-feature in the fusion coding feature; wherein the fusion coding feature includes multiple fusion sub-features, the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; determining a reference sub-feature corresponding to the first sub-feature based on the first sub-feature, the target first sub-feature, and the target fusion sub-feature; and performing normalization processing on the reference sub-feature corresponding to the first sub-feature to obtain the weight of the first sub-feature.

[0021] In some embodiments, the first determining module is further configured to multiply the norm of the first sub-feature, the target first sub-feature, and the first sub-feature to obtain a first multiplication result; determine the inner product of the first sub-feature and the target first sub-feature, and determine the norm of the inner product; multiply the norm of the inner product, the target fused sub-feature, and the target first sub-feature to obtain a second multiplication result; and subtract the first multiplication result from the second multiplication result to obtain a reference sub-feature corresponding to the first sub-feature.

[0022] In some embodiments, the first determining module is further configured to obtain a first encoding feature of the first historical behavior sequence and a fusion encoding feature of the fusion behavior sequence, wherein the fusion encoding feature includes a plurality of fusion sub-features, and the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence; based on the first encoding feature and the fusion encoding feature, determine the weight of each fusion sub-feature in the fusion encoding feature; and perform a weighted summation of each fusion sub-feature according to its weight to obtain the fusion sequence feature.

[0023] In some embodiments, the first determining module is further configured to perform the following processing on each of the fusion sub-features in the fusion coding feature: obtaining a target first sub-feature in the first coding feature and a target fusion sub-feature in the fusion coding feature; wherein, the first coding feature includes a plurality of first sub-features, the first sub-features correspond one-to-one with the first sequence elements in the first historical behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; determining a reference sub-feature corresponding to the fusion sub-feature based on the fusion sub-feature, the target first sub-feature, and the target fusion sub-feature; and performing normalization processing on the reference sub-feature corresponding to the fusion sub-feature to obtain the weight of the fusion sub-feature.

[0024] In some embodiments, the second determining module is further configured to obtain a first encoding feature of the first historical behavior sequence, a fusion encoding feature of the fusion behavior sequence, and a second encoding feature of the second historical behavior sequence, wherein the second encoding feature includes a plurality of second sub-features, and the second sub-features correspond one-to-one with the second sequence elements in the second historical behavior sequence; based on the first encoding feature, the second encoding feature, and the fusion encoding feature, determine the weight of each second sub-feature in the second encoding feature; and perform a weighted summation of each second sub-feature according to its weight to obtain the second sequence feature.

[0025] In some embodiments, the second determining module is further configured to perform the following processing on each of the second sub-features in the second coding feature: obtaining a target first sub-feature in the first coding feature and a target fusion sub-feature in the fusion coding feature; wherein the fusion coding feature includes multiple fusion sub-features, the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; determining a reference sub-feature corresponding to the second sub-feature based on the second sub-feature, the target first sub-feature, and the target fusion sub-feature; and performing normalization processing on the reference sub-feature corresponding to the second sub-feature to obtain the weight of the second sub-feature.

[0026] In some embodiments, the information recommendation module is further configured to obtain information features of each piece of information to be recommended in the target domain; compare the object features with each of the information features to obtain the credibility of each information feature; determine the information to be recommended corresponding to the highest credibility as the target recommendation information for the target object in the target domain; and send the target recommendation information to the terminal corresponding to the target object.

[0027] In some embodiments, the target information recommendation model includes a first target encoding layer, a second target encoding layer, a third target encoding layer, and a target prediction layer; the first determining module is further configured to call the first target encoding layer to encode the first historical behavior sequence to obtain the first sequence feature; the first determining module is further configured to call the second target encoding layer to encode the fused behavior sequence to obtain the fused sequence feature; the second determining module is further configured to call the third target encoding layer to encode the second historical behavior sequence by combining the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence to obtain the second sequence feature; the prediction module is further configured to call the target prediction layer to perform feature prediction on the target object by combining the first sequence feature, the second sequence feature, and the fused sequence feature to obtain the object feature of the target object in the target domain.

[0028] In some embodiments, the information recommendation device further includes: a training module, configured to acquire an information recommendation model, the information recommendation model including a first encoding layer, a second encoding layer, a third encoding layer, and a prediction layer; to fuse a first historical behavior sample sequence of a sample object in the target domain and a second historical behavior sample sequence of the sample object in the source domain to obtain a fused behavior sample sequence; to call the first encoding layer to encode the first historical behavior sample sequence to obtain a first sequence sample feature, and to call the second encoding layer to encode the fused behavior sample sequence to obtain a fused sequence sample feature; and to call the third encoding layer to combine the first historical behavior sample sequence, the second historical behavior sample sequence, and the fused behavior sample sequence to obtain a second historical behavior sample feature. The sequence is encoded to obtain second sequence sample features; a prediction network is invoked, and the first sequence sample features, the second sequence sample features, and the fused sequence sample features are combined to predict the features of the target object, obtaining object sample features of the sample object in the target domain; a first loss value is determined by combining the object sample features and the information features of each piece of information to be recommended in the target domain; a second loss value and a third loss value are determined based on the first sequence sample features, the fused sequence sample features, and the second sequence sample features; the first loss value, the second loss value, and the third loss value are weighted and summed to obtain a target loss value; the model parameters of the information recommendation model are updated based on the target loss value to obtain the target information recommendation model.

[0029] In some embodiments, the training module described above is further configured to determine the credibility between the object sample features and each of the information features; and to perform the following processing for each credibility: multiplying the credibility by the logarithm of the credibility to obtain a first multiplication result; determining the difference between the credibility and 1 as a target difference; multiplying the target difference by the logarithm of the target difference to obtain a second multiplication result; adding the first multiplication result and the second multiplication result to obtain a credibility sum; adding the credibility sums of each credibility to obtain a sum result, and determining the negative of the sum result as the first loss value.

[0030] In some embodiments, the training module is further configured to determine a first contrast loss value between the first sequence sample features and the fused sequence sample features; and determine a second contrast loss value between the first sequence sample features and the second sequence sample features; and determine a third contrast loss value between the fused sequence sample features and the second sequence sample features; and perform a weighted summation of the first contrast loss value, the second contrast loss value, and the third contrast loss value to obtain the second loss value.

[0031] In some embodiments, the training module is further configured to determine a first feature distance between the first sequence sample features and the second sequence sample features, and to determine a second feature distance between the fused sequence sample features and the second sequence sample features; and to determine a third loss value based on the difference between the first feature distance and the second feature distance.

[0032] This application provides an electronic device, including:

[0033] Memory is used to store executable instructions or computer programs.

[0034] When a processor executes computer-executable instructions or computer programs stored in the memory, it implements the information recommendation method provided in the embodiments of this application.

[0035] This application provides a computer-readable storage medium storing computer-executable instructions for inducing a processor to execute and implement the information recommendation method provided in this application.

[0036] This application provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. The processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the information recommendation method described in this application.

[0037] The embodiments of this application have the following beneficial effects:

[0038] By fusing the first and second historical behavior sequences, a fused behavior sequence is obtained. The first sequence features corresponding to the first historical behavior sequence, the second sequence features corresponding to the second historical behavior sequence, and the fused sequence features corresponding to the fused behavior sequence are determined. Based on these features, the object features of the target object in the target domain are predicted. Combining these object features with the information features of each piece of information to be recommended in the target domain, information recommendation is performed on the target object within the target domain. Thus, because the fused behavior sequence fully integrates the historical behavior of the target object in both the source and target domains, the prediction of the target object's features in the target domain fully considers this historical behavior, effectively improving the accuracy of the determined object features. Consequently, when recommending information to the target object based on these object features within the target domain, the accuracy of information recommendation is significantly improved. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the architecture of the information recommendation system provided in the embodiments of this application;

[0040] Figure 2 This is a schematic diagram of the structure of an electronic device for information recommendation provided in an embodiment of this application;

[0041] Figures 3 to 8 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application;

[0042] Figures 9 to 10 This is a schematic diagram illustrating the principle of the information recommendation method provided in the embodiments of this application;

[0043] Figure 11 This is a schematic diagram illustrating the principle of the target information recommendation model provided in the embodiments of this application;

[0044] Figure 12 This is a comparison chart of the effects of a universal experiment provided in the embodiments of this application;

[0045] Figure 13 This is a schematic diagram showing the results of the parameter experiment provided in the embodiments of this application;

[0046] Figure 14 This is a schematic diagram of the results of multi-domain representation visualization provided in the embodiments of this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is 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.

[0049] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0051] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0052] 1) Artificial Intelligence (AI): This refers to the theories, methods, technologies, and application systems that utilize digital computers or computers-controlled machines to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. AI technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics.

[0053] 2) Convolutional Neural Networks (CNNs): These are a class of feedforward neural networks (FNNs) that incorporate convolutional computations and have a deep structure. They are one of the representative algorithms of deep learning. CNNs possess representation learning capabilities, enabling them to perform shift-invariant classification of input images according to their hierarchical structure.

[0054] 3) Convolutional Layers: Each convolutional layer in a convolutional neural network consists of several convolutional units. The parameters of each convolutional unit are optimized through backpropagation. The purpose of convolution is to extract different features from the input. The first convolutional layer may only extract some low-level features such as edges, lines, and corners. More layers of the network can iteratively extract more complex features from low-level features.

[0055] 4) Pooling Layer: After feature extraction in the convolutional layer, the output feature map is passed to the pooling layer for feature selection and information filtering. The pooling layer contains a predefined pooling function, which replaces the result of a single point in the feature map with the feature map statistics of its neighboring regions. The selection of pooling regions by the pooling layer is the same as the step of the convolutional kernel scanning the feature map, and is controlled by the pooling size, stride, and padding.

[0056] 5) Fully-Connected Layer: A fully-connected layer in a convolutional neural network is equivalent to a hidden layer in a traditional feedforward neural network. The fully-connected layer is located at the end of the hidden layers in a convolutional neural network and only transmits signals to other fully-connected layers. Feature maps lose their spatial topology in the fully-connected layer, are unfolded into vectors, and pass through an activation function.

[0057] 6) Information to be recommended: This is information to be delivered. Information to be recommended has various forms of dissemination, such as news and advertisements. Among them, advertisements can carry purchase links.

[0058] 7) Recommendation metrics: Metrics that measure the positive response of the target audience to the recommended information, such as click-through rate and conversion rate.

[0059] 8) Recommendation Systems: These systems utilize e-commerce websites to provide customers with product information and suggestions, helping users decide what products to buy and mimicking the salesperson's assistance in completing the purchase process. Personalized recommendations recommend information and products that users are likely to be interested in based on their interests and purchasing behavior. With the continuous expansion of e-commerce and the rapid growth in the number and variety of products, customers need to spend a significant amount of time finding what they want to buy. This process of browsing through large amounts of irrelevant information and products undoubtedly leads to a continuous loss of consumers overwhelmed by information overload. To solve these problems, personalized recommendation systems have emerged. Personalized recommendation systems are advanced business intelligence platforms built on massive data mining to help e-commerce websites provide fully personalized decision support and information services for their customers' shopping.

[0060] 9) Contrastive Learning: Originating from unsupervised learning, this approach faces greater challenges compared to supervised learning algorithms due to the lack of labeled guidance, making it more difficult for the network to learn the features of samples during training. The core idea of ​​contrastive learning is to construct diversity in the original samples through data augmentation. The loss function is designed to narrow the distance between positive and anchor samples and increase the distance between positive and negative samples. In this process, the network more easily learns the common features shared by multiple samples after data augmentation, features that are likely essential to the original samples.

[0061] During the implementation of the embodiments of this application, the applicant discovered the following problems with the related technology:

[0062] In related technologies, the main works include Recommendation Systems (RS), Cross-domain Sequential Recommendation (CDSR), and Contrastive Learning (CL) for recommendation. Recommendation systems, when used to study the imbalance and sparsity of interaction information in single-domain sequences, struggle to achieve ideal results. Cross-domain Sequential Recommendation, which incorporates multi-domain information, focuses only on the transfer of target object preferences from source domain to target domain or from global preferences to single-domain preferences, neglecting to model the mixed behavioral sequences of target objects and the relationships between source domain preferences, target domain preferences, and mixed domain preferences. Existing works on Contrastive Learning rarely construct contrastive tasks from the perspective of cross-domain sequence modeling. Therefore, in cross-domain recommendation processes, the lack of modeling of the relationships between target domain preferences and mixed domain preferences leads to poor recommendation accuracy.

[0063] This application provides an information recommendation method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can effectively improve the accuracy of information recommendation. The following describes an exemplary application of the information recommendation system provided in this application.

[0064] See Figure 1 , Figure 1 This is a schematic diagram of the architecture of the information recommendation system 100 provided in the embodiments of this application. The terminal (terminal 400 is shown as an example) connects to the server 200 through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0065] Terminal 400 is used by the target user to access client 410 and displays recommended information on a graphical interface 410-1 (graphical interface 410-1 is shown as an example). Terminal 400 and server 200 are interconnected via wired or wireless network.

[0066] In some embodiments, server 200 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Terminal 400 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smart TV, smartwatch, in-vehicle terminal, etc., but is not limited to these. The electronic device provided in this application embodiment can be implemented as a terminal or a server. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited in this application embodiment.

[0067] In some embodiments, server 200 obtains a first historical behavior sequence of the target object in the target domain and a second historical behavior sequence of the target object in the source domain. Based on the first and second historical behavior sequences, it predicts the object features of the target object in the target domain. Combining the object features with the information features of each piece of information to be recommended in the target domain, it recommends information about the target object in the target domain and sends the target recommendation information to the terminal 400 corresponding to the target object.

[0068] In other embodiments, terminal 400 acquires a first historical behavior sequence of the target object in the target domain and a second historical behavior sequence of the target object in the source domain. Based on the first and second historical behavior sequences, it predicts the object features of the target object in the target domain and sends the object features to server 200. Server 200 combines the object features with the information features of each piece of information to be recommended in the target domain to recommend information to the target object in the target domain and sends the target recommendation information to the terminal 400 corresponding to the target object.

[0069] In other embodiments, the embodiments of this application can be implemented with the aid of cloud technology, which refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or local area network to realize the computation, storage, processing, and sharing of data.

[0070] Cloud technology is a general term encompassing network technology, information technology, integration technology, management platform technology, and application technology based on the cloud computing business model. It can form resource pools, allowing for on-demand use with flexibility and convenience. Cloud computing technology will become a crucial support. The backend services of cloud computing systems require substantial computing and storage resources.

[0071] See Figure 2 , Figure 2This is a schematic diagram of the structure of an electronic device 500 for information recommendation provided in an embodiment of this application, wherein, Figure 2 The electronic device 500 shown can be Figure 1 Server 200 or terminal 400 in the middle, Figure 2 The illustrated electronic device 500 includes at least one processor 410, a memory 450, and at least one network interface 420. The various components in the electronic device 500 are coupled together via a bus system 440. It is understood that the bus system 440 is used to implement communication between these components. In addition to a data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 440.

[0072] The 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. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0073] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices physically located away from the processor 410.

[0074] The memory 450 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 450 described in this application embodiment is intended to include any suitable type of memory.

[0075] In some embodiments, memory 450 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0076] Operating system 451 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0077] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420, such as Bluetooth, WiFi, and Universal Serial Bus.

[0078] In some embodiments, the information recommendation device provided in this application can be implemented in software. Figure 2 An information recommendation device 455 stored in memory 450 is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: a fusion module 4551, a first determination module 4552, a second determination module 4553, a prediction module 4554, and an information recommendation module 4555. These modules are logically connected and can therefore be arbitrarily combined or further divided according to their implemented functions. The functions of each module will be described below.

[0079] In other embodiments, the information recommendation device provided in this application can be implemented in hardware. As an example, the information recommendation device provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the information recommendation method provided in this application. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0080] In some embodiments, the terminal or server can implement the information recommendation method provided in this application by running a computer program or computer-executable instructions. For example, the computer program can be a native program in the operating system (e.g., a dedicated information recommendation program) or a software module, such as an information recommendation module that can be embedded in any program (e.g., an instant messaging client, a photo album program, an electronic map client, a navigation client); or it can be a native application (APP), i.e., a program that needs to be installed in the operating system to run. In summary, the above-mentioned computer program can be any form of application, module, or plugin.

[0081] The information recommendation method provided in this application will be described in conjunction with exemplary applications and implementations of the server or terminal provided in the embodiments of this application.

[0082] See Figure 3 , Figure 3 This is a flowchart illustrating the information recommendation method provided in the embodiments of this application, which will be combined with... Figure 3 Steps 101 to 106 are described below. The information recommendation method provided in this application embodiment can be implemented by the server or the terminal alone, or by the server and the terminal working together. The following description will take the implementation by the server alone as an example.

[0083] In step 101, the first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain are fused to obtain a fused behavior sequence.

[0084] In some embodiments, the target domain refers to a domain or use case where the historical behavior of the target object is relatively sparse, and the source domain refers to a domain or use case where the historical behavior of the target object is relatively dense. In transfer learning, the source domain refers to the domain or use case from which knowledge is transferred out, and the target domain refers to the domain or use case into which knowledge is transferred in. By transferring the relatively rich historical behavior of the target object in the source domain to the target domain, target recommendation information for the target object in the target domain is predicted. Thus, in the prediction process, the rich historical behavior of the target object in the source domain is fully utilized, effectively improving the accuracy of the predicted target recommendation information.

[0085] In some embodiments, within the same application scenario, the determination of the source domain and the target domain depends on the relative sparsity of the historical behaviors of the target objects. The sparsity of the historical behaviors of the target objects in the source domain is less than or equal to the sparsity of the historical behaviors of the target objects in the target domain. The determination of the source domain and the target domain also depends on the direction of historical behavior migration.

[0086] As an example, in the application scenario of online shopping, the source domain can be a specific purchase area for game items, and the target domain can be a specific purchase area for physical toy items.

[0087] Continuing with the previous example, the second historical behavior sequence of the target object in the source domain (the subdivided purchase domain of game items) can be: {The target object purchases game item A on January 5th, the target object purchases game item B on May 10th, the target object purchases game item A on October 20th, the target object purchases game item B on November 8th, the target object purchases game item A on November 9th, the target object purchases game item B on November 10th, and the target object purchases game item A on November 11th}.

[0088] Continuing with the previous example, the first historical behavior sequence of the target object in the target domain (the subdivided purchase domain of physical toy items) can be: {The target object purchased physical toy A on January 5th, and the target object purchased physical toy B on November 11th}.

[0089] As an example, in the application scenario of online shopping, the source domain can be a specific purchase area for movies, and the target domain can be a specific purchase area for books.

[0090] Continuing with the previous example, the second historical behavior sequence of the target object in the source domain (the specific purchase area of ​​the movie) can be: {The target object buys movie tickets for movie A on January 5th, the target object buys movie tickets for movie B on May 10th, the target object buys movie tickets for movie A on October 20th, the target object buys movie merchandise for movie B on November 8th, the target object buys movie tickets for movie A on November 9th, the target object buys movie tickets for movie B on November 10th, and the target object buys movie merchandise for movie A on November 11th}.

[0091] Continuing with the previous example, the first historical behavior sequence of the target object in the target domain (the subdivided purchase domain of books) can be: {The target object purchased book A on January 5th, and the target object purchased book B on November 11th}.

[0092] As an example, in multimedia playback applications, the source domain can be a sub-category of audio playback, and the target domain can be a sub-category of video playback.

[0093] Continuing with the previous example, the second historical behavior sequence of the target object in the source domain (the subcategories of audio playback) can be: {The target object clicked to listen to audio on January 5th, the target object clicked to listen to audio on May 10th, the target object clicked to listen to audio on October 20th, the target object clicked to listen to audio on November 8th, the target object clicked to listen to audio on November 9th, the target object clicked to listen to audio on November 10th, and the target object clicked to listen to audio on November 11th}.

[0094] Continuing from the previous example, the first historical behavior sequence of the target object in the target domain (the subcategories of video playback) can be: {the target object clicked to watch the video on January 5th, and the target object clicked to watch the video on November 11th}.

[0095] In some embodiments, see Figure 4 , Figure 4 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application. Figure 4 Step 101 shown can be achieved by performing steps 1011 to 1013.

[0096] In step 1011, each first sequence element in the first historical behavior sequence and the timestamp of each first sequence element are obtained.

[0097] In some embodiments, the first historical behavior sequence includes at least two first sequence elements, the timestamp of the first first sequence element in the first historical behavior sequence is the earliest timestamp in the first historical behavior sequence, and each first sequence element in the first historical behavior sequence is arranged in the order of its timestamps, with each first sequence element corresponding to a timestamp.

[0098] In some embodiments, the timestamp of the first sequence element represents the historical time node corresponding to when the target object triggers the historical behavior corresponding to the first sequence element.

[0099] As an example, when the first historical behavior sequence is: {the target object clicked to watch the video on January 5th, the target object clicked to watch the video on November 11th}, the first sequence element is: the target object clicked to watch the video on January 5th, the target object clicked to watch the video on January 5th; where, the first sequence element: the target object clicked to watch the video on January 5th has a corresponding timestamp of: January 5th; the first sequence element: the target object clicked to watch the video on November 11th has a corresponding timestamp of: November 11th.

[0100] In step 1012, each second sequence element in the second historical behavior sequence and the timestamp of each second sequence element are obtained.

[0101] In some embodiments, the second historical behavior sequence includes at least two second sequence elements, the timestamp of the first second sequence element in the second historical behavior sequence is the earliest timestamp in the second historical behavior sequence, and the second sequence elements in the second historical behavior sequence are arranged in chronological order of their timestamps, with each second sequence element corresponding to a timestamp.

[0102] In some embodiments, the timestamp of the second sequence element represents the historical time node corresponding to when the target object triggers the historical behavior corresponding to the second sequence element.

[0103] As an example, when the second historical behavior sequence is: {the target object clicked to listen to the audio on January 5th, the target object clicked to listen to the audio on May 10th, the target object clicked to listen to the audio on October 20th, the target object clicked to listen to the audio on November 8th, the target object clicked to listen to the audio on November 9th, the target object clicked to listen to the audio on November 10th, and the target object clicked to listen to the audio on November 11th}, the second sequence elements are: (1) the target object clicked to listen to the audio on January 5th, (2) the target object clicked to listen to the audio on May 10th, and (3) the target object clicked on October 20th. Listen to the audio, (4) the target object clicks to listen to the audio on November 8, (5) the target object clicks to listen to the audio on November 9, (6) the target object clicks to listen to the audio on November 10, (7) the target object clicks to listen to the audio on November 11; among which, the second sequence element: the target object clicks to listen to the audio on January 5, and the corresponding timestamp is: January 5; the second sequence element: the target object clicks to listen to the audio on May 10, and the corresponding timestamp is: May 10; the second sequence element: the target object clicks to listen to the audio on October 20, and the corresponding timestamp is: October 20.

[0104] In step 1013, the first sequence elements and the second sequence elements are sorted according to the order of their timestamps, and the sorted first sequence elements and second sequence elements are determined as the fusion behavior sequence.

[0105] In some embodiments, the fusion behavior sequence includes each first sequence element and each second sequence element, that is, the fusion sequence element in the fusion behavior sequence can be either a first sequence element or a second sequence element.

[0106] As an example, when the first historical behavior sequence is: {the target audience clicked to watch a video on January 5th, the target audience clicked to watch a video on November 11th}, and the second historical behavior sequence is: {the target audience clicked to listen to audio on January 5th, the target audience clicked to listen to audio on May 10th, the target audience clicked to listen to audio on October 20th, the target audience clicked to listen to audio on November 8th, the target audience clicked to listen to audio on November 9th, the target audience clicked to listen to audio on November 10th, and the target audience clicked to listen to audio on November 11th}; according to the chronological order of the timestamps, each first sequence... Sort the first and second sequence elements, and determine the sorted first and second sequence elements as the fused behavior sequence: {Target object clicked to watch video on January 5, target object clicked to listen to audio on January 5, target object clicked to listen to audio on May 10, target object clicked to listen to audio on October 20, target object clicked to listen to audio on November 8, target object clicked to listen to audio on November 9, target object clicked to listen to audio on November 10, target object clicked to listen to audio on November 11, target object clicked to watch video on November 11}.

[0107] Thus, by fusing the first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain, a fused behavior sequence is obtained. Through the fused behavior sequence, target recommendation information for the target object in the target domain is predicted. In this way, the historical behavior of the target object in the source domain and the target domain is fully combined during the prediction process, which effectively improves the accuracy of the predicted target recommendation information.

[0108] In step 102, the first sequence feature of the first historical behavior sequence is determined.

[0109] In some embodiments, the first sequence feature of the first historical behavior sequence is determined based on the first historical behavior sequence and the fused behavior sequence. The first sequence feature of the first historical behavior sequence is used to characterize the overall historical behavior of the target object in the target domain, and the degree to which the overall historical behavior of the target object in the target domain is influenced by the historical behavior of the target object in the source domain.

[0110] In some embodiments, see Figure 5 , Figure 5 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application. Figure 5 Step 102 shown can be achieved by performing steps 1021 to 1023.

[0111] In step 1021, the first coding feature of the first historical behavior sequence and the fusion coding feature of the fusion behavior sequence are obtained.

[0112] In some embodiments, step 1021 above can be implemented as follows: encoding the first historical behavior sequence to obtain the first encoded feature of the first historical behavior sequence; encoding the fused behavior sequence to obtain the fused encoded feature.

[0113] As an example, see Figure 10 , Figure 10 This is a schematic diagram illustrating the principle of the information recommendation method provided in the embodiments of this application, for the first historical behavior sequence. Encoding is performed to obtain the first encoded feature H of the first historical behavior sequence. T , for the fusion behavior sequence Encode the features to obtain the fused encoded features H. M .

[0114] In some embodiments, encoding the first historical behavior sequence to obtain the first encoded feature of the first historical behavior sequence can be achieved by: determining the first feature matrix D of the first historical behavior sequence. T Linear projection is performed on the first feature matrix to obtain the query feature Q, key feature K, and value feature V of the first historical behavior sequence. Based on the query feature, key feature, and value feature of the first historical behavior sequence, a self-attention network is invoked for self-attention encoding to obtain the implicit matrix of the first historical behavior sequence. The activation function is called to activate the implicit matrix, thus obtaining the first encoded feature of the first historical behavior sequence.

[0115] In some embodiments, the implicit matrix of the first historical behavior sequence The expression can be:

[0116]

[0117] Where Q = D T W Q K=D T W K V=D T W V W Q W K W V These represent linear mapping layers.

[0118] In some embodiments, the expression for the first encoded feature of the first historical behavior sequence can be:

[0119]

[0120] Among them, H TThe first encoded feature represents the first historical behavior sequence, ReLU represents the activation function, and w1, w2, b1, and b2 represent the weight matrix and bias vector, respectively.

[0121] In some embodiments, the fused coding feature H M Second coding feature H s The method of determining it is the same as the first coding feature H mentioned above. T The method for determining it is similar and will not be elaborated here.

[0122] In some embodiments, the first coding feature includes a plurality of first sub-features, and the first sub-features correspond one-to-one with the first sequence elements in the first historical behavior sequence.

[0123] In step 1022, the weights of each first sub-feature in the first coding feature are determined based on the first coding feature and the fused coding feature.

[0124] In some embodiments, the weight of the first sub-feature is used to characterize the importance of the first sub-feature in the first coding feature, and the weight of the first sub-feature is positively correlated with the importance of the first sub-feature in the first coding feature.

[0125] In some embodiments, step 1022 above can be implemented by performing the following processing on each first sub-feature in the first coding feature: obtaining the target first sub-feature in the first coding feature and the target fusion sub-feature in the fusion coding feature; wherein, the fusion coding feature includes multiple fusion sub-features, the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; based on the first sub-feature, the target first sub-feature, and the target fusion sub-feature, determining the reference sub-feature corresponding to the first sub-feature; and normalizing the reference sub-feature corresponding to the first sub-feature to obtain the weight of the first sub-feature.

[0126] As an example, when the first encoded feature is: The fusion coding features are: At that time, the first sub-feature of the target is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence. The target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion history behavior sequence.

[0127] In some embodiments, the determination of the reference sub-feature corresponding to the first sub-feature based on the first sub-feature, the target first sub-feature, and the target fused sub-feature can be achieved as follows: multiplying the norm of the first sub-feature, the target first sub-feature, and the first sub-feature to obtain a first multiplication result; determining the inner product of the first sub-feature and the target first sub-feature, and determining the norm of the inner product; multiplying the norm of the inner product, the target fused sub-feature, and the target first sub-feature to obtain a second multiplication result; and subtracting the first multiplication result from the second multiplication result to obtain the reference sub-feature corresponding to the first sub-feature.

[0128] In some embodiments, the expression for the reference sub-feature corresponding to the first sub-feature can be:

[0129]

[0130] in, The reference sub-feature that represents the first sub-feature Representing the result of the first multiplication, Characterizing the result of the second multiplication, The first sub-feature representing the target The norm representing the first sub-feature Characterizing the first sub-feature, Characterize the target fusion sub-features, The norm of the inner product of the first sub-feature and the target first sub-feature is used to characterize a two-layer fully connected network.

[0131] In some embodiments, the expression for the weight of the first sub-feature can be:

[0132]

[0133] in, The weights represent the first sub-features, and Softmax represents the normalization function.

[0134] In step 1023, the first sub-features are weighted and summed according to their respective weights to obtain the first sequence features.

[0135] In some embodiments, the expression for the first sequence feature described above can be:

[0136]

[0137] Wherein, sT represents the features of the first sequence. The weights representing the first sub-feature, Characterize the first sub-feature.

[0138] Thus, by determining the weights of each first sub-feature in the first coding feature based on the first coding feature and the fused coding feature, the fused coding feature is fully considered in the process of determining the weights of each first sub-feature in the first coding feature. Since the fused coding feature can more comprehensively reflect the historical behavior of the target object in the source domain and the target domain, the weights of each first sub-feature are more accurate, effectively improving the accuracy of the determined weights of the first sub-features. Then, according to the weights of each first sub-feature, the first sub-features are weighted and summed to obtain the first sequence feature, effectively improving the accuracy of the determined first sequence feature.

[0139] In step 103, the fusion sequence characteristics of the fusion behavior sequence are determined.

[0140] In some embodiments, the fusion sequence features of the fusion behavior sequence are determined based on the first historical behavior sequence and the fusion behavior sequence, and the fusion sequence features of the fusion behavior sequence are used to characterize the overall historical behavior of the target object in the source domain and the target domain.

[0141] In some embodiments, see Figure 6 , Figure 6 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application. Figure 6 Step 103 shown can be achieved by performing steps 1031 to 1033.

[0142] In step 1031, the first coding feature of the first historical behavior sequence and the fusion coding feature of the fusion behavior sequence are obtained.

[0143] In some embodiments, the fusion coding feature includes multiple fusion sub-features, and the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence.

[0144] In step 1032, the weights of each fusion sub-feature in the fusion coding feature are determined based on the first coding feature and the fusion coding feature.

[0145] In some embodiments, the weight of the fusion sub-feature is used to characterize the importance of the fusion sub-feature in the fusion coding feature, and the weight of the fusion sub-feature is positively correlated with the importance of the fusion sub-feature in the fusion coding feature.

[0146] In some embodiments, step 1032 above can be implemented by performing the following processing on each fusion sub-feature in the fusion coding feature: obtaining the target first sub-feature in the first coding feature and the target fusion sub-feature in the fusion coding feature; wherein, the first coding feature includes multiple first sub-features, the first sub-features correspond one-to-one with the first sequence elements in the first historical behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; based on the fusion sub-feature, the target first sub-feature, and the target fusion sub-feature, determining the reference sub-feature corresponding to the fusion sub-feature; and normalizing the reference sub-feature corresponding to the fusion sub-feature to obtain the weight of the fusion sub-feature.

[0147] As an example, when the first encoded feature is: The fusion coding features are: At that time, the first sub-feature of the target is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence. The target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion history behavior sequence.

[0148] In some embodiments, the determination of the reference sub-feature corresponding to the fusion sub-feature based on the fusion sub-feature, the target first sub-feature, and the target fusion sub-feature can be achieved as follows: multiplying the norm of the fusion sub-feature, the target fusion sub-feature, and the fusion sub-feature to obtain a third multiplication result; determining the inner product of the fusion sub-feature and the target fusion sub-feature, and determining the norm of the inner product; multiplying the norm of the inner product, the target fusion sub-feature, and the target fusion sub-feature to obtain a fourth multiplication result; and subtracting the third multiplication result from the fourth multiplication result to obtain the reference sub-feature corresponding to the fusion sub-feature.

[0149] In some embodiments, the expression for the reference sub-feature corresponding to the above-mentioned fused sub-feature can be:

[0150]

[0151] in, The reference sub-features that represent the fused sub-features Characterizing the result of the third multiplication, The result of the fourth multiplication is represented by MLPa, which represents a two-layer fully connected network.

[0152] In some embodiments, the expression for the weights of the above-mentioned fused sub-features can be:

[0153]

[0154] in, The weights represent the features of the fusion sub-features, and Softmax represents the normalization function.

[0155] In step 1033, the fusion sub-features are weighted and summed according to their respective weights to obtain the fusion sequence features.

[0156] In some embodiments, the expression for the above-mentioned fused sequence features can be:

[0157]

[0158] Among them, sM represents the characteristics of the fusion sequence. Weights representing the features of the fusion sub-features Characterize the features of the fusion component.

[0159] Thus, by determining the weights of each fusion sub-feature in the fusion coding feature based on the first coding feature and the fusion coding feature, the fusion coding feature is fully considered in the process of determining the weights of each fusion sub-feature. Since the fusion coding feature can more comprehensively reflect the historical behavior of the target object in the source domain and the target domain, the weights of each fusion sub-feature are more accurate, which effectively improves the accuracy of the determined weights of the fusion sub-features. Then, according to the weights of each fusion sub-feature, the fusion sub-features are weighted and summed to obtain the fusion sequence feature, which effectively improves the accuracy of the determined fusion sequence feature.

[0160] In step 104, based on the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence, the second sequence feature of the second historical behavior sequence is determined.

[0161] In some embodiments, the second sequence feature of the second historical behavior sequence is determined based on the first historical behavior sequence, the fused behavior sequence, and the second historical behavior sequence. The second sequence feature of the second historical behavior sequence is used to characterize the overall historical behavior of the target object in the source domain.

[0162] In some embodiments, see Figure 7 , Figure 7 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application. Figure 7 Step 104 shown can be achieved by performing steps 1041 to 1043.

[0163] In step 1041, the first coding feature of the first historical behavior sequence, the fusion coding feature of the fusion behavior sequence, and the second coding feature of the second historical behavior sequence are obtained.

[0164] In some embodiments, the second coding feature includes a plurality of second sub-features, each of which corresponds one-to-one with a second sequence element in the second historical behavior sequence.

[0165] In some embodiments, the encoding process of the first encoding feature, the fused encoding feature, and the second encoding feature in step 1041 above has been described in detail above and will not be repeated here.

[0166] In step 1042, the weights of each second sub-feature in the second coding feature are determined based on the first coding feature, the second coding feature, and the fused coding feature.

[0167] In some embodiments, the weight of the second sub-feature is used to characterize the importance of the second sub-feature in the second coding feature, and the weight of the second sub-feature is positively correlated with the importance of the second sub-feature in the second coding feature.

[0168] In some embodiments, step 1042 above can be implemented by performing the following processing on each second sub-feature in the second coding feature: obtaining the target first sub-feature in the first coding feature and the target fusion sub-feature in the fusion coding feature; wherein, the fusion coding feature includes multiple fusion sub-features, the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; based on the second sub-feature, the target first sub-feature, and the target fusion sub-feature, determining the reference sub-feature corresponding to the second sub-feature; and normalizing the reference sub-feature corresponding to the second sub-feature to obtain the weight of the second sub-feature.

[0169] As an example, when the first encoded feature is: The fusion coding features are: At that time, the first sub-feature of the target is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence. The target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion history behavior sequence.

[0170] In some embodiments, the determination of the reference sub-feature corresponding to the second sub-feature based on the second sub-feature, the target first sub-feature, and the target fused sub-feature can be achieved as follows: multiplying the norm of the second sub-feature, the target fused sub-feature, and the second sub-feature to obtain a fifth multiplication result; determining the inner product of the second sub-feature and the target fused sub-feature, and determining the norm of the inner product; multiplying the norm of the inner product, the target fused sub-feature, and the target first sub-feature to obtain a sixth multiplication result; and subtracting the fifth multiplication result from the sixth multiplication result to obtain the reference sub-feature corresponding to the second sub-feature.

[0171] In some embodiments, the expression for the reference sub-feature corresponding to the second sub-feature can be:

[0172]

[0173] in, The reference sub-feature that represents the second sub-feature Characterizes the result of the fifth multiplication. Characterizing the result of the sixth multiplication, The first sub-feature representing the target The norm that characterizes the second sub-feature Characterizing the second sub-feature, Characterize the target fusion sub-features, The norm of the inner product of the first sub-feature and the target first sub-feature is used to characterize a two-layer fully connected network.

[0174] In some embodiments, the expression for the weight of the second sub-feature can be:

[0175]

[0176] in, The weights represent the second sub-features, and Softmax represents the normalization function.

[0177] In step 1043, the second sub-features are weighted and summed according to their respective weights to obtain the second sequence features.

[0178] In some embodiments, the expression for the second sequence feature described above can be:

[0179]

[0180] Where ss represents the second sequence feature, The weights representing the second sub-features Characterizes the second sub-feature.

[0181] Thus, by determining the weights of each second sub-feature in the second coding feature based on the first coding feature, the fused coding feature, and the second coding feature, the fused coding feature is fully considered in the process of determining the weights of each second sub-feature in the second coding feature. Since the fused coding feature can more comprehensively reflect the historical behavior of the target object in the source domain and the target domain, the weights of each determined second sub-feature are more accurate, effectively improving the accuracy of the determined weights of the second sub-features. Then, according to the weights of each second sub-feature, the second sub-features are weighted and summed to obtain the second sequence feature, effectively improving the accuracy of the determined second sequence feature.

[0182] In step 105, based on the first sequence features, the second sequence features, and the fused sequence features, the object features of the target object in the target domain are predicted.

[0183] As an example, see Figure 10 Based on the first sequence features s T Second sequence features s s and fusion sequence features s M Through the prediction layer, the object features U of the target object in the target domain are predicted. T .

[0184] In step 106, information recommendation is performed on the target object within the target domain by combining the object features and the information features of each piece of information to be recommended in the target domain.

[0185] As an example, when the first historical behavior sequence of the target object is: {the target object clicked to watch a video on January 5th, the target object clicked to watch a video on November 11th}, the recommended information in the target domain can be various types of videos such as: Video A, Video B, Video C, Video D, Video E, etc.

[0186] In some embodiments, see Figure 8 , Figure 8 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application. Figure 8 Step 106 shown can be achieved by performing steps 1061 to 1063.

[0187] In step 1061, the information features of each piece of information to be recommended in the target domain are obtained.

[0188] In some embodiments, step 1061 above can be implemented as follows: obtain each piece of information to be recommended in the target domain, encode each piece of information to be recommended, and obtain the information features of each piece of information to be recommended.

[0189] In step 1062, the object features are compared with each information feature to obtain the credibility of each information feature.

[0190] In some embodiments, the credibility of an information feature represents the degree of matching between the information feature and the object feature, and the degree of matching is positively correlated with the credibility value.

[0191] In some embodiments, step 1062 above can be implemented as follows: performing processing for each information feature separately: determining the credibility between the information feature and the object feature.

[0192] In step 1063, the information to be recommended corresponding to the highest credibility is determined as the target recommendation information for the target object under the target domain.

[0193] As an example, the credibility of information A to be recommended is 0.6, the credibility of information B to be recommended is 0.5, the credibility of information C to be recommended is 0.1, and the credibility of information D to be recommended is 0.9. Then, information D with the highest credibility of 0.9 is determined as the target recommendation information for the target object in the target domain.

[0194] In step 1064, target recommendation information is sent to the terminal corresponding to the target object.

[0195] In some embodiments, after the server determines the target recommendation information from the various recommendation information to be recommended in the target domain, the server sends the target recommendation information to the terminal corresponding to the target object.

[0196] In this way, by comparing the accurate object features with the information features of each piece of information to be recommended, the target recommendation information that best matches the target object is accurately determined, and the target recommendation information is sent to the terminal corresponding to the target object, thereby effectively improving the recommendation accuracy.

[0197] In some embodiments, the information recommendation method provided in this application can be implemented based on a target information recommendation model, which includes a first target encoding layer, a second target encoding layer, a third target encoding layer, and a target prediction layer.

[0198] As an example, see Figure 11 , Figure 11 This is a schematic diagram of the target information recommendation model provided in the embodiments of this application. The information recommendation method provided in the embodiments of this application can be implemented based on the target information recommendation model, which includes a first target encoding layer, a second target encoding layer, a third target encoding layer, and a target prediction layer.

[0199] In some embodiments, step 102 above can be implemented as follows: call the first target encoding layer to encode the first historical behavior sequence to obtain the first sequence features.

[0200] As an example, see Figure 11 Invoke the first target encoding layer to process the first historical behavior sequence S T Encode the first sequence feature H. T .

[0201] In some embodiments, step 103 above can be implemented as follows: call the second target encoding layer to encode the fusion behavior sequence to obtain fusion sequence features.

[0202] As an example, see Figure 11 The second target encoding layer is invoked to process the fusion behavior sequence S. M Encode the fused sequence features H. M .

[0203] In some embodiments, step 104 above can be implemented as follows: calling the third target encoding layer, combining the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence, to encode the second historical behavior sequence to obtain the second sequence feature.

[0204] As an example, see Figure 11 The third target encoding layer is invoked, combined with the first historical behavior sequence S. T Second historical behavior sequence S S and fusion behavior sequence S M For the second historical behavior sequence S S Encode the second sequence feature H. S .

[0205] In some embodiments, step 105 above can be implemented as follows: call the target prediction layer, combine the first sequence features, the second sequence features and the fused sequence features to perform feature prediction on the target object, and obtain the object features of the target object in the target domain.

[0206] As an example, see Figure 11 Call the target prediction layer and combine it with the first sequence features H T Second sequence feature H S and fusion sequence features H M Feature prediction is performed on the target object to obtain the object feature U of the target object in the target domain. T .

[0207] In some embodiments, the information recommendation method provided in the application embodiments can be implemented based on a target information recommendation model. The target information recommendation model can be obtained by training the information recommendation model. The training process of the information recommendation model is described in detail below.

[0208] In some embodiments, prior to step 102 above, the target information recommendation model can be trained as follows: An information recommendation model is obtained, comprising a first encoding layer, a second encoding layer, a third encoding layer, and a prediction layer; the first historical behavior sample sequence of the sample object in the target domain and the second historical behavior sample sequence of the sample object in the source domain are fused to obtain a fused behavior sample sequence; the first encoding layer is invoked to encode the first historical behavior sample sequence to obtain first sequence sample features, and the second encoding layer is invoked to encode the fused behavior sample sequence to obtain fused sequence sample features; and the third encoding layer is invoked to combine the first historical behavior sample sequence, the second historical behavior sample sequence, and the fused behavior sample sequence... The system encodes the second historical behavior sample sequence to obtain the second sequence sample features; it then calls a prediction network to combine the first sequence sample features, the second sequence sample features, and the fused sequence sample features to predict the features of the target object, obtaining the object sample features of the sample object in the target domain; it then determines the first loss value by combining the object sample features and the information features of each piece of information to be recommended in the target domain; based on the first sequence sample features, the fused sequence sample features, and the second sequence sample features, it determines the second loss value and the third loss value; it then performs a weighted sum of the first, second, and third loss values ​​to obtain the target loss value; and finally, it updates the model parameters of the information recommendation model based on the target loss value to obtain the target information recommendation model.

[0209] In some embodiments, the target information recommendation model and the information recommendation model have the same model structure but different model parameters. The model parameters of the target information recommendation model are obtained by updating the model parameters of the information recommendation model.

[0210] In some embodiments, the sample object and the target object are different objects. It is understood that different objects have different historical behavior sequences in the target domain and different objects have different historical behavior sequences in the source domain.

[0211] In some embodiments, the determination of the first loss value by combining the object sample features and the information features of each piece of information to be recommended in the target domain can be achieved as follows: determine the credibility between the object sample features and each information feature; perform the following processing on each credibility: multiply the credibility by the logarithm of the credibility to obtain the first multiplication result; determine the difference between the credibility and 1 as the target difference; multiply the target difference by the logarithm of the target difference to obtain the second multiplication result; add the first multiplication result and the second multiplication result to obtain the credibility sum; add the credibility sums of each credibility to obtain the sum result, and determine the negative number of the sum result as the first loss value.

[0212] In some embodiments, the expression for the first loss value can be:

[0213]

[0214] in, Characterizing the first loss value, Characterizes the reliability of object sample features and the relationships between various information features. The logarithm, which represents credibility, Characterize the target difference, The logarithm represents the difference between the target values, and RT is the training set of the target domain. and These represent positive samples and corresponding randomly sampled negative samples, respectively.

[0215] In some embodiments, the determination of the second loss value based on the first sequence sample features, the fused sequence sample features, and the second sequence sample features can be achieved as follows: determining a first contrast loss value between the first sequence sample features and the fused sequence sample features; determining a second contrast loss value between the first sequence sample features and the second sequence sample features; determining a third contrast loss value between the fused sequence sample features and the second sequence sample features; and weighted summing the first contrast loss value, the second contrast loss value, and the third contrast loss value to obtain the second loss value.

[0216] In some embodiments, the expression for the second loss value can be:

[0217]

[0218] Among them, L CSM Characterizing the second loss value, Characterizing the second contrastive loss value, Characterizing the third contrast loss value, The first contrastive loss value is represented by λ1, λ2, and λ3, which represent the loss weights.

[0219] In some embodiments, the expression for the first contrastive loss value can be:

[0220]

[0221] Where B represents the batch obtained from the sampling. This represents the cosine similarity.

[0222] In some embodiments, the expression for the second contrastive loss value can be:

[0223]

[0224] Where B represents the batch obtained from the sampling. This represents the cosine similarity.

[0225] In some embodiments, the expression for the third contrast loss value can be:

[0226]

[0227] Where B represents the batch obtained from the sampling. This represents the cosine similarity.

[0228] In some embodiments, the determination of the third loss value can be achieved by: determining a first feature distance between the first sequence sample features and the second sequence sample features, and determining a second feature distance between the fused sequence sample features and the second sequence sample features; and determining the third loss value based on the difference between the first feature distance and the second feature distance.

[0229] In some embodiments, the expression for the third loss value can be:

[0230]

[0231] Among them, LFDM represents the third loss value. Characterizing the first feature distance, The second feature distance is represented by γ, which represents the distance parameter to be controlled by the ternary loss.

[0232] In some embodiments, the expression for the target loss value can be:

[0233] L = L CTR +λ CSM L CSM +λ FDM L FDM (18)

[0234] Where L represents the target loss value, L FDM Characterizing the third loss value, L CSM Characterizing the second loss value, The first loss value, λ CSM and λ FDM L CSM and L FDM The weight.

[0235] Thus, by fusing the first and second historical behavior sequences to obtain a fused behavior sequence, the first sequence features corresponding to the first historical behavior sequence, the second sequence features corresponding to the second historical behavior sequence, and the fused sequence features corresponding to the fused behavior sequence are determined. Based on the first sequence features, the second sequence features, and the fused sequence features, the object features of the target object in the target domain are predicted. Combining the object features with the information features of each piece of information to be recommended in the target domain, information recommendation is performed for the target object in the target domain. Because the fused behavior sequence fully integrates the historical behavior of the target object in both the source and target domains, the prediction of the target object's features in the target domain fully considers the historical behavior of the target object in both domains, thereby effectively improving the accuracy of the determined object features. This, in turn, effectively improves the accuracy of information recommendation when recommending information about the target object in the target domain based on the object features.

[0236] The following will describe an exemplary application of the embodiments of this application in a real-world cross-domain recommendation scenario.

[0237] In a practical cross-domain recommendation application scenario, see [link to relevant documentation]. Figure 9 , Figure 9 This is a schematic diagram illustrating the principle of the information recommendation method provided in this application embodiment. The second historical behavior sequence 4 of target object 1 in the source domain can be: target object reads anime books at the time node corresponding to timestamp 11, and target object reads Go books at the time node corresponding to timestamp 14; the first historical behavior sequence 2 of target object 1 in the target domain can be: target object watches movie A at the time node corresponding to timestamp 12, and target object watches movie B at the time node corresponding to timestamp 13; the fused behavior sequence 3 of target object 1 is: target object reads anime books at the time node corresponding to timestamp 11, target object watches movie A at the time node corresponding to timestamp 12, target object watches movie B at the time node corresponding to timestamp 13, and target object reads Go books at the time node corresponding to timestamp 14. Based on the above-mentioned first historical behavior sequence features, second historical behavior sequence features, and fused sequence features, information recommendation is performed on the target object in the target domain using the information recommendation method provided in this application embodiment.

[0238] This application proposes an effective and universally applicable ternary learning method for cross-domain recommendation. It is the first to propose modeling learning of ternary sequences, that is, in cross-domain recommendation tasks, in addition to the behavioral sequences of the source domain and the target domain, the behavioral information of the target object in the mixed domain is also considered. The relevance and difference of the three-domain sequence representations are modeled at multiple granularities through a ternary cross-domain attention mechanism and a ternary contrastive learning strategy, thereby improving the cross-domain recommendation performance of existing models.

[0239] This application's embodiments, through comparison with existing algorithms and ablation experiments, found that the proposed Tri-CDR achieves greater recommendation performance improvements across settings ranging from games (dense) to toys (sparse) and movies (dense) to books (sparse). This aligns with the fundamental assumptions of cross-domain recommendation and also confirms the practical significance of Tri-CDR in cross-domain positive knowledge transfer. Therefore, this application's embodiments can be applied to specific scenarios, namely, recommending target objects with sparse consumption behavior under the video tab in the "Watch" section. By utilizing the rich consumption behavior information of target objects in video scenarios, this application assists in the recommendation process of target objects with sparse interactive behavior in video tab scenarios.

[0240] First, define the source behavior sequence and the target behavior sequence in the source domain and the target domain, respectively. and Where p and q represent the number of behaviors of the target object in the source domain and the target domain, respectively. and These represent behavioral embeddings. In Tri-CDR, a hybrid behavioral sequence of the target object is also proposed. As a supplement to the target object's behavior sequences in the source and target domains, it is a complete sequence of target object behavior, containing all behaviors of the target object in the source and target domains and arranged strictly in chronological order. Given three behavior sequences S s S T and S M Tri-CDR attempts to recommend the next interactive item for the target object in the target domain.

[0241] The following describes the model framework proposed in the embodiments of this application—a triadic sequence learning method (Tri-CDR) for cross-domain recommendation. This framework jointly models the behavioral sequences of the target object in the source domain, target domain, and mixed domain to improve cross-domain recommendation performance. See also Figure 10 , Figure 10This is the overall structure of Tri-CDR. Specifically, firstly, three sequence encoders are used to model the behavioral sequences of the target object in the source domain, target domain, and mixed domain, respectively, thereby generating corresponding latent representations. Then, a ternary cross-domain attention strategy (TCA) is proposed. When learning the sequence representations of the target object in the three domains, knowledge related to the target object's target domain preferences and global interests is highlighted, thereby mitigating negative knowledge transfer. To better simulate the correlation between multi-domain sequence representations, a ternary cross-domain contrastive learning mechanism (TCL) is designed, which includes two contrastive constraints: (a) coarse-grained similarity modeling, which makes the source domain, target domain, and mixed domain sequence representations from the same target object more similar; (b) fine-grained distance modeling, which considers the diversity and relevance of multiple domains, and maintains the information gain of the source domain and mixed domain sequences relative to the target domain sequence by controlling the distance between the ternary sequences, thereby learning better multi-domain representations to capture the target object's preferences.

[0242] In some embodiments, inspired by the self-attention mechanism in sequence recommendation, SASRec is used as the sequence encoder for Tri-CDR. For the target domain sequence S... T Establish the input matrix Each behavior is embedded It consists of a learnable item index embedding and a position embedding. Then, the sequence encoder will... T Linear projection is a method for feeding queries, keys, and values ​​into attention calculation, and its definition is as follows:

[0243]

[0244] Where Q = D T W Q K=D T W K V=D T W V W Q W K W V These represent the linear mapping layers, and the implicit matrix H of the target domain is also obtained through a point-based feedforward network. T Its definition is as follows:

[0245]

[0246] Where w1, w2, b1, and b2 represent the weight matrix and bias vector, respectively. The implicit matrix H of the target domain... T Similarly, the implicit matrix H of the target object's behavioral sequence in the source and mixed domains s H MIt is also built based on the domain-specific behavioral interactions of the target object. It is worth noting that when learning sequence representations, even the same items in different domains are assigned different item index embeddings. This avoids overly homogeneous representations across different domains, thereby achieving better representation capabilities.

[0247] The ternary cross-domain attention strategy is based on the fundamental assumption that different historical behaviors of a target object in three domains should have varying importance for that target object's next interaction item in the target domain. More precisely, it aims to ensure that the information emphasized in the three behavioral sequences of the target object is related both to the target object's preferences in the target domain and to its global interests, so that the model can complete the target domain prediction task based on a comprehensive understanding of the target object's preferences. Therefore, in the three implicit behavior matrices H... T H s and H M The above uses a ternary cross-domain attention strategy (TCA) to learn three sequence representations s from the source domain, target domain, and mixed domain. T s S and s M This enables positive knowledge transfer.

[0248] Specifically, for the source domain, given the implicit behavior matrix and target projects TCA calculates the implicit embedding of the target object in the i-th behavior of the source domain. Attention weight The specific formula is as follows:

[0249]

[0250] Where ⊙ denotes the vector dot product operation, ||.|| denotes the concatenation operation, and MLP... a (.) is a two-layer fully connected network, with each layer followed by a PReLU activation function. and These represent the implicit embeddings of the target object's last behavior in the target domain and the hybrid domain, respectively. It's worth noting that, for efficiency reasons, the embedding of the target item was not used. Instead, it utilizes This represents the target object's preference in the target domain. Furthermore, it is assumed that the hidden embedding of the target object's last behavior in the mixed domain is closely related to the target object's global interest, and this is emphasized in the cross-domain attention computation strategy. The significance of this is that, based on the aforementioned attention calculation method, the implicit embedding of the target object's behavior in all the source domains is considered. The sequence representation ss of the source domain is obtained by summarizing:

[0251]

[0252] Sequence representations of the source domain s s The computational methods are consistent, and the sequence representations of the mixed domain and the target domain are... M and s T The expressions are as follows:

[0253]

[0254]

[0255] The key to conventional cross-domain recommendation research lies in capturing the inter-domain correlation between source and target domain behaviors. Tri-CDR, however, faces a more challenging task: comprehensively understanding the triple correlation between the target object's behavioral sequences across the source, target, and mixed domains. Therefore, this application proposes a novel ternary contrastive learning mechanism (TCL) to comprehensively establish the correlation between the three sequence representations. Specifically, the TCL incorporates two CL tasks: coarse-grained similarity modeling and fine-grained distance modeling. The former CL task captures the coarse-grained similarity between any two sequence representations of the same target object and the sequence representations of other target objects; the latter CL task performs fine-grained modeling of the distances between the multi-domain sequence representations of the target object, thus preserving information diversity.

[0256] The behavioral sequences of a target object in different domains naturally exhibit certain potential similarities. Therefore, a coarse-grained similarity modeling method (CSM) is designed to model the coarse-grained similarity between the sequence representations of the same target object across three domains. Specifically, the sequence representations s T s s and s M Linear projector P through a specific domain S (.), P T (.) and P M (.) Project onto the corresponding feature space. This yields the projection sequence representation. and Then, the contrastive loss LCL is calculated assuming any two of them are positive examples. and For example, following the classic InfoNCE approach, the contrast loss is defined as:

[0257]

[0258] Where B represents the batch obtained from sampling, j∈B\i represents the negative sample in B corresponding to i, and τ represents the temperature coefficient. This represents the cosine similarity.

[0259] Finally, CSM lost L CSM It can be defined as:

[0260]

[0261] Where λ1, λ2 and λ3 represent the loss weights.

[0262] CSM assumes that preferences for the same target object should be more similar across different domains, and monotonically narrows the multivariate representations of the same target object. However, subtle differences do exist between the preferences of the target object across different domains, leading to over-optimization of L... CSM This will cause the model to collapse to an approximation point, meaning that the source domain behavior sequence and the mixed domain behavior sequence of the target object are unlikely to provide additional information gain for the overly similar target domain behavior sequence, which will instead affect the model's ternary sequence modeling process.

[0263] To address this issue, the specific composition of the target object's hybrid domain behavior sequence is considered, encompassing all behavior sequences of the target object in both the source and target domains. Therefore, the distances between the source domain sequence representation and the hybrid domain sequence representation, and between the target domain sequence representation and the hybrid domain sequence representation, should both be smaller than the distance between the source domain sequence representation and the target domain sequence representation. Fine-grained distance modeling (FDM) is proposed based on a triplet loss based on marginal distance, taking into account... and The distance between them should be at least greater than and The distance between them is the large distance parameter γ. Formally, the loss L of FDM is... FDM Defined as:

[0264]

[0265] in Measured by L2 distance and The similarity between them. γ represents the distance parameter to be controlled by the ternary loss. Through joint optimization of L... CSM and L FDM Tri-CDR can find a balance that allows the multi-domain preferences of the target object to become more consistent while maintaining subtle differences, thereby learning more information gains from other domains.

[0266] Following TCA and TCL, Tri-CDR represents the ternary sequence s. T s s and s M The concatenated sequences are fed into a multivariate sequence representation aggregation layer to generate the final target object representation u. T The multivariate sequence characterization aggregation layer can be defined as:

[0267] u T =MLP f(s M ||S S ||s T (28)

[0268] MLP f (.) indicates a two-layer fully connected network including a Leaky ReLU activation layer in the middle. Finally, based on the final target object representation uT and item embedding... To calculate the target object u and the project The reliability of predictions between The binary cross-entropy loss L CTR The statement is as follows:

[0269]

[0270] Among them, R T It is the training set for the target domain. and These represent positive samples and corresponding randomly sampled negative samples, respectively. Let L represent the predicted probability of (u, d). To optimize ternary cross-domain sequence recommendation in conjunction with the CL task, this embodiment defines the overall objective function L as L0. CTR L CSM and L FDM Linear combination:

[0271] L = L CTR +λ CsM L CsM +λ FDM L FDM (30)

[0272] Where, λ CSM and λ FDM L CSM and L FDM The weight.

[0273] The embodiments of this application comprehensively analyze the effectiveness and universality of the proposed Tri-CDR framework through various experiments, including comparison with existing SOTA algorithms, ablation experiments, universality tests, parameter analysis, and visualization. These will be described in turn below.

[0274] See Table 1 below. Table 1 is a comparison table of experimental results for cross-domain recommendation provided in the embodiments of this application.

[0275]

[0276] In some embodiments, as shown in Table 1 above, compared with related technologies 1 and 2, the embodiments of this application compare the Tri-CDR framework with existing single-domain sequence recommendation algorithms (related technology 1) and cross-domain sequence recommendation algorithms (related technology 2), and the results are shown in Table 1 above. First, it is found that compared with related technologies, Tri-CDR achieves the best performance on all datasets and evaluation metrics, which proves the effectiveness of the TCL proposed in this application embodiment in the CDR task. In addition, it also shows that Tri-CDR can accurately model the correlation between the source, target, and mixed behavior sequences of the target object, and successfully capture useful information related to the prediction of the target domain from all domains. Second, it is found that Tri-CDR outperforms all existing cross-domain sequence recommendation methods, which confirms the importance of incorporating mixed behavior sequences containing the global behavior patterns of the target object into the cross-domain modeling process and modeling the coarse-grained similarity and fine-grained differences between the multi-domain preferences of the target object. This also means that compared with traditional CDR methods, the proposed TCA and TCL can perform better positive knowledge transfer. Finally, by comparing the performance improvement ratios of different cross-domain settings, it was found that Tri-CDR is more advantageous in settings from Amazon Games to Amazon Toys and from Amazon Movies to Amazon Books. This reflects that Tri-CDR performs better when transferring learning from a relatively dense source domain to a sparse target domain, reflecting the basic assumptions of cross-domain recommendation and confirming the application significance of Tri-CDR's positive cross-domain knowledge transfer.

[0277] See Table 2 below, which is a comparison table of the results of the cross-domain recommended ablation experiments provided in the embodiments of this application.

[0278] Table 2 Comparison of ablation experiment results for cross-domain recommendations

[0279]

[0280] In some embodiments, for ablation experiments, this application compares and analyzes the effectiveness of different domains and components in Tri-CDR, and the results are shown in Table 2 above. Here, S, T, and M represent the information of the source domain, target domain, and hybrid domain, respectively. The practical significance of the TCA module and FDM module proposed in this application is analyzed through Tri-CDR without TCA and Tri-CDR without FDM. First, by comparing the results of SASRec(T), SASRec(M), and SASRec(S+T) under various cross-domain settings, it is found that SASRec(S+T) and SASRec(M) outperform SASRec(T) in most metrics, demonstrating the importance of source domain information in CDR tasks. Simultaneously, it is found that SASRec(M), which performs simple modeling of the hybrid sequence of the target object, can also achieve better results than SASRec(S+T) in some cases, implying the potential superiority of hybrid domain sequences. Secondly, it was found that SASRec(S+T+M) does not always outperform the three experimental versions mentioned above. This confirms that the mixed sequences of the target object not only contain useful information but also some noise information, and simply jointly modeling the behavioral information of the three domains is not necessarily feasible. Thirdly, it was found that Tri-CDR w / o TCA significantly outperforms SASRec(S+T+M) (i.e., Tri-CDR w / o TCA & TCL), which proves the effectiveness of TCL. Ternary contrastive learning can extend the consistency of CL from a single domain to a ternary domain, which is beneficial to the multi-domain preference learning process of the target object. Fourthly, it was observed that Tri-CDR further improves the performance of Tri-CDR w / o TCA. This indicates that TCA can capture information related to the item to be predicted and the overall interest of the target object through a ternary cross-domain attention strategy, thereby achieving positive knowledge transfer. Finally, through the performance comparison of Tri-CDR and Tri-CDR w / o FDM, the importance of fine-grained distance modeling in TCL is further discovered. It can not only help Tri-CDR learn better multi-domain sequence representations but also make the model more robust and stable under different parameters.

[0281] In some embodiments, for general experiments, see [link to relevant documentation]. Figure 12 , Figure 12 This is a comparison chart of the effects of a general experiment provided in the embodiments of this application. In addition to SASRec, the effectiveness of the Tri-CDR framework was also verified on GRU4Rec, and the experimental results are as follows. Figure 5As shown in the diagram. First, Tri-CDR achieves consistent and significant improvements over the base model GR U4Rec across all cross-domain settings and metrics, demonstrating the universality of the Tri-CDR framework. Second, similar to the conclusion in Section 4.2, while mixed sequences include much positive information beneficial to the prediction of target items, they also add noise. Simple fusion modeling may actually lead to a performance decrease. However, the TCA and TCL proposed in this application can help model the correlation between ternary sequences with fine granularity, thereby achieving a consistent improvement in recommendation performance. Finally, the performance of Tri-CDR using the current state-of-the-art single-domain sequence recommendation model CL4SRec as the sequence encoder is evaluated. Compared to the original CL4SRec, Tri-CDR achieves improvements of 4.99% and 7.56% on the NDCG@10 metric and 4.41% and 6.14% on the HR@10 metric on the Game and Toy datasets, further demonstrating the universality of the Tri-CDR framework.

[0282] In some embodiments, for parameter experiments, see [link to documentation]. Figure 13 , Figure 13 This is a schematic diagram of the results of the parameter experiment provided in the embodiments of this application. Parameter analysis was performed under the cross-domain setting of Amazon Game→Amazon Toy to study the influence of different loss weights λCSM, λFDM, and distance parameter γ on the Tri-CDR results. The results are as follows: Figure 13 As shown in the figure, firstly, it was found that the performance of Tri-CDR first increases and then decreases with the increase of λCSM, achieving the best effect at λCSM = 0.1. Secondly, it was observed that excessively small or large λFDM may interfere with the modeling process of ternary sequence correlations in Tri-CDR, and Tri-CDR is not sensitive to the weight λCSM of the fine-grained distance modeling loss LFDM. Finally, by setting the fine-grained distance, γ = 0 indicates that the model only needs to satisfy the condition that the distance between the source domain representation and the mixed domain representation is less than the distance between the domain representation and the target domain representation, while γ = 100 indicates that the model wants the distance between them to be as large as possible. Tri-CDR performs relatively poorly at these two extreme values, thus demonstrating the importance of setting an appropriate distance γ in FDM.

[0283] For multi-domain representation visualization, see [link / reference]. Figure 14 , Figure 14 This is a schematic diagram illustrating the results of multi-domain representation visualization provided in this application embodiment. To intuitively demonstrate the correlation between ternary sequence representations in this application embodiment, several target objects were selected, and... Figure 14The image shows the visualization of the ternary sequence representations of these target objects in SASRec(S+T+M), Tri-CDRw / oFDM, and Tri-CDR. Figure 14 The first line focuses on the representation distribution of different domains, while the second line uses different colors to emphasize the representation distribution of different target objects. Firstly, through... Figure 14 A comparison of the three figures in the first row reveals that in SASRec(S+T+M), the ternary sequence representations of most target objects are based on natural domain clustering, and the distances between the source sequence representation, mixed sequence representation, and target sequence representation of the same target object are relatively large, making it difficult to fully utilize the information from the source and mixed domains. Secondly, it is observed that the ternary sequence representations in figures (e) and (f) are naturally clustered based on the target objects, indicating that TCL's coarse-grained similarity modeling can make the ternary sequence representations of the same target object similar. Finally, it is also found that in figure (e), the ternary sequence representations of some target objects form small acute-angled triangles in the feature space. These overly homogeneous ternary representations may weaken the additional information gain from the source and mixed domains, thereby reducing the model's cross-domain knowledge transfer ability. In contrast, the ternary representations corresponding to the target objects in figure (f) are more discriminative, constructing reasonable obtuse-angled triangles, which, with the help of FDM, improves the cross-domain recommendation performance of Tri-CDR.

[0284] It is understood that in the embodiments of this application, data related to historical behavior sequences are involved. When the embodiments of this application are applied to specific products or technologies, permission or consent from the target object is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0285] The following description continues to illustrate the exemplary structure of the information recommendation device 455 provided in the embodiments of this application as a software module. In some embodiments, such as... Figure 3As shown, the software modules stored in the information recommendation device 455 in the memory 450 may include: a fusion module 4551, used to fuse a first historical behavior sequence of the target object in the target domain and a second historical behavior sequence of the target object in the source domain to obtain a fused behavior sequence; a first determination module 4552, used to determine a first sequence feature of the first historical behavior sequence and determine a fused sequence feature of the fused behavior sequence; a second determination module 4553, used to determine a second sequence feature of the second historical behavior sequence based on the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence; a prediction module 4554, used to predict the object features of the target object in the target domain based on the first sequence features, the second sequence features, and the fused sequence features; and an information recommendation module 4555, used to combine the object features and the information features of each piece of information to be recommended in the target domain to recommend information about the target object in the target domain.

[0286] In some embodiments, the fusion module 4551 is further configured to obtain each first sequence element in the first historical behavior sequence and the timestamp of each first sequence element; obtain each second sequence element in the second historical behavior sequence and the timestamp of each second sequence element; sort each first sequence element and each second sequence element according to the order of the timestamps, and determine the sorted first sequence elements and second sequence elements as the fused behavior sequence.

[0287] In some embodiments, the first determining module 4552 is further configured to obtain a first encoding feature of the first historical behavior sequence and a fusion encoding feature of the fusion behavior sequence, wherein the first encoding feature includes a plurality of first sub-features, and the first sub-features correspond one-to-one with the first sequence elements in the first historical behavior sequence; based on the first encoding feature and the fusion encoding feature, determine the weight of each first sub-feature in the first encoding feature; and perform a weighted summation of each first sub-feature according to the weight of each first sub-feature to obtain the first sequence feature.

[0288] In some embodiments, the first determining module 4552 is further configured to perform the following processing on each of the first sub-features in the first coding feature: obtaining a target first sub-feature in the first coding feature and a target fusion sub-feature in the fusion coding feature; wherein the fusion coding feature includes multiple fusion sub-features, the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; determining a reference sub-feature corresponding to the first sub-feature based on the first sub-feature, the target first sub-feature, and the target fusion sub-feature; and performing normalization processing on the reference sub-feature corresponding to the first sub-feature to obtain the weight of the first sub-feature.

[0289] In some embodiments, the first determining module 4552 is further configured to multiply the norm of the first sub-feature, the target first sub-feature, and the first sub-feature to obtain a first multiplication result; determine the inner product of the first sub-feature and the target first sub-feature, and determine the norm of the inner product; multiply the norm of the inner product, the target fused sub-feature, and the target first sub-feature to obtain a second multiplication result; and subtract the first multiplication result from the second multiplication result to obtain a reference sub-feature corresponding to the first sub-feature.

[0290] In some embodiments, the first determining module 4552 is further configured to obtain a first encoding feature of the first historical behavior sequence and a fusion encoding feature of the fusion behavior sequence, wherein the fusion encoding feature includes a plurality of fusion sub-features, and the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence; based on the first encoding feature and the fusion encoding feature, determine the weight of each fusion sub-feature in the fusion encoding feature; and perform a weighted summation of each fusion sub-feature according to its weight to obtain the fusion sequence feature.

[0291] In some embodiments, the first determining module 4552 is further configured to perform the following processing on each of the fusion sub-features in the fusion coding feature: obtaining a target first sub-feature in the first coding feature and a target fusion sub-feature in the fusion coding feature; wherein, the first coding feature includes a plurality of first sub-features, the first sub-features correspond one-to-one with the first sequence elements in the first historical behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; determining a reference sub-feature corresponding to the fusion sub-feature based on the fusion sub-feature, the target first sub-feature, and the target fusion sub-feature; and performing normalization processing on the reference sub-feature corresponding to the fusion sub-feature to obtain the weight of the fusion sub-feature.

[0292] In some embodiments, the second determining module 4553 is further configured to obtain a first coding feature of the first historical behavior sequence, a fusion coding feature of the fusion behavior sequence, and a second coding feature of the second historical behavior sequence, wherein the second coding feature includes a plurality of second sub-features, and the second sub-features correspond one-to-one with the second sequence elements in the second historical behavior sequence; based on the first coding feature, the second coding feature, and the fusion coding feature, determine the weight of each second sub-feature in the second coding feature; and perform a weighted summation of each second sub-feature according to its weight to obtain the second sequence feature.

[0293] In some embodiments, the second determining module 4553 is further configured to perform the following processing on each of the second sub-features in the second coding feature: obtaining a target first sub-feature in the first coding feature and a target fusion sub-feature in the fusion coding feature; wherein the fusion coding feature includes multiple fusion sub-features, the fusion sub-features correspond one-to-one with the fusion sequence elements in the fusion behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion historical behavior sequence; determining a reference sub-feature corresponding to the second sub-feature based on the second sub-feature, the target first sub-feature, and the target fusion sub-feature; and performing normalization processing on the reference sub-feature corresponding to the second sub-feature to obtain the weight of the second sub-feature.

[0294] In some embodiments, the information recommendation module 4555 is further configured to obtain information features of each piece of information to be recommended in the target domain; compare the object features with each of the information features to obtain the credibility of each information feature; determine the information to be recommended corresponding to the highest credibility as the target recommendation information for the target object in the target domain; and send the target recommendation information to the terminal corresponding to the target object.

[0295] In some embodiments, the target information recommendation model includes a first target encoding layer, a second target encoding layer, a third target encoding layer, and a target prediction layer; the first determining module 4552 is further configured to call the first target encoding layer to encode the first historical behavior sequence to obtain the first sequence feature; the first determining module 4552 is further configured to call the second target encoding layer to encode the fused behavior sequence to obtain the fused sequence feature; the second determining module 4553 is further configured to call the third target encoding layer to encode the second historical behavior sequence by combining the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence to obtain the second sequence feature; the prediction module 4554 is further configured to call the target prediction layer to perform feature prediction on the target object by combining the first sequence feature, the second sequence feature, and the fused sequence feature to obtain the object feature of the target object in the target domain.

[0296] In some embodiments, the information recommendation device 455 further includes: a training module, configured to acquire an information recommendation model, the information recommendation model including a first encoding layer, a second encoding layer, a third encoding layer, and a prediction layer; to fuse a first historical behavior sample sequence of a sample object in the target domain and a second historical behavior sample sequence of the sample object in the source domain to obtain a fused behavior sample sequence; to call the first encoding layer to encode the first historical behavior sample sequence to obtain a first sequence sample feature, and to call the second encoding layer to encode the fused behavior sample sequence to obtain a fused sequence sample feature; and to call the third encoding layer to combine the first historical behavior sample sequence, the second historical behavior sample sequence, and the fused behavior sample sequence to obtain a second historical behavior sample feature. The sample sequence is encoded to obtain a second sequence sample feature; a prediction network is invoked, and the first sequence sample feature, the second sequence sample feature, and the fused sequence sample feature are combined to predict the features of the target object, obtaining the object sample feature of the sample object in the target domain; a first loss value is determined by combining the object sample feature and the information features of each piece of information to be recommended in the target domain; a second loss value and a third loss value are determined based on the first sequence sample feature, the fused sequence sample feature, and the second sequence sample feature; the first loss value, the second loss value, and the third loss value are weighted and summed to obtain a target loss value; the model parameters of the information recommendation model are updated based on the target loss value to obtain the target information recommendation model.

[0297] In some embodiments, the training module described above is further configured to determine the credibility between the object sample features and each of the information features; and to perform the following processing for each credibility: multiplying the credibility by the logarithm of the credibility to obtain a first multiplication result; determining the difference between the credibility and 1 as a target difference; multiplying the target difference by the logarithm of the target difference to obtain a second multiplication result; adding the first multiplication result and the second multiplication result to obtain a credibility sum; adding the credibility sums of each credibility to obtain a sum result, and determining the negative of the sum result as the first loss value.

[0298] In some embodiments, the training module is further configured to determine a first contrast loss value between the first sequence sample features and the fused sequence sample features; and determine a second contrast loss value between the first sequence sample features and the second sequence sample features; and determine a third contrast loss value between the fused sequence sample features and the second sequence sample features; and perform a weighted summation of the first contrast loss value, the second contrast loss value, and the third contrast loss value to obtain the second loss value.

[0299] In some embodiments, the training module is further configured to determine a first feature distance between the first sequence sample features and the second sequence sample features, and to determine a second feature distance between the fused sequence sample features and the second sequence sample features; and to determine a third loss value based on the difference between the first feature distance and the second feature distance.

[0300] This application provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. The processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the information recommendation method described in this application.

[0301] This application provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are executed by a processor, they cause the processor to execute the information recommendation method provided in this application, for example... Figure 3 The information recommendation method is shown.

[0302] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEP ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of electronic devices including one or any combination of the above-mentioned memories.

[0303] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, 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 stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0304] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, 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 co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0305] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0306] In summary, the embodiments of this application have the following beneficial effects:

[0307] (1) By fusing the first historical behavior sequence and the second historical behavior sequence, a fused behavior sequence is obtained. The first sequence feature corresponding to the first historical behavior sequence, the second sequence feature corresponding to the second historical behavior sequence, and the fused sequence feature corresponding to the fused behavior sequence are determined. Based on the first sequence feature, the second sequence feature, and the fused sequence feature, the object features of the target object in the target domain are predicted. Combining the object features and the information features of each piece of information to be recommended in the target domain, information recommendation is performed on the target object in the target domain. In this way, since the fused behavior sequence fully integrates the historical behavior of the target object in the source domain and the target domain, when predicting the object features of the target object in the target domain, the historical behavior of the target object in the source domain and the target domain is fully considered, thereby effectively improving the accuracy of the determined object features. Thus, when recommending information on the target object in the target domain based on the object features, the accuracy of information recommendation is effectively improved.

[0308] (2) By fusing the first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain, a fused behavior sequence is obtained. Through the fused behavior sequence, target recommendation information for the target object in the target domain is predicted. Thus, in the prediction process, the historical behavior of the target object in the source domain and the target domain is fully combined, which effectively improves the accuracy of the predicted target recommendation information.

[0309] (3) By determining the weight of each first sub-feature in the first coding feature based on the first coding feature and the fused coding feature, the fused coding feature is fully considered in the process of determining the weight of each first sub-feature in the first coding feature. Since the fused coding feature can more comprehensively reflect the historical behavior of the target object in the source domain and the target domain, the weight of each first sub-feature is more accurate, which effectively improves the accuracy of the weight of the first sub-feature. Then, according to the weight of each first sub-feature, the first sub-feature is weighted and summed to obtain the first sequence feature, which effectively improves the accuracy of the determined first sequence feature.

[0310] (4) By determining the weight of each fusion sub-feature in the fusion coding feature based on the first coding feature and the fusion coding feature, the fusion coding feature is fully considered in the process of determining the weight of each fusion sub-feature in the fusion coding feature. Since the fusion coding feature can more comprehensively reflect the historical behavior of the target object in the source domain and the target domain, the weight of each fusion sub-feature is more accurate, which effectively improves the accuracy of the weight of the determined fusion sub-feature. Then, according to the weight of each fusion sub-feature, the fusion sub-feature is weighted and summed to obtain the fusion sequence feature, which effectively improves the accuracy of the determined fusion sequence feature.

[0311] (5) By determining the weights of each second sub-feature in the second coding feature based on the first coding feature, the fused coding feature, and the second coding feature, the fused coding feature is fully considered in the process of determining the weights of each second sub-feature in the second coding feature. Since the fused coding feature can more comprehensively reflect the historical behavior of the target object in the source domain and the target domain, the weights of each determined second sub-feature are more accurate, which effectively improves the accuracy of the determined weights of the second sub-features. Then, according to the weights of each second sub-feature, the second sub-features are weighted and summed to obtain the second sequence feature, which effectively improves the accuracy of the determined second sequence feature.

[0312] (6) By comparing the accurate object features and the information features of each information to be recommended, the target recommendation information that best matches the target object is accurately determined, and the target recommendation information is sent to the terminal corresponding to the target object, thereby effectively improving the recommendation accuracy.

[0313] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. An information recommendation method, characterized in that, The method includes: The first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain are fused to obtain a fused behavior sequence; wherein the historical behavior corresponding to the first historical behavior sequence and the second historical behavior sequence includes one of the following: purchase, click to listen to audio, click to watch video; Determine the first sequence features of the first historical behavior sequence, and determine the fusion sequence features of the fused behavior sequence; Based on the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence, the second sequence feature of the second historical behavior sequence is determined; The step of determining the second sequence feature of the second historical behavior sequence based on the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence includes: obtaining the first encoding feature of the first historical behavior sequence, the fused encoding feature of the fused behavior sequence, and the second encoding feature of the second historical behavior sequence, wherein the second encoding feature includes multiple second sub-features, and the second sub-features correspond one-to-one with the second sequence elements in the second historical behavior sequence; determining the weight of each second sub-feature in the second encoding feature based on the first encoding feature, the second encoding feature, and the fused encoding feature; and performing a weighted summation of each second sub-feature according to its weight to obtain the second sequence feature. Based on the first sequence features, the second sequence features, and the fused sequence features, predict the object features of the target object in the target domain; By combining the object features and the information features of each piece of information to be recommended in the target domain, information recommendation is performed on the target object within the target domain.

2. The method according to claim 1, characterized in that, The process of fusing the first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain to obtain a fused behavior sequence includes: Obtain each first sequence element in the first historical behavior sequence, and the timestamp of each first sequence element; Obtain each second sequence element in the second historical behavior sequence, and the timestamp of each second sequence element; According to the order of the timestamps, each of the first sequence elements and each of the second sequence elements are sorted, and the sorted first sequence elements and second sequence elements are determined as the fusion behavior sequence.

3. The method according to claim 1, characterized in that, The first sequence feature for determining the first historical behavior sequence includes: Obtain the first encoding feature of the first historical behavior sequence and the fusion encoding feature of the fusion behavior sequence, wherein the first encoding feature includes a plurality of first sub-features, and the first sub-features correspond one-to-one with the first sequence elements in the first historical behavior sequence; Based on the first coding feature and the fused coding feature, determine the weight of each first sub-feature in the first coding feature; The first sequence features are obtained by weighting and summing the first sub-features according to their respective weights.

4. The method according to claim 3, characterized in that, The step of determining the weight of each first sub-feature in the first coding feature based on the first coding feature and the fused coding feature includes: For each of the first sub-features in the first encoded feature, the following processing is performed: Obtain the target first sub-feature in the first encoded feature and the target fusion sub-feature in the fused encoded feature; The fusion coding feature includes multiple fusion sub-features, each of which corresponds one-to-one with a fusion sequence element in the fusion behavior sequence. The target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion behavior sequence. Based on the first sub-feature, the target first sub-feature, and the target fused sub-feature, a reference sub-feature corresponding to the first sub-feature is determined; The reference sub-feature corresponding to the first sub-feature is normalized to obtain the weight of the first sub-feature.

5. The method according to claim 4, characterized in that, The step of determining the reference sub-feature corresponding to the first sub-feature based on the first sub-feature, the target first sub-feature, and the target fused sub-feature includes: Multiply the norm of the first sub-feature, the target first sub-feature, and the first sub-feature together to obtain the first multiplication result; Determine the inner product of the first sub-feature and the target first sub-feature, and determine the norm of the inner product; The norm of the inner product, the target fusion sub-feature, and the target first sub-feature are multiplied together to obtain the second multiplication result; Subtract the first multiplication result from the second multiplication result to obtain the reference sub-feature corresponding to the first sub-feature.

6. The method according to claim 1, characterized in that, The determination of the fusion sequence features of the fusion behavior sequence includes: Obtain the first encoding feature of the first historical behavior sequence and the fusion encoding feature of the fused behavior sequence, wherein the fusion encoding feature includes multiple fusion sub-features, and the fusion sub-features correspond one-to-one with the fusion sequence elements in the fused behavior sequence; Based on the first coding feature and the fused coding feature, determine the weight of each fused sub-feature in the fused coding feature; The fusion sequence features are obtained by weighting and summing the fusion sub-features according to their respective weights.

7. The method according to claim 6, characterized in that, The step of determining the weights of each fused sub-feature in the fused coding feature based on the first coding feature and the fused coding feature includes: The following processing is performed on each of the fused sub-features in the fused coding features: Obtain the target first sub-feature in the first encoded feature and the target fusion sub-feature in the fused encoded feature; Wherein, the first encoding feature includes a plurality of first sub-features, each of which corresponds one-to-one with a first sequence element in the first historical behavior sequence, the target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion behavior sequence; Based on the fusion sub-feature, the target first sub-feature, and the target fusion sub-feature, a reference sub-feature corresponding to the fusion sub-feature is determined; The reference sub-features corresponding to the fused sub-features are normalized to obtain the weights of the fused sub-features.

8. The method according to claim 1, characterized in that, The step of determining the weight of each second sub-feature in the second coding feature based on the first coding feature, the second coding feature, and the fused coding feature includes: For each of the second sub-features in the second encoded feature, the following processing is performed: Obtain the target first sub-feature in the first encoded feature and the target fusion sub-feature in the fused encoded feature; The fusion coding feature includes multiple fusion sub-features, each of which corresponds one-to-one with a fusion sequence element in the fusion behavior sequence. The target first sub-feature is the first sub-feature corresponding to the last first sequence element in the first historical behavior sequence, and the target fusion sub-feature is the fusion sub-feature corresponding to the last fusion sequence element in the fusion behavior sequence. Based on the second sub-feature, the first sub-feature of the target, and the fused sub-feature of the target, a reference sub-feature corresponding to the second sub-feature is determined; The reference sub-feature corresponding to the second sub-feature is normalized to obtain the weight of the second sub-feature.

9. The method according to claim 1, characterized in that, The step of combining the object features and the information features of each piece of information to be recommended in the target domain to recommend information about the target object includes: Obtain the information features of each piece of information to be recommended in the target domain; The object features are compared with each of the information features to obtain the credibility of each information feature; The recommendation information corresponding to the highest credibility is determined as the target recommendation information for the target object under the target domain; The target recommendation information is sent to the terminal corresponding to the target object.

10. The method according to claim 1, characterized in that, The information recommendation method is implemented based on a target information recommendation model, which includes a first target encoding layer, a second target encoding layer, a third target encoding layer, and a target prediction layer. The first sequence feature for determining the first historical behavior sequence includes: The first target encoding layer is invoked to encode the first historical behavior sequence to obtain the first sequence features; The determination of the fusion sequence features of the fusion behavior sequence includes: The second target encoding layer is invoked to encode the fusion behavior sequence, thereby obtaining the fusion sequence features; The step of determining the second sequence feature of the second historical behavior sequence based on the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence includes: The third target encoding layer is invoked to encode the second historical behavior sequence by combining the first historical behavior sequence, the second historical behavior sequence, and the fused behavior sequence, thereby obtaining the second sequence feature; The step of predicting the object features of the target object in the target domain based on the first sequence features, the second sequence features, and the fused sequence features includes: The target prediction layer is invoked to combine the first sequence features, the second sequence features, and the fused sequence features to perform feature prediction on the target object, thereby obtaining the object features of the target object in the target domain.

11. The method according to claim 10, characterized in that, Before determining the first sequence feature of the first historical behavior sequence, the method further includes: An information recommendation model is obtained, which includes a first encoding layer, a second encoding layer, a third encoding layer, and a prediction layer; The first historical behavior sample sequence of the sample object in the target domain and the second historical behavior sample sequence of the sample object in the source domain are fused to obtain a fused behavior sample sequence. The first encoding layer is invoked to encode the first historical behavior sample sequence to obtain the first sequence sample feature; the second encoding layer is invoked to encode the fused behavior sample sequence to obtain the fused sequence sample feature; and the third encoding layer is invoked to combine the first historical behavior sample sequence, the second historical behavior sample sequence, and the fused behavior sample sequence to encode the second historical behavior sample sequence to obtain the second sequence sample feature. The prediction network is invoked to combine the features of the first sequence sample, the features of the second sequence sample, and the fused sequence sample features to perform feature prediction on the target object, thereby obtaining the object sample features of the sample object in the target domain; By combining the object sample features and the information features of each piece of information to be recommended in the target domain, a first loss value is determined; based on the first sequence sample features, the fused sequence sample features, and the second sequence sample features, a second loss value is determined, and a third loss value is determined. The first loss value, the second loss value, and the third loss value are weighted and summed to obtain the target loss value; Based on the target loss value, the model parameters of the information recommendation model are updated to obtain the target information recommendation model.

12. The method according to claim 11, characterized in that, The step of determining the first loss value by combining the features of the object sample and the information features of each piece of information to be recommended in the target domain includes: Determine the credibility between the object sample features and each of the information features; For each confidence level, the following processing is performed: multiply the confidence level by its logarithm to obtain a first multiplication result; determine the difference between the confidence level and 1 as a target difference; multiply the target difference by its logarithm to obtain a second multiplication result; add the first multiplication result and the second multiplication result to obtain a confidence sum. The credibility values ​​of each credibility level are summed together to obtain a summed result, and the negative number of the summed result is determined as the first loss value.

13. The method according to claim 11, characterized in that, The step of determining the second loss value based on the features of the first sequence sample, the features of the fused sequence sample, and the features of the second sequence sample includes: A first contrast loss value is determined between the first sequence sample features and the fused sequence sample features; a second contrast loss value is determined between the first sequence sample features and the second sequence sample features; and a third contrast loss value is determined between the fused sequence sample features and the second sequence sample features. The second loss value is obtained by weighted summing of the first contrast loss value, the second contrast loss value, and the third contrast loss value.

14. The method according to claim 11, characterized in that, Determining the third loss value includes: Determine a first feature distance between the first sequence sample features and the second sequence sample features, and determine a second feature distance between the fused sequence sample features and the second sequence sample features; The third loss value is determined based on the difference between the first feature distance and the second feature distance.

15. An information recommendation device, characterized in that, The device includes: The fusion module is used to fuse the first historical behavior sequence of the target object in the target domain and the second historical behavior sequence of the target object in the source domain to obtain a fused behavior sequence; wherein, the historical behavior corresponding to the first historical behavior sequence and the second historical behavior sequence includes one of the following: purchase, click to listen to audio, click to watch video; The first determining module is used to determine the first sequence feature of the first historical behavior sequence and to determine the fusion sequence feature of the fusion behavior sequence; The second module is determined. Obtain the first encoding feature of the first historical behavior sequence, the fusion encoding feature of the fusion behavior sequence, and the second encoding feature of the second historical behavior sequence, wherein the second encoding feature includes multiple second sub-features, and the second sub-features correspond one-to-one with the second sequence elements in the second historical behavior sequence; Based on the first coding feature, the second coding feature, and the fused coding feature, determine the weight of each second sub-feature in the second coding feature; According to the weight of each second sub-feature, the second sub-features are weighted and summed to obtain the second sequence features; The prediction module is used to predict the object features of the target object in the target domain based on the first sequence features, the second sequence features, and the fused sequence features; The information recommendation module is used to combine the object features and the information features of each piece of information to be recommended in the target domain to recommend information about the target object within the target domain.

16. An electronic device, characterized in that, The electronic device includes: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the information recommendation method according to any one of claims 1 to 14.

17. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed by a processor, they implement the information recommendation method according to any one of claims 1 to 14.

18. A computer program product comprising a computer program or computer-executable instructions, characterized in that, When the computer program or computer-executable instructions are executed by a processor, the information recommendation method according to any one of claims 1 to 14 is implemented.

Images