Cross-domain sequential recommendation method and device based on hypergraph convolutional neural network

Abstract

Embodiments of the present specification provide a cross-domain sequence recommendation method and device based on a hypergraph convolutional neural network, wherein the cross-domain sequence recommendation method based on the hypergraph convolutional neural network comprises: acquiring historical user interaction item time sequence of multiple domains; constructing local hypergraphs and global hypergraphs corresponding to different time periods, the local hypergraph being a hypergraph constructed based on interaction items in the historical user interaction item time sequence of the target domain within the corresponding time period, the global hypergraph being a hypergraph constructed based on interaction items in the historical user interaction item time sequence of the multiple domains within the corresponding time period, the target domain being any domain in the multiple domains, and any hyperedge in the local hypergraph and the global hypergraph representing an interaction item time sequence subsequence of interaction items of any user within the corresponding time period; inputting the local hypergraph and the global hypergraph into a cross-domain sequence recommendation model to predict a to-be-interacted item of the user. The item recommendation accuracy can be improved.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of recommendation technology, and in particular to a cross-domain sequence recommendation method and apparatus based on hypergraph convolutional neural networks. Background Technology

[0002] Sequence recommendation is one of the core tasks in the field of recommender systems. Its core objective is to capture users' dynamically evolving interests by deeply analyzing their historical behavior sequences, thereby providing accurate recommendations for their next item. However, when faced with scenarios where user behavior sequences are short or data is sparse, most existing sequence recommendation models still face significant challenges in accurately depicting users' personalized preferences.

[0003] Existing sequence recommendation schemes still suffer from inaccurate item recommendations. Therefore, improving the accuracy of item recommendations based on users' historical behavior sequences is a problem that needs to be solved. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a cross-domain sequence recommendation method based on hypergraph convolutional neural networks. One or more embodiments of this specification also relate to cross-domain sequence recommendation apparatus, recommendation system, computing device, computer-readable storage medium, and computer program product based on hypergraph convolutional neural networks, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a cross-domain sequence recommendation method based on a hypergraph convolutional neural network is provided, comprising:

[0006] Obtain a time-series sequence of historical user interaction items across multiple domains, where each time-series sequence includes interaction items corresponding to multiple users. Construct local and global hypergraphs for different time periods, where the local hypergraph is a hypergraph constructed based on interaction items within the time-series sequence of historical user interaction items in the target domain within the corresponding time period, and the global hypergraph is a hypergraph constructed based on interaction items within the time-series sequence of historical user interaction items in the multiple domains within the corresponding time period. The target domain is any one of the multiple domains, and any hyperedge in the local and global hypergraphs represents an interaction item composed of interaction items of any user within the corresponding time period. The model uses a time-series subsequence; it inputs the local hypergraph and the global hypergraph into a cross-domain sequence recommendation model to predict the user's expected interaction items; wherein, the cross-domain sequence recommendation model is used to extract collaboration information between and within the time-series subsequences of interaction items of different users based on the local hypergraph and the global hypergraph using a hypergraph convolutional network, to obtain the item features of the interaction items containing the collaboration information, to determine the user item interaction features of any user based on the item features corresponding to the interaction items in the time-series subsequences of interaction items in multiple time periods, and to predict the expected interaction items based on the user item interaction features.

[0007] According to a second aspect of the embodiments of this specification, a method for training a cross-domain sequence recommendation model is provided, comprising:

[0008] The training sample set is used to obtain time-series samples of user interaction items from multiple domains, along with corresponding label information. The time-series samples include interaction items for multiple users. Local and global hypergraphs are constructed for different time periods. The local hypergraph is a hypergraph constructed based on interaction items within the time-series samples of the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on interaction items within the time-series samples of the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local and global hypergraphs represents a time-series subsequence of interaction items for any user within the corresponding time period. The local and global hypergraphs are input into a preset model to predict the user's interaction items. The preset model is used to predict the user's interaction items using a hypergraph convolutional network based on local... The hypergraph and global hypergraph extract collaborative information between and within the time-series subsequences of interactive items for different users, obtaining item features of the interactive items that include the collaborative information. Based on the item features corresponding to the interactive items in the time-series subsequences of the interactive items for any user in multiple time periods, the user item interaction features are determined, and the item to be interacted is predicted based on the user item interaction features. The training loss is calculated using the item to be interacted and the label information. The parameters of the preset model are updated based on the training loss to obtain the preset model with updated parameters. If the training termination condition is not met, the steps of obtaining user interaction item time-series samples from multiple domains and the label information corresponding to the user interaction item time-series samples from the training sample set are continued. When the training termination condition is met, the current preset model with updated parameters is determined as a cross-domain sequence recommendation model.

[0009] According to a third aspect of the embodiments of this specification, a cross-domain sequence recommendation device based on a hypergraph convolutional neural network is provided, comprising:

[0010] The sequence acquisition module is used to acquire the time-series sequence of historical user interaction items in multiple domains, wherein the time-series sequence of historical user interaction items includes interaction items corresponding to multiple users;

[0011] The hypergraph construction module is used to construct local hypergraphs and global hypergraphs corresponding to different time periods. The local hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local hypergraph and the global hypergraph represents the time sequence subsequence of interaction items formed by the interaction items of any user in the corresponding time period.

[0012] The project prediction module is used to input the local hypergraph and global hypergraph into the cross-domain sequence recommendation model to predict the user's interactive items;

[0013] The cross-domain sequence recommendation model is used to extract collaboration information between and within the time-series subsequences of interaction items of different users based on local and global hypergraphs using a hypergraph convolutional network, thereby obtaining the item features of the interaction items containing the collaboration information. Based on the item features corresponding to the interaction items of any user in the time-series subsequences of the interaction items in multiple time periods, the user item interaction features of the user are determined, and the item to be interacted is predicted based on the user item interaction features.

[0014] According to a fourth aspect of the embodiments of this specification, a recommendation system is provided, including a server and a client device, wherein,

[0015] The server is configured to: acquire a time-series sequence of historical user interaction items from multiple domains, the time-series sequence of historical user interaction items including interaction items corresponding to multiple users; construct a local hypergraph and a global hypergraph corresponding to different time periods, the local hypergraph being a hypergraph constructed based on interaction items within the time-series sequence of historical user interaction items in the target domain within the corresponding time period, the global hypergraph being a hypergraph constructed based on interaction items within the time-series sequence of historical user interaction items in the multiple domains within the corresponding time period, the target domain being any one of the multiple domains, and any hyperedge in the local hypergraph and global hypergraph representing a time-series subsequence of interaction items composed of interaction items of any user within the corresponding time period; and convert the local hypergraph into a time-series sequence of interaction items. The local and global hypergraphs are input into a cross-domain sequence recommendation model to predict the user's items to be interacted with; recommendation information corresponding to the items to be interacted with is sent to the user's corresponding client device. The cross-domain sequence recommendation model is used to extract collaboration information between and within the time-series subsequences of interaction items of different users based on the local and global hypergraphs through a hypergraph convolutional network, to obtain the item features of the interaction items containing the collaboration information, to determine the user's item interaction features based on the item features corresponding to the interaction items in the time-series subsequences of the interaction items of any user in multiple time periods, and to predict the items to be interacted with based on the user's item interaction features.

[0016] The client device is used to receive the recommendation information and display the interactive items through a user interface.

[0017] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0018] Memory and processor;

[0019] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the cross-domain sequence recommendation method based on hypergraph convolutional neural networks described above, and / or the cross-domain sequence recommendation model training method described above.

[0020] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the cross-domain sequence recommendation method based on a hypergraph convolutional neural network described above, and / or the cross-domain sequence recommendation model training method described above.

[0021] According to a seventh aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instruction that, when executed by a processor, implements the steps of the above-described cross-domain sequence recommendation method based on a hypergraph convolutional neural network, and / or the above-described cross-domain sequence recommendation model training method.

[0022] One embodiment of this specification implements an information extraction method, including: obtaining a time-series sequence of historical user interaction items across multiple domains, wherein the time-series sequence includes interaction items corresponding to multiple users; constructing a local hypergraph and a global hypergraph for different time periods, wherein the local hypergraph is a hypergraph constructed based on interaction items within the time-series sequence of historical user interaction items in the target domain within the corresponding time period, and the global hypergraph is a hypergraph constructed based on interaction items within the time-series sequence of historical user interaction items in the multiple domains within the corresponding time period, the target domain being any one of the multiple domains, and any hyperedge in the local hypergraph and the global hypergraph representing any user within the corresponding time period. Interactive items constitute a temporal subsequence of interactive items; the local hypergraph and the global hypergraph are input into a cross-domain sequence recommendation model to predict the user's interactive items; wherein, the cross-domain sequence recommendation model is used to extract collaboration information between and within the temporal subsequences of interactive items of different users based on the local hypergraph and the global hypergraph through a hypergraph convolutional network, to obtain the item features of the interactive items containing the collaboration information, to determine the user's item interaction features based on the item features corresponding to the interactive items in the temporal subsequences of interactive items of any user in multiple time periods, and to predict the interactive items based on the user's item interaction features.

[0023] As can be seen, this specification constructs a local hypergraph of a single domain and a global hypergraph containing multiple domains based on the historical user interaction item time series of multiple users. Hyperedges represent the time series subsequence of interaction items composed of any user's interaction items within a corresponding time period. Through hypergraph convolutional networks, collaborative information between and within the time series subsequences of interaction items of different users is extracted. It considers the influence relationship between interaction items of the same user within a certain period, as well as the influence relationship between interaction items of different users within a certain period, to obtain richer information to predict the items to be interacted with, and then to make item recommendations, which can improve the accuracy of item recommendations and thus improve the user experience. Attached Figure Description

[0024] Figure 1 This document presents a recommendation system architecture diagram for a cross-domain sequence recommendation method based on a hypergraph convolutional neural network.

[0025] Figure 2 This is a flowchart of a cross-domain sequence recommendation method based on a hypergraph convolutional neural network, as provided in this specification.

[0026] Figure 3 This is a schematic diagram of a cross-domain sequence recommendation model structure provided in this manual;

[0027] Figure 4 This is a schematic diagram of a multi-dimensional attention transfer gating module structure provided in this manual;

[0028] Figure 5 This is a flowchart of a cross-domain sequence recommendation model training method provided in this manual;

[0029] Figure 6 This is a schematic diagram of the structure of a cross-domain sequence recommendation device based on a hypergraph convolutional neural network provided in this specification;

[0030] Figure 7 This is a structural block diagram of a computing device provided in this specification. Detailed Implementation

[0031] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0032] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0033] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0034] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0035] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0036] A hypergraph is a high-level data structure used to represent complex relationships and multi-dimensional connectivity. Unlike traditional graphs, edges in a hypergraph can connect more than two nodes; these edges are called hyperedges. This allows hypergraphs to represent one-to-many or many-to-many relationships between multiple entities, making them ideal for complex scenarios that require capturing high-dimensional relationships and intricate interactions.

[0037] The degree matrix is ​​an important concept in graph theory, primarily used to describe the degree of each vertex in a graph. A degree matrix is ​​a diagonal matrix, where the elements on the diagonal represent the degree of each vertex in the graph. In an undirected graph, the degree of a vertex is the number of edges directly connected to that vertex. In a directed graph, the degree of a vertex is divided into out-degree and in-degree, representing the number of edges originating from that vertex and the number of edges pointing to that vertex, respectively.

[0038] Graph convolution is a deep learning technique for graph-structured data. Its core idea is to use the neighbor information of nodes to update the feature representation of nodes, while preserving the global structural information of the graph.

[0039] This specification provides a cross-domain sequence recommendation method based on hypergraph convolutional neural networks. This specification also relates to a cross-domain sequence recommendation model training method, a cross-domain sequence recommendation device based on hypergraph convolutional neural networks, a recommendation system, a computing device, and a computer-readable storage medium, which are described in detail in the following embodiments.

[0040] See Figure 1 , Figure 1 This is a recommendation system architecture diagram of a cross-domain sequence recommendation method based on a hypergraph convolutional neural network, provided in one embodiment of this specification. It may include client devices and servers. The client devices may be terminal devices such as mobile phones, tablets, and desktop computers, and the servers include, but are not limited to, physical servers and cloud servers.

[0041] The server is configured to acquire historical user interaction item time-series sequences from multiple domains, the historical user interaction item time-series sequences including interaction items corresponding to multiple users; construct local hypergraphs and global hypergraphs corresponding to different time periods, the local hypergraph being a hypergraph constructed based on interaction items within the historical user interaction item time-series sequences of the target domain within the corresponding time period, and the global hypergraph being a hypergraph constructed based on interaction items within the historical user interaction item time-series sequences of the multiple domains within the corresponding time period, the target domain being any one of the multiple domains, and any hyperedge in the local hypergraph and global hypergraph representing a time-series subseries of interaction items formed by any user's interaction items within the corresponding time period; and to transfer the local hypergraph to the global hypergraph. The local and global hypergraphs are input into a cross-domain sequence recommendation model to predict the user's expected interaction items. Recommendation information corresponding to the expected interaction items is sent to the user's corresponding client device. The cross-domain sequence recommendation model is used to extract collaboration information between and within the time-series subsequences of interaction items of different users based on the local and global hypergraphs through a hypergraph convolutional network, to obtain the item features of the interaction items containing the collaboration information. Based on the item features corresponding to the interaction items of any user in the time-series subsequences of interaction items in multiple time periods, the user item interaction features of the user are determined, and the expected interaction items are predicted based on the user item interaction features.

[0042] The client device is used to receive the recommendation information and display the interactive items through a user interface.

[0043] In addition, the historical user interaction item time sequence can be a sequence of interaction items sent from the client device to the server.

[0044] Taking novel recommendation as an example, multiple domains are involved, including novel and movie domains. Historical user interaction time-series sequences for both domains are obtained, with a time interval of 5 days and a period of 1 day. Local and global hypergraphs are constructed for each of the five period periods. For each period, a local hypergraph for the novel domain is constructed, including the novels clicked by users during that period. Each user corresponds to a hyperedge, including the effects clicked by that user during that period. Similarly, a local hypergraph for the movie domain, as well as global hypergraphs for both the novel and movie domains, are constructed, including the novels and movies clicked by users during that period. Any hyperedge includes the novels and movies clicked by the user corresponding to that hyperedge during that period. The local and global hypergraphs are input into a cross-domain sequence recommendation model to predict the novels users want to interact with. Recommendation information for the novels to be interacted with is sent to the user's corresponding client device, which displays the novel through the user interface. Additionally, the model can also predict the movies users want to interact with and push those movies to the client device.

[0045] In this process, the client device and the server establish a connection via a network. The network provides the medium for communication between the client device and the server. Networks can include various connection types, such as wired or wireless communication links or fiber optic cables. Data transmitted by the client device may need to undergo encoding, transcoding, compression, and other processing before being published to the server.

[0046] Servers can include servers that provide various services, such as servers that provide communication services to multiple clients, or servers that process data sent by clients. It should be noted that servers can be implemented as a distributed server cluster composed of multiple servers, or as a single server. Servers can also be servers in distributed systems, or servers integrated with blockchain. Servers can also be cloud servers 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, or intelligent cloud computing servers or intelligent cloud hosts with artificial intelligence technology.

[0047] See Figure 2 , Figure 2 A flowchart of a cross-domain sequence recommendation method based on a hypergraph convolutional neural network according to an embodiment of this specification is shown, which specifically includes the following steps.

[0048] Step 202: Obtain the time sequence of historical user interaction items for multiple domains, wherein the time sequence of historical user interaction items includes interaction items corresponding to multiple users.

[0049] Here, "domain" can be understood as a field, such as a book domain, a movie domain, an e-commerce domain, etc. "Multiple" can be understood as two or more. "Interactive item" can be understood as an item where a user interacts with it, such as a book clicked in the book domain, a movie clicked in the movie domain, or a novel clicked in the novel domain. It can also refer to different items a user can click on in different apps. The historical user interaction item time sequence can be understood as a time sequence of interactive items corresponding to multiple users within any of the multiple domains. In this time sequence, the interactive items are arranged in chronological order, and different domains correspond to different historical user interaction item time sequences.

[0050] Step 204: Construct local hypergraphs and global hypergraphs corresponding to different time periods. The local hypergraph is a hypergraph constructed based on the interaction items in the historical user interaction item time sequence of the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on the interaction items in the historical user interaction item time sequence of the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local hypergraph and the global hypergraph represents the interaction item time sequence subsequence composed of the interaction items of any user within the corresponding time period.

[0051] In this context, a time period can be understood as a segment of time within the time interval corresponding to the historical user interaction item time sequence, divided according to a preset interval. The preset time interval can be set according to the actual scenario, such as 1 day, 1 hour, 0.5 hours, 10 minutes, etc. For example, if the time interval corresponding to the historical user interaction item time sequence is the past 1 day, with a preset time interval of 2 hours, it can be divided into 6 time periods, namely T1, T2, T3, T4, T5, and T6, each time period corresponding to a different local hypergraph and global hypergraph.

[0052] A local hypergraph can be understood as a hypergraph constructed based on the time sequence of historical user interaction items in a single domain within any given time period. For example, the local hypergraph corresponding to time period T1 in the movie domain is constructed based on the movie items in the time sequence of historical user interaction items in the movie domain within time period T1. For example, time period T1 includes movie A clicked by user 1 and user 3, movie B clicked by user 2 and user 3, movie C clicked by user 4 and user 5, and movie D clicked by user 1 and user 5. The local hypergraph is constructed based on movie A, movie B, movie C, and movie D.

[0053] A global hypergraph can be understood as a hypergraph constructed based on the interaction items in the time sequence of historical user interaction items in multiple domains within any given time period. For example, the movie domain and the book domain correspond to a global hypergraph in time period T1, which is constructed based on the movie items in the time sequence of historical user interaction items in the movie domain and the book items in the time sequence of historical user interaction items in the book domain within time period T1.

[0054] Both local and global hypergraphs include a vertex set, a hypergraph edge set, an incidence matrix, degree matrices of vertices and hyperedges, and a time period corresponding to the hypergraph. The vertex set includes all interactive items within that time period; each vertex represents an interactive item. The hypergraph edge set includes the hyperedges corresponding to the user within that time period. An item interacted with by a user within that time period can be represented as multiple hypernodes, and these hypernodes, when connected, form a hyperedge. The incidence matrix represents the association between corresponding vertices and hyperedges. The degree matrix of a vertex is a diagonal matrix, with the diagonal elements representing the degree of each vertex, indicating the number of edges associated with that vertex. The degree matrix of a hyperedge is also a diagonal matrix, with the diagonal elements representing the degree of each hyperedge, indicating the number of vertices associated with that hyperedge.

[0055] An interactive item time sequence subsequence can be understood as an interactive item time sequence subsequence of any user within the time interval corresponding to the historical user interactive item time sequence. For example, with 6 time periods, any user's interactive item time sequence can also be divided into 6 subsequences. Any user's interactive item time sequence includes the user's interactive items. The items a user interacted with within a certain time period constitute an interactive item time sequence subsequence. A complete user sequence is divided into many subsequences according to time periods. For example, a user has a sequence of item 1, item 2, item 3, item 4, item 5, item 6, item 7, item 8, and item 9. Divided into three segments according to time: item 1, item 2, item 3; item 4, item 5, item 6; item 7, item 8, and item 9, each segment is a subsequence, and each hyperedge corresponds to a unique subsequence.

[0056] Understandably, in this specification, local hypergraphs of different time periods within the same domain constitute a local hypergraph flow. This flow is used to capture dynamic preferences within the time-series sequence of historical user interaction items in a single domain. First, it's necessary to construct a dynamic hypergraph, converting the time-series sequence of historical user interaction items in a single domain into a series of dynamic hypergraphs. Each hypergraph corresponds to a specific time interval, containing all user interaction items within that time interval. The hypergraph corresponding to a time interval in a single domain is thus a local hypergraph. Formally, taking domain A as an example, the time-series sequence of historical user interaction items is represented as S. A =[v1,v2,...,v i ,...], where v i This represents the user's interaction items in domain A. This specification divides the entire historical user interaction item time sequence into several time periods with a time period of t: T1 = [0, t1], T2 = [t1, t2], ..., T... n =[t n-1 ,t n ], will be in each time period T iThe local hypergraph corresponding to the built-in hypergraph is defined as follows:

[0057]

[0058] Where A represents the domain A, and i represents the time period. A hypergraph consists of a set of vertices. Hypergraph edge set represents all interactive items within that time period. It represents the set of super edges connected by multiple super nodes (i.e. interactive items) within this time period. Each super edge represents a unique subsequence of interaction, and one user corresponds to one super edge. This represents the incidence matrix of the hypergraph, where a value of 1 indicates that the corresponding node belongs to the corresponding hyperedge. and Let T represent the angle matrices of the vertices and hyperedges, respectively. i This represents the time period corresponding to the hypergraph. The association matrix, which characterizes the relationship between interactive items and hyperedges in the hypergraph, is a two-dimensional matrix with only 0s and 1s. Assuming there are n users (n hyperedges) and m items (m hypernodes), the association matrix is ​​m*n dimensional. A value of 1 indicates that the corresponding node belongs to the corresponding hyperedge, meaning the user has clicked on that item.

[0059] Over time, the topology of these dynamic hypergraphs changes continuously, eventually forming local hypergraph flows:

[0060]

[0061] Where A represents domain A (and domain B is similar), and 1-n represent the time period.

[0062] This specification demonstrates how global hypergraph flow captures cross-domain sequence patterns. Unlike local hypergraph flow, global hypergraph flow models cross-domain user interaction sequences within a shared feature space. Specifically, it merges the historical user interaction item time-series sequences from domain A and domain B into a cross-domain sequence S. M ={B1,A1,B2,...,A i B j ,...}, and in each time period T i The corresponding global dynamic hypergraph flow is constructed internally. The construction of the global hypergraph can refer to that of the local hypergraph, but its goal is to model cross-domain sequence patterns.

[0063]

[0064] Where S represents the cross-domain sequence of a user in domains A and B, and 1-n represent the time periods in which they are located.

[0065] Step 206: Input the local hypergraph and the global hypergraph into the cross-domain sequence recommendation model to predict the user's interactive items; wherein, the cross-domain sequence recommendation model is used to extract the collaboration information between and within the temporal subsequences of interactive items of different users based on the local hypergraph and the global hypergraph using a hypergraph convolutional network, to obtain the item features of the interactive items containing the collaboration information, to determine the user's item interaction features based on the item features corresponding to the interactive items in the temporal subsequences of the interactive items of any user in multiple time periods, and to predict the interactive items based on the user's item interaction features.

[0066] In this context, the cross-domain sequence recommendation model can be understood as a model trained on a neural network using a large number of user interaction item time-series training samples. The item to be interacted with can be understood as the predicted next item the user wants to click, and the hypergraph convolutional network can be understood as a graph convolutional network that processes hypergraph data. Collaboration information can be understood as the mutual influence information between interaction items of different users and the mutual influence information between interaction items of the same user. Item features can be understood as the features of interaction items containing collaboration information, which can be embedded vectors. User item interaction features can be understood as the features representing a user's item interaction characteristics within the time interval corresponding to the historical user interaction item time-series sequence; item interaction features can be understood as features containing the user's item interaction interests. Multiple time periods can be understood as time periods within the time interval corresponding to the historical user interaction item time-series sequence.

[0067] In one optional embodiment of this specification, the hypergraph convolutional network includes different convolutional structures corresponding to different time periods. Different local hypergraphs and global hypergraphs within any given time period correspond to different convolutional structures. The step of extracting collaboration information between and within the temporal subsequences of interaction items from different users using the hypergraph convolutional network based on the local and global hypergraphs to obtain item features of the interaction items containing the collaboration information may include: determining the input data for the convolutional structure, wherein the input data includes item features of interaction items within the local or global hypergraph corresponding to the convolutional structure, and the convolutional structure corresponding to the global hypergraph. Any item feature in the input data of the structure is a static embedding representation of the interaction item. Any item feature in the input data of the convolutional structure corresponding to the local hypergraph is a feature determined based on the item features output by the convolutional structure in previous time periods. If the interaction item appears for the first time in the historical user interaction item time sequence, then the item feature of the interaction item in the input data is a static embedding representation of the interaction item. The input data is input into the convolutional structure to extract the collaboration information between and within the interaction item time sequence subsequences of different users, and to obtain the item features of the interaction item output by the convolutional structure that include the collaboration information.

[0068] Here, a convolutional structure can be understood as a structure containing multiple hypergraph convolutional layers, and item features can be understood as features corresponding to interactive items. The input data may also include item features of interactive items within other local or global hypergraphs processed by the convolutional structure in previous time periods. The convolutional structure of a local hypergraph performs convolution processing on the item features of its corresponding local or global hypergraph interactive items, but does not perform convolution processing on the item features of interactive items within other local or global hypergraphs, and transmits them to the convolutional structure of subsequent time periods. The convolutional processing is a graph convolutional processing process, that is, extracting the collaboration information between and within the temporal subsequences of interactive items from different users, and obtaining the item features of the interactive items output by the convolutional structure that contain the collaboration information. After processing by the convolutional structure, the item features can contain richer information, increasing the number of item features. Furthermore, the convolutional structure corresponding to a local hypergraph transmits the item features of interactive items corresponding to other local hypergraphs to the convolutional structure corresponding to the local hypergraph in the same domain in the next time period. Any item feature in the input data of the convolutional structure corresponding to the local hypergraph is a feature determined based on the item features output by the convolutional structure corresponding to the local hypergraph of the same domain in the previous time period and the item features output by the convolutional structure corresponding to the global hypergraph.

[0069] Static embedding representations can be understood as vectorizing interactive items to obtain initial embedding vectors, which can be randomly initialized vectors. Each interactive item has a corresponding ID (identifier), and static embedding representations for each item can be generated and stored in a table. Each ID corresponds to one static embedding representation, and when a static embedding representation is needed, it is retrieved based on the ID index. For the global hypergraph, the convolutional structure of each time period retrieves the corresponding static embedding representation from this table according to the interactive items contained in the global hypergraph. For the local hypergraph, the local hypergraph of the first time period can retrieve the corresponding static embedding representation according to the interactive items contained in the local hypergraph. For the local hypergraph of subsequent time periods, the first appearance of an interactive item can also retrieve the static embedding representation from this table.

[0070] In this specification, different convolutional structures corresponding to different time periods are used to propagate the information extracted from the local and global hypergraphs in the short term backward until the project features of the complete time interval are obtained. This can capture the user's dynamic preferences and thus improve the accuracy of project prediction.

[0071] In this specification, to explore the collaborative relationships between items and sequences in a local hypergraph, a hypergraph convolutional network is designed to propagate information within the hypergraph through a message-passing mechanism:

[0072]

[0073] Where A represents the domain A, i represents the time period, and l represents the l-th convolutional layer. T represents i The node representation (i.e., item features) of the (l+1)th layer of the hypergraph convolutional layer in the local hypergraph of domain A during a given time period. and These represent the angle matrices of the vertices and hyperedges, respectively. This represents the incidence matrix of the hypergraph. This represents the hyperedge weight matrix, with all its weights assigned to 1. Let represent the trainable weight matrix associated with the l-th layer, and δ be the activation function.

[0074] In this specification, the item representation can be initialized as a static embedding representation, and dynamically updated within each time period. By stacking multiple convolutional layers, higher-order connections in the hypergraph topology can be extracted, clarifying the collaborative relationships within and between sequences. At time T i Within, after L layers of convolution, the node embeddings in the local hypergraph of domain A are represented as follows: Node embedding represents the project characteristics corresponding to the interactive item.

[0075] In the global hypergraph, the same information propagation method as in the local hypergraph is used, and a hypergraph convolutional network is employed to extract the collaborative relationships between cross-domain items and sequences. At time T... i Inside, after L layers of convolution, the node embedding in the global hypergraph is represented as:

[0076] In one optional embodiment of this specification, the cross-domain sequence recommendation model further includes a gating module connected to the output of the convolutional structures corresponding to the local hypergraph and the global hypergraph in the same time period. Different time periods correspond to different gating modules. The gating module is used to: for any interactive item, determine internal and external features based on the item features output by the convolutional structures corresponding to the target local hypergraph and the target global hypergraph; and fuse the internal and external features to obtain gating output features. The gating output features are used to fuse with the item features output by the convolutional structure corresponding to the target local hypergraph to obtain the item features corresponding to that interactive item, so that the next convolutional structure corresponding to the local hypergraph containing that interactive item will be affected by it. The project features are processed by convolution. The target local hypergraph and the target global hypergraph are the local hypergraph and global hypergraph corresponding to the same time period as the gating module. Furthermore, the internal features are the features obtained by fusing the project features of the interactive project output by the convolutional structure corresponding to the target local hypergraph and the target global hypergraph. The external features are the features obtained by fusing external node features and external hyperedge features. The external node features are the features obtained by fusing the project features of cross-domain projects. The cross-domain projects are cross-domain interactive projects selected based on the similarity with the interactive project. The external hyperedge features are the features obtained based on the features of the hyperedge where the interactive project is located. The features of the hyperedge are the features determined based on the project features of the interactive projects on the hyperedge.

[0077] The gating module can be understood as a module for cross-domain information transmission. It fuses the project features corresponding to the local hypergraph and the project features corresponding to the global hypergraph according to a preset logic and passes them to the convolutional structure corresponding to the local hypergraph in subsequent time periods. Connecting to the output of the convolutional structures corresponding to the local and global hypergraphs in the same time period can be understood as inputting the project features output by the convolutional structures corresponding to the local and global hypergraphs into the gating module in the same time period. Specifically, the gating output features are used to fuse with the project features of any interactive project output by the convolutional structure corresponding to the target local hypergraph to obtain the project features corresponding to that interactive project. Fusion can be understood as the process of merging or integrating different features to generate a more comprehensive and robust feature representation, thereby improving the model's performance and accuracy. Internal features can be understood as features corresponding to the same interactive project, while external features can be understood as correlation features between other projects. In addition to the inherent attributes of the project (internal features), the project also expresses some sequence features related to the sequence, i.e., external features. External features can be understood as a project exhibiting different features in different sequences, or the weights of various features of a project differing across sequences. A cross-domain project can be understood as an interactive project in a different domain than any of the interactive projects mentioned above.

[0078] This specification fuses the features of the same item in the convolutional structures corresponding to the local and global hypergraphs, enhancing intra-domain sequence learning. It also considers the correlation with other items across domains, as well as the correlation with different user preferences, and performs external feature alignment. This ensures that useful information is preserved to the maximum extent and irrelevant or harmful interference is reduced during cross-domain information transmission, thereby improving the recommendation effect.

[0079] In an optional embodiment of this specification, determining the internal features may include: determining the attention scores corresponding to the target local hypergraph and the target global hypergraph based on the item features of any interactive item output by the convolutional structures corresponding to the target local hypergraph and the target global hypergraph; calculating the retention ratio corresponding to the target local hypergraph and the transfer ratio corresponding to the target global hypergraph based on the attention scores; and fusing the item features of any interactive item output by the convolutional structures corresponding to the target local hypergraph and the target global hypergraph based on the retention ratio and the transfer ratio to obtain the internal features.

[0080] The attention score can be understood as an intermediate value before calculating the attention weights, reflecting the model's level of attention to different data. The retention ratio can be understood as the weights of retaining the item features corresponding to the local hypergraph, and the propagation ratio can be understood as the weights of propagating the item features corresponding to the global hypergraph backward.

[0081] In this specification, the weights of the corresponding item features in the local hypergraph and global hypergraph are determined based on the attention mechanism, and then fused. This ensures that the internal features after fusion retain more useful information and improves the recommendation effect.

[0082] In an optional implementation of this specification, determining external node features may include: calculating a first similarity between the project features corresponding to any interactive item in the target global hypergraph and the project features of cross-domain interactive items in the target global hypergraph; selecting a preset number or preset proportion of cross-domain interactive items with the highest first similarity to obtain a cross-domain item set; and aggregating the project features of the cross-domain item set to obtain external node features.

[0083] The first similarity score characterizes the degree of similarity between the item features of any interactive item and the item features of interactive items in the target global hypergraph that do not belong to the same domain as the item features of the interactive item. Vector similarity, such as cosine similarity, can be calculated between item features to determine the first similarity score. It is understood that the similarity calculation can use item features output from the convolutional structure corresponding to the global hypergraph. This specification can select a preset number or a preset proportion of cross-domain interactive items to aggregate item features and obtain external node features.

[0084] In this specification, project features of cross-domain interactive projects are selected based on similarity and then integrated to ensure the stability of cross-domain information transmission and avoid introducing interfering information.

[0085] In an optional embodiment of this specification, determining the external hyperedge features may include: determining the hyperedge where any interactive item in the target local hypergraph is located to obtain a first target hyperedge; aggregating the features of the first target hyperedge to obtain local hyperedge features; determining the hyperedge where any interactive item in the target global hypergraph is located to obtain a second target hyperedge; aggregating the features of the second target hyperedge to obtain global hyperedge features; and fusing the local hyperedge features and the global hyperedge features to obtain the external hyperedge features of any interactive item.

[0086] It is understood that this specification constructs local and global hypergraphs based on the time sequence of historical user interaction items that contain interaction items corresponding to multiple users. In the local and global hypergraphs, any interaction item can exist on multiple hyperedges.

[0087] In this specification, for any interactive item, by integrating the features of its superedge, it is possible to obtain the influence information between different users with similar interests, thereby improving the recommendation effect.

[0088] The gating module enables stable information transfer and integration between the local and global hypergraph flows. It is a key component connecting the local and global hypergraph flows, and its design considers the dynamic nature of user preferences and the complexity of cross-domain information, aiming to avoid negative transfer and improve the performance of cross-domain recommendations. Fine-grained knowledge transfer can be achieved through the following methods:

[0089] Internal feature fusion aims to integrate inherent item attributes that manifest differently in local and global spaces to enhance intra-domain sequence learning. This process is achieved by calculating attention scores, which determine the proportion of local information retained and the proportion of global information transferred during fusion.

[0090] First, calculate the attention score between the local and global candidate sets.

[0091]

[0092] Among them, e b,x and e s,x These represent the node embedding representations of interactive item x as a hypergraph node in the local and global hypergraphs, respectively. and W I These are all trainable weight parameters, and σ is the tanh activation function. Using s l and s g Let represent the attention score of the node embedding representation in the local and global hypergraphs, and [] represent the concatenation operation.

[0093] Based on the attention score, node embedding representations in the local and global hypergraph flows are fused.

[0094] e ii,x =re b,x +(1-r)e s,x ;

[0095]

[0096] Where r is the proportion of local information retained, and 1-r is the proportion of global information transmitted. ii,x Let exp(x) represent the internal characteristics of the obtained project x.

[0097] External feature alignment aims to explore the correlations between externalized features of common items in different sequence contexts and reduce domain bias. This process involves two proxies: external node proxies (i.e., selected interactive items) and external hyperedge proxies (i.e., selected hyperedges), used for item-level transfer and user-level transfer, respectively. External features can be understood as features extracted and aggregated from the two proxies, and alignment can be understood as the operation of processing data with different data distributions or different feature spaces, thereby eliminating the differences between data to a certain extent.

[0098] External node proxies: Node proxies are selected by measuring the implicit relationships between items in the global dynamic hypergraph, and then an attention mechanism is used to aggregate the node proxies.

[0099]

[0100] Among them, W N It is a learnable parameter, e np，x This indicates the fusion state of the node proxy, i.e., the external node characteristics corresponding to the interactive item. The set of node proxies selected to measure the implicit relationships between items in the global dynamic hypergraph, i.e., the cross-domain item set. s,x This represents the node embedding representation of the interactive item x as a hypergraph node in the global hypergraph, e. s,i Let represent the node embedding representation of the interaction item i in the node proxy set as a hypergraph node in the global hypergraph, i.e., the external node feature. a(x,i) represents the weights designed according to the attention mechanism, with the aim of better information extraction. s,j The node embedding representation represents the interactive items in the node proxy set. This specification traverses each node in the node proxy set. For each item x in domain B of the global candidate set, based on the similarity of the supernodes, the Top K most similar nodes in domain A of the global set are selected as the node proxy set.

[0101] External hyperedge proxies obtain hyperedge representations by extracting the dynamic representations after hypergraph convolution, and then aggregate local and global hyperedge proxies to generate an integrated sequence representation for external features.

[0102] First, we obtain a dynamic representation of all hyperedges by aggregating the nodes on all hyperedges:

[0103]

[0104] The resulting matrix Indicates a local hypergraph The embedding representation of all hyperedges (i.e., the features of the hyperedges). The superscript 2 indicates the second time period, the local hypergraph in this embodiment. For example, the representation of hyperedges in other local hypergraphs and global hypergraphs can refer to this example. For each candidate node x in the local hypergraph, extract the hyperedges to which x belongs. All hyperedges are used as local hyperedge proxies.

[0105]

[0106] Then, as shown in the formula above, the fusion state u of the hyperedge proxy of the local hypergraph is obtained through an average pooling layer. l,x ub,t To represent a local hyperedge proxy, the fusion state u of hyperedge proxies on the global hypergraph can be obtained using the same method. g,x .

[0107] e ep,x =ru l,x +(1-r)u g,x ;

[0108] Where r is the proportion of local information retained, and 1-r is the proportion of global information transmitted. ep,x The external hyperedge features are generated by combining the local and global fusion states of the hyperedge proxy (i.e., local hyperedge features and global hyperedge features).

[0109] Furthermore, by utilizing the fused internal and external node features, as well as external hyperedge features, a dynamic item embedding is generated. This dynamic item embedding is then fused with the output of the convolutional structure corresponding to the local hypergraph in the same time period, and used as the input to the convolutional structure of the local hypergraph in the next time period. In this way, stable transmission and integration of cross-domain information is achieved through dynamic hypergraph flow.

[0110] In an optional embodiment of this specification, determining the user's user item interaction features based on the item features corresponding to the interaction items in the time-series subsequences of the interaction items in multiple time periods may include: combining the time-series subsequences of the interaction items of any user in multiple time periods into a time-series sequence of user interaction items in the target domain; fusing the item features output by the convolutional structure corresponding to the local hypergraph of the interaction items in the time-series sequence of user interaction items into the user's user item interaction features based on fusion weights; wherein, the fusion weights are determined based on the item features of a specified interaction item in the time-series sequence of user interaction items.

[0111] In this context, the target domain can be understood as any one of multiple domains. This specification can predict the items a user will interact with in the target domain. That is, the fusion involves the item features output by the convolutional structure corresponding to the local hypergraph of the target domain. The fusion weights can be understood as the weights of the item features when fusing multiple item features. The preset time period can be understood as a specified time period. For example, the first preset time period is the time period during which the user last interacted with that item within the time interval corresponding to the historical user interaction item time sequence. In other words, the first preset time period for any interaction item can be specified as the time period during which the user last clicked on that item. Different interaction items can correspond to different time periods.

[0112] In this manual, by presetting a time period, you can select a time period close to the predicted interactive item for item feature fusion. You can specify an interactive item or a close interactive item for weight calculation. The closer to the interactive item, the higher the correlation with the interactive item. In addition, it combines the user's short-term and long-term interests, which can improve the accuracy of the interactive item prediction.

[0113] In this specification, a sequence representation containing rich information is generated. Combining the user's short-term and long-term interests, a complete single-domain sequence is formed by recombining subsequences, and a self-attention layer is used to extract long-term interest patterns.

[0114]

[0115] x a,u Let Su represent the final representation of sequence Su in domain A, i.e., the user item interaction characteristics. Su represents the time sequence of user interaction items for user u. W Q W K and W v These are learnable training parameters, x a,m This represents the node embedding representation of the time period corresponding to the last item in the time sequence of user interaction items. This time period can be the last time period in which the user clicked on this item. a,i This represents the node embedding representation of sequence Su within a specific interactive item in domain A, where the time period can be the last time period during which the user clicked on that item. D is the embedding dimension and a scaling factor. This is to avoid excessively large inner product values, especially when the dimension is very high.

[0116] In one optional embodiment of this specification, predicting the item to be interacted based on the user item interaction features may include: calculating a second similarity between the user item interaction features and the item features output by the convolutional structure corresponding to the local hypergraph of the interactive item in a second preset time period; and predicting the item to be interacted based on the second similarity.

[0117] The second similarity can characterize the degree of similarity between item features, and can be the vector similarity between item features. The second similarity can be calculated between the user item interaction features and the item features output by the convolutional structure corresponding to the local hypergraph of the target domain's interaction items during a second preset time period. Furthermore, in an optional implementation, the interaction item with the highest second similarity can be determined as the item to be interacted with.

[0118] Specifically, the second preset time period can be the time period during which the interaction item last appeared within the time interval corresponding to the historical user interaction item time sequence. Using the item features output from the most recent time period can improve the accuracy of item prediction. For example, if the time period to be predicted is T15, item 1 last occurred at T13, and item 2 last occurred at T14, this specification calculates the similarity between the item features of item 1 output by convolution at T13 and the user's item interaction features, and calculates the similarity between the item features of item 2 output by convolution at T14 and the user's item interaction features. The closer the time periods, the more similar the item features of the items.

[0119] In this specification, the accuracy of predicting the interactive item can be improved by based on the similarity between the user's item interaction characteristics and the project characteristics of the interactive item.

[0120] As can be seen, this specification constructs a local hypergraph of a single domain and a global hypergraph containing multiple domains based on the historical user interaction item time series of multiple users. Hyperedges represent the time series subsequence of interaction items composed of any user's interaction items within a corresponding time period. Through hypergraph convolutional networks, collaborative information between and within the time series subsequences of interaction items of different users is extracted. It considers the influence relationship between interaction items of the same user within a certain period, as well as the influence relationship between interaction items of different users within a certain period, to obtain richer information to predict the items to be interacted with, and then to make item recommendations, which can improve the accuracy of item recommendations and thus improve the user experience.

[0121] This specification extracts complex collaborative relationships within domain sequences, considering both internal (within the same sequence) and external (between different sequences) collaborations in user behavior. Within the same sequence, items interacting with by users over a certain period often share similar characteristics. Therefore, even discontinuous transitions within a sequence can still be collaborative. Between different sequences, users with shared interests are more likely to execute similar behavioral patterns hidden beneath their interaction trajectories. Both aspects of collaborative relationships can profoundly influence the representation learning of items and sequences. This specification fully leverages the complex dependencies between sequences. Furthermore, it captures and stably transfers interaction patterns between domains; transitions within a sequence can indicate the evolution of user preferences, which do not always follow stable patterns. From a fine-grained perspective, users' short-term interests may occasionally fluctuate, and directly transferring each pattern (i.e., useful information) across domains may introduce noise into both domains, affecting recommendation effectiveness. Simultaneously, the factors determining interaction are a complex combination of internal and external attributions. To effectively utilize cross-domain information, this embodiment distinguishes between internal and external features and adopts different transmission strategies accordingly. This helps to maximize the retention of useful information and reduce irrelevant or harmful interference during cross-domain information transmission, better meeting the needs of real-world scenarios.

[0122] See Figure 3 , Figure 3 This diagram illustrates a cross-domain sequence recommendation model structure according to an embodiment of this specification, wherein the static item embedding of domain A includes static embedding representations corresponding to all interaction items in the time-series sequence of historical user interaction items in domain A, constructing a local hypergraph flow for domain A. Let represent a local hypergraph of domain A during time intervals from 1 to n, where each local hypergraph corresponds to a multi-layer convolutional structure. Including interactive projects: A1, A2, A3, The interactive items include: A1, A2, A3, A4, and A5. A5, A2, and A4 form a hyperedge, and A1, A2, and A3 form another hyperedge. The interactive items include: A1, A2, A3, A4, A5, A6, and A7. The static item embeddings in domain B include the static embedding representations corresponding to all interactive items in the time-series sequence of historical user interactions in domain B, constructing a local hypergraph flow for domain B. Let B1 represent the local hypergraph of domain B during time intervals 1 to n, and B6 represent the interactive items in domain B. Construct the global hypergraph flow. Represents the global super time interval from 1 to n. Figures 1 to 7 The interactive items are represented by the multi-dimensional attention transfer gating module mentioned above. The flow interest generator is used to generate user item interaction features and then make the next prediction. For example, if domain A is the movie domain, predict the next movie; if domain B is the book domain, predict the next book.

[0123] See Figure 4 , Figure 4 This diagram illustrates the structure of a multi-dimensional attention transfer gating module according to an embodiment of this specification. Taking the multi-dimensional attention transfer gating module corresponding to the second time period as an example, the global candidate node is item 2, and the local candidate node is B2. The global candidate node and the local candidate node are the same item. Internal feature fusion is performed, that is, the item features of the global candidate node and the local candidate node are fused to obtain the internal feature E. ii , This includes projects 1 through 5, from Select the global superedge proxy, including the superedge consisting of Project 1, Project 1, Project 2, Project 2 and Project 3, and the superedge consisting of Project 4, Project 1, Project 3 and Project 2. Including items 1 through 4, represented by B1, B2, B3, and B4, from... We select local hyperedge proxies, including the hyperedge formed by items 1 and 2, and the hyperedge formed by items 4, 3, and 2. We treat the global hyperedge proxy and the local hyperedge proxy as external hyperedge proxies, and fuse the features of the external hyperedge proxies to obtain the external hyperedge feature E. epSelect the external node proxy for Project 2, perform external node proxy transfer, and obtain the external node feature E. np Perform state fusion, and combine the features obtained from the state fusion with... The output features of the corresponding convolutional structures are fused to obtain the fused features. These fused features, along with those selected from the static embeddings, are then combined... Correspondingly, the static embeddings of items (B5, B6) that do not appear in the preceding hypergraph represent the input. The corresponding convolutional structure.

[0124] This specification describes dual hypergraph flow modeling. By constructing local and global hypergraph flows, information propagation is achieved, comprehensively capturing and modeling intra-domain and cross-domain sequential collaboration relationships. Through multivariate attention-based transfer gating, information transfer and integration are realized between local and global hypergraph flows, further achieving internal feature fusion and external feature alignment. This ensures the stability and accuracy of cross-domain pattern transmission. Furthermore, a flow interest generator produces information-rich representations containing long-term user preference patterns, thereby improving the performance of the recommendation system.

[0125] As can be seen, this specification adopts a cross-domain sequence recommendation approach. By utilizing shared sequence patterns across different domains, it effectively alleviates the data sparsity problem. By leveraging user interaction data across multiple domains, it helps the model to more comprehensively understand user interests and preferences, thereby improving the accuracy of sequence recommendations. To effectively mine and integrate cross-domain sequence patterns, avoid introducing noise, and achieve stable cross-domain information transmission, a cross-domain sequence recommendation method based on hypergraph convolutional neural networks is provided. This method models the sequence collaboration relationships between domains by constructing dual hypergraph flows for user sequence segments within each domain in different time periods. A multivariate attention transfer gating system is designed to achieve internal feature fusion and external feature alignment at the hypergraph connections, completing stable cross-domain information transmission and fusion, avoiding negative propagation, and effectively alleviating the single-domain sequence data sparsity problem. Finally, a stream interest generator is used to generate information-rich sequential representations, thereby improving the overall performance of the recommendation system.

[0126] See Figure 5 , Figure 5 A flowchart illustrating a cross-domain sequence recommendation model training method according to an embodiment of this specification is shown, including:

[0127] Step 502: Obtain time-series sequence samples of user interaction items from multiple domains and the label information corresponding to the time-series sequence samples of user interaction items from the training sample set, wherein the time-series sequence samples of user interaction items include interaction items corresponding to multiple users.

[0128] The tag information can be the next item clicked by each user. It is understandable that the next item clicked by a user will be different in different domains.

[0129] Step 504: Construct local hypergraphs and global hypergraphs corresponding to different time periods. The local hypergraph is a hypergraph constructed based on the interaction items in the time-series sample of the user interaction items in the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on the interaction items in the time-series sample of the user interaction items in the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local hypergraph and global hypergraph represents the time-series subsequence of interaction items formed by the interaction items of any user in the corresponding time period.

[0130] Step 506: Input the local hypergraph and global hypergraph into a preset model to predict the user's interactive items. The preset model is used to extract collaboration information between and within the temporal subsequences of interactive items of different users based on the local and global hypergraphs using a hypergraph convolutional network. This yields the item features of the interactive items containing the collaboration information. Based on the item features corresponding to the interactive items in the temporal subsequences of the interactive items of any user in multiple time periods, the user's item interaction features are determined. The user's item interaction features are then used to predict the interactive items based on the user's item interaction features.

[0131] Step 508: Calculate the training loss using the interactive item and label information.

[0132] Step 510: Update the parameters of the preset model based on the training loss to obtain the preset model with updated parameters.

[0133] If the training termination condition is not met, the step of obtaining time-series sequence samples of user interaction items from multiple domains and the corresponding label information of the time-series sequence samples of user interaction items from the training sample set continues.

[0134] Step 512: When the training termination condition is met, the current preset model with updated parameters is determined as the cross-domain sequence recommendation model.

[0135] The training termination condition can be that the training loss is less than a preset threshold, or that the number of model iterations reaches a preset threshold.

[0136] As can be seen, during the training process, this specification constructs a local hypergraph of a single domain and a global hypergraph containing multiple domains based on the historical user interaction item time-series sequences corresponding to multiple users. Hyperedges represent the time-series subsequences of interaction items formed by any user's interaction items within a corresponding time period. The hypergraph convolutional network extracts the collaborative information between and within the time-series subsequences of interaction items of different users, considering the influence relationship between interaction items of the same user within a certain period, as well as the influence relationship between interaction items of different users within a certain period, to obtain richer information to predict the items to be interacted with. The model learns richer information, which can improve the accuracy of item recommendation, thereby improving the user experience.

[0137] Corresponding to the above method embodiments, this specification also provides embodiments of a cross-domain sequence recommendation device based on a hypergraph convolutional neural network. Figure 6 This diagram illustrates a schematic representation of a cross-domain sequence recommendation device based on a hypergraph convolutional neural network, according to one embodiment of this specification. Figure 6 As shown, the device includes:

[0138] The sequence acquisition module 602 is used to acquire a time-series sequence of historical user interaction items from multiple domains, wherein the time-series sequence of historical user interaction items includes interaction items corresponding to multiple users.

[0139] The hypergraph construction module 604 is used to construct local hypergraphs and global hypergraphs corresponding to different time periods. The local hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local hypergraph and the global hypergraph represents the time sequence subsequence of interaction items formed by the interaction items of any user in the corresponding time period.

[0140] The project prediction module 606 is used to input the local hypergraph and global hypergraph into the cross-domain sequence recommendation model to predict the user's interactive items;

[0141] The cross-domain sequence recommendation model is used to extract collaboration information between and within the time-series subsequences of interaction items of different users based on local and global hypergraphs using a hypergraph convolutional network, thereby obtaining the item features of the interaction items containing the collaboration information. Based on the item features corresponding to the interaction items of any user in the time-series subsequences of the interaction items in multiple time periods, the user item interaction features of the user are determined, and the item to be interacted is predicted based on the user item interaction features.

[0142] In an optional embodiment of this specification, the hypergraph convolutional network includes different convolutional structures corresponding to different time periods. Different local hypergraphs and global hypergraphs within any given time period correspond to different convolutional structures. The item prediction module 606 is specifically used to determine the input data of the convolutional structure. The input data includes item features of interactive items within the local or global hypergraph corresponding to the convolutional structure. Any item feature in the input data of the convolutional structure corresponding to the global hypergraph is a static embedding representation of the interactive item. Any item feature in the input data of the convolutional structure corresponding to the local hypergraph is a feature determined based on the item features output by the convolutional structure in previous time periods. If the interactive item appears for the first time in the historical user interactive item time sequence, then the item feature of the interactive item in the input data is a static embedding representation of the interactive item. The input data is input into the convolutional structure to extract the collaboration information between and within the time sequence subsequences of interactive items of different users, thereby obtaining the item features of the interactive items output by the convolutional structure that include the collaboration information.

[0143] Furthermore, the cross-domain sequence recommendation model also includes a gating module connected to the output of the convolutional structures corresponding to the local hypergraph and global hypergraph in the same time period. Different time periods correspond to different gating modules. The gating module is used to: for any interactive item, determine internal and external features based on the item features output by the convolutional structures corresponding to the target local hypergraph and the target global hypergraph, and fuse the internal and external features to obtain gating output features. The gating output features are used to fuse with the item features output by the convolutional structure corresponding to the target local hypergraph to obtain the item features corresponding to the interactive item, so that the next convolutional structure corresponding to the local hypergraph containing the interactive item performs convolution processing on the item features. The target local hypergraph and the target global hypergraph are the local hypergraph and global hypergraph in the same time period as the gating module.

[0144] Furthermore, the internal features are the features obtained by fusing the project features of the interactive project output by the convolutional structure corresponding to the target local hypergraph and the target global hypergraph. The external features are the features obtained by fusing external node features and external hyperedge features. The external node features are the features obtained by fusing the project features of cross-domain projects. The cross-domain projects are cross-domain interactive projects selected based on their similarity to the interactive project. The external hyperedge features are the features obtained based on the features of the hyperedge where the interactive project is located. The hyperedge features are the features determined based on the project features of the interactive projects on the hyperedge.

[0145] The gating module includes:

[0146] The internal feature determination submodule is specifically used to determine the attention scores corresponding to the target local hypergraph and the target global hypergraph based on the item features of any interactive item output by the convolutional structures corresponding to the target local hypergraph and the target global hypergraph; calculate the retention ratio corresponding to the target local hypergraph and the transfer ratio corresponding to the target global hypergraph based on the attention scores; and fuse the item features of any interactive item output by the convolutional structures corresponding to the target local hypergraph and the target global hypergraph based on the retention ratio and the transfer ratio to obtain internal features.

[0147] The external node feature determination submodule is specifically used to calculate the first similarity between the project feature corresponding to any interactive item in the target global hypergraph and the project feature of the cross-domain interactive item in the target global hypergraph; to filter the cross-domain interactive items with the highest first similarity to obtain a cross-domain item set; and to aggregate the project features of the cross-domain item set to obtain the external node features.

[0148] The external hyperedge feature determination submodule is specifically used for: determining the hyperedge where any interactive item in the target local hypergraph is located, obtaining a first target hyperedge; aggregating the features of the first target hyperedge to obtain local hyperedge features; determining the hyperedge where any interactive item in the target global hypergraph is located, obtaining a second target hyperedge; aggregating the features of the second target hyperedge to obtain global hyperedge features; and fusing the local hyperedge features and the global hyperedge features to obtain the external hyperedge features of any interactive item.

[0149] Furthermore, the cross-domain sequence recommendation model also includes a user item interaction feature determination module, specifically used to combine the time-series subsequences of interaction items of any user in multiple time periods into a time-series sequence of user interaction items in the target domain; and to fuse the item features output by the convolutional structure corresponding to the local hypergraph of the interaction items in the time-series sequence of the user interaction items into the user item interaction features of the user based on the fusion weights; wherein, the fusion weights are determined based on the item features of a specified interaction item in the time-series sequence of the user interaction items.

[0150] Furthermore, the cross-domain sequence recommendation model also includes an item prediction module, specifically used to: calculate the second similarity between the user item interaction features and the item features output by the convolutional structure corresponding to the local hypergraph of the interactive item in the second preset time period; and predict the item to be interacted based on the second similarity.

[0151] The above is an illustrative scheme of a cross-domain sequence recommendation device based on a hypergraph convolutional neural network according to this embodiment. It should be noted that the technical solution of this cross-domain sequence recommendation device based on a hypergraph convolutional neural network belongs to the same concept as the technical solution of the cross-domain sequence recommendation method based on a hypergraph convolutional neural network described above. Details not described in detail in the technical solution of the cross-domain sequence recommendation device based on a hypergraph convolutional neural network can be found in the description of the technical solution of the cross-domain sequence recommendation method based on a hypergraph convolutional neural network described above.

[0152] As can be seen, this specification constructs a local hypergraph of a single domain and a global hypergraph containing multiple domains based on the historical user interaction item time series of multiple users. Hyperedges represent the time series subsequence of interaction items composed of any user's interaction items within a corresponding time period. Through hypergraph convolutional networks, collaborative information between and within the time series subsequences of interaction items of different users is extracted. It considers the influence relationship between interaction items of the same user within a certain period, as well as the influence relationship between interaction items of different users within a certain period, to obtain richer information to predict the items to be interacted with, and then to make item recommendations, which can improve the accuracy of item recommendations and thus improve the user experience.

[0153] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0154] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0155] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0156] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 700 can also be a mobile or stationary server.

[0157] The processor 720 is used to execute the following computer program / instruction, which, when executed by the processor, implements the steps of the cross-domain sequence recommendation method based on the hypergraph convolutional neural network described above, and / or the cross-domain sequence recommendation model training method described above.

[0158] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computing device embodiments are relatively simple in description because they are fundamentally similar to the cross-domain sequence recommendation method embodiments of hypergraph convolutional neural networks. Relevant details can be found in the descriptions of the cross-domain sequence recommendation method embodiments of hypergraph convolutional neural networks.

[0159] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the aforementioned cross-domain sequence recommendation method based on a hypergraph convolutional neural network, and / or the aforementioned cross-domain sequence recommendation model training method.

[0160] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computer-readable storage medium embodiment is described simply because it is fundamentally similar to the cross-domain sequence recommendation method embodiment of the hypergraph convolutional neural network. Relevant details can be found in the description of the cross-domain sequence recommendation method embodiment of the hypergraph convolutional neural network.

[0161] An embodiment of this specification also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of the aforementioned cross-domain sequence recommendation method of a hypergraph convolutional neural network, and / or the aforementioned cross-domain sequence recommendation model training method.

[0162] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the cross-domain sequence recommendation method of the hypergraph convolutional neural network described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the cross-domain sequence recommendation method of the hypergraph convolutional neural network described above.

[0163] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0164] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0165] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0166] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0167] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A cross-domain sequence recommendation method based on a hypergraph convolutional neural network, characterized in that, include: Obtain a time-series sequence of historical user interaction items from multiple domains, wherein the time-series sequence of historical user interaction items includes interaction items corresponding to multiple users; Construct local and global hypergraphs corresponding to different time periods. The local hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local and global hypergraphs represents the time sequence subsequence of interaction items formed by the interaction items of any user in the corresponding time period. The local hypergraph and the global hypergraph are input into the cross-domain sequence recommendation model to predict the user's items to be interacted with. The cross-domain sequence recommendation model is used to determine the input data of the convolutional structure based on the local and global hypergraphs using a hypergraph convolutional network. The input data is then fed into the convolutional structure to extract collaboration information between and within the temporal subsequences of interaction items for different users, obtaining item features of the interaction items that include the collaboration information. Based on the item features corresponding to the interaction items within the temporal subsequences of the interaction items in multiple time periods, the user's user item interaction features are determined, and the item to be interacted with is predicted based on the user item interaction features. The hypergraph convolutional network includes different convolutional structures corresponding to different time periods. Different local and global hypergraphs within any time period correspond to different convolutional structures. The input data includes item features of interaction items within the local or global hypergraph corresponding to the convolutional structure. Any item feature in the input data of the convolutional structure corresponding to the global hypergraph is a static embedding representation of the interaction item, and any item feature in the input data of the convolutional structure corresponding to the local hypergraph is a feature determined based on the item features output by the convolutional structure in previous time periods.

2. The method according to claim 1, characterized in that, If the interactive item appears for the first time in the time sequence of the historical user interactive items, then the item feature of the interactive item in the input data is the static embedded representation of the interactive item.

3. The method according to claim 2, characterized in that, The cross-domain sequence recommendation model further includes a gating module connected to the output of the convolutional structures corresponding to the local hypergraph and the global hypergraph in the same time period. Different time periods correspond to different gating modules, wherein the gating module is used for: For any interactive item, internal and external features are determined based on the item features output by the convolutional structure corresponding to the target local hypergraph and the target global hypergraph. The internal and external features are then fused to obtain the gated output features. The gated output feature is used to fuse with the item feature output by the convolutional structure corresponding to the target local hypergraph to obtain the item feature corresponding to the interactive item. The next convolutional structure corresponding to the local hypergraph containing the interactive item performs convolution processing on the item feature. The target local hypergraph and the target global hypergraph are the local hypergraph and global hypergraph corresponding to the same time period as the gated module. Furthermore, the internal features are the features obtained by fusing the project features of the interactive project output by the convolutional structure corresponding to the target local hypergraph and the target global hypergraph. The external features are the features obtained by fusing external node features and external hyperedge features. The external node features are the features obtained by fusing the project features of cross-domain projects. The cross-domain projects are cross-domain interactive projects selected based on their similarity to the interactive project. The external hyperedge features are the features obtained based on the features of the hyperedge where the interactive project is located. The hyperedge features are the features determined based on the project features of the interactive projects on the hyperedge.

4. The method according to claim 3, characterized in that, Determine internal characteristics, including: Based on the item features of any interactive item output by the convolutional structures corresponding to the target local hypergraph and the target global hypergraph, determine the attention scores corresponding to the target local hypergraph and the target global hypergraph; Based on the attention score, calculate the retention ratio of the target local hypergraph and the transfer ratio of the target global hypergraph; Based on the retention ratio and the transfer ratio, the project features of any interactive item output by the convolutional structure corresponding to the target local hypergraph and the target global hypergraph are fused to obtain internal features.

5. The method according to claim 3, characterized in that, Determine the characteristics of external nodes, including: Calculate the first similarity between the item features corresponding to any interactive item in the target global hypergraph and the item features of cross-domain interactive items in the target global hypergraph; The cross-domain item set is obtained by selecting the cross-domain interactive items with the highest first similarity from a preset number or preset proportion. The project features of the cross-domain project set are aggregated to obtain external node features.

6. The method according to claim 3, characterized in that, Determine the features of the external hyperedge, including: Determine the hyperedge where any interactive item in the target local hypergraph is located to obtain the first target hyperedge, and aggregate the features of the first target hyperedge to obtain the local hyperedge features; Determine the hyperedge containing any interactive item in the target global hypergraph to obtain the second target hyperedge, and aggregate the features of the second target hyperedge to obtain the global hyperedge features; By fusing the local hyperedge features and the global hyperedge features, the external hyperedge features of any interactive item are obtained.

7. The method according to claim 1, characterized in that, The process of determining a user's user item interaction characteristics based on the item characteristics corresponding to the interaction items within the time-series sub-sequences of the interaction items in multiple time periods includes: Combine the time-series subsequences of any user's interaction items across multiple time periods into a time-series sequence of user interaction items for the target domain; Based on the fusion weight, the item features output by the convolutional structure corresponding to the local hypergraph of the user interaction item time sequence are fused into the user item interaction features of the user. The fusion weight is determined based on the project characteristics of a specified interaction project within the time sequence of the user interaction project.

8. The method according to claim 1, characterized in that, The step of predicting the item to be interacted with based on the user item interaction features includes: Calculate the second similarity between the user item interaction features and the item features output by the convolutional structure corresponding to the local hypergraph of the interactive item in the second preset time period; The item to be interacted with is predicted based on the second similarity.

9. A method for training a cross-domain sequence recommendation model, characterized in that, include: The training sample set is used to obtain time-series sequence samples of user interaction items from multiple domains and the label information corresponding to the time-series sequence samples of user interaction items. The time-series sequence samples of user interaction items include interaction items corresponding to multiple users. Construct local and global hypergraphs corresponding to different time periods. The local hypergraph is a hypergraph constructed based on the interaction items in the time sequence samples of the user interaction items in the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on the interaction items in the time sequence samples of the user interaction items in the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local and global hypergraphs represents the time sequence subsequence of interaction items formed by the interaction items of any user in the corresponding time period. The local and global hypergraphs are input into a preset model to predict the user's interactive items. The preset model is used to determine the input data of the convolutional structure based on the local and global hypergraphs using a hypergraph convolutional network. The input data is then input into the convolutional structure to extract collaboration information between and within the temporal subsequences of interactive items for different users, obtaining item features of the interactive items containing the collaboration information. Based on the item features corresponding to the interactive items in the temporal subsequences of the interactive items for any user across multiple time periods, the user's item interaction features are determined, and the interactive items are predicted based on these user item interaction features. The hypergraph convolutional network includes different convolutional structures corresponding to different time periods. Different local and global hypergraphs within any time period correspond to different convolutional structures. The input data includes item features of interactive items within the local or global hypergraphs corresponding to the convolutional structure. Any item feature in the input data of the convolutional structure corresponding to the global hypergraph is a static embedding representation of the interactive item, and any item feature in the input data of the convolutional structure corresponding to the local hypergraph is a feature determined based on the item features output by the convolutional structure in previous time periods. The training loss is calculated using the interactive items and label information. The parameters of the preset model are updated based on the training loss to obtain the preset model with updated parameters. If the training termination condition is not met, the step of obtaining time-series samples of user interaction items from multiple domains and the label information corresponding to the time-series samples of user interaction items from the training sample set continues. When the training termination condition is met, the current preset model with updated parameters is determined as the cross-domain sequence recommendation model.

10. A cross-domain sequence recommendation device based on a hypergraph convolutional neural network, characterized in that, include: The sequence acquisition module is used to acquire the time-series sequence of historical user interaction items in multiple domains, wherein the time-series sequence of historical user interaction items includes interaction items corresponding to multiple users; The hypergraph construction module is used to construct local hypergraphs and global hypergraphs corresponding to different time periods. The local hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the target domain within the corresponding time period. The global hypergraph is a hypergraph constructed based on the interaction items in the time sequence of the historical user interaction items in the multiple domains within the corresponding time period. The target domain is any one of the multiple domains. Any hyperedge in the local hypergraph and the global hypergraph represents the time sequence subsequence of interaction items formed by the interaction items of any user in the corresponding time period. The project prediction module is used to input the local hypergraph and global hypergraph into the cross-domain sequence recommendation model to predict the user's interactive items; The cross-domain sequence recommendation model is used to determine the input data of the convolutional structure based on local and global hypergraphs using a hypergraph convolutional network. The input data is then fed into the convolutional structure to extract collaboration information between and within the temporal subsequences of interaction items for different users, obtaining item features of the interaction items that include the collaboration information. Based on the item features corresponding to the interaction items within the temporal subsequences of the interaction items in multiple time periods, the model determines the user's item interaction features and predicts the item to be interacted with based on these user item interaction features. The hypergraph convolutional network includes different convolutional structures corresponding to different time periods. Different local and global hypergraphs within any time period correspond to different convolutional structures. The input data includes item features of interaction items within the local or global hypergraph corresponding to the convolutional structure. Any item feature in the input data of the convolutional structure corresponding to the global hypergraph is a static embedding representation of the interaction item, and any item feature in the input data of the convolutional structure corresponding to the local hypergraph is a feature determined based on the item features output by the convolutional structure in previous time periods.

11. A recommendation system, characterized in that, Including server and client devices, among which, The server is configured to acquire historical user interaction item time-series sequences across multiple domains, each sequence including interaction items corresponding to multiple users; construct local and global hypergraphs for different time periods, where the local hypergraph is a hypergraph constructed based on interaction items within the historical user interaction item time-series sequences of the target domain within the corresponding time period, and the global hypergraph is a hypergraph constructed based on interaction items within the historical user interaction item time-series sequences of the multiple domains within the corresponding time period. The target domain is any one of the multiple domains, and any hyperedge in the local and global hypergraphs represents a temporal subsequence of interaction items formed by any user's interaction items within the corresponding time period; input the local and global hypergraphs into a cross-domain sequence recommendation model to predict the user's pending interaction items; and send recommendation information corresponding to the pending interaction items to the user's corresponding client device. The cross-domain sequence recommendation model is configured to determine the input data of the convolutional structure based on the local and global hypergraphs using a hypergraph convolutional network. The input data is input into a convolutional structure to extract collaboration information between and within the temporal subsequences of interaction items for different users, thereby obtaining item features of the interaction items that include the collaboration information. Based on the item features corresponding to the interaction items in the temporal subsequences of the interaction items for any user in multiple time periods, the user item interaction features of that user are determined, and the item to be interacted with is predicted based on the user item interaction features. The hypergraph convolutional network includes different convolutional structures corresponding to different time periods. Different local hypergraphs and global hypergraphs in any time period correspond to different convolutional structures. The input data includes item features of interaction items in the local hypergraph or global hypergraph corresponding to the convolutional structure. Any item feature in the input data of the convolutional structure corresponding to the global hypergraph is a static embedding representation of the interaction item, and any item feature in the input data of the convolutional structure corresponding to the local hypergraph is a feature determined based on the item features output by the convolutional structure in previous time periods. The client device is used to receive the recommendation information and display the interactive items through a user interface.

12. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the cross-domain sequence recommendation method based on hypergraph convolutional neural networks as described in any one of claims 1 to 8, and / or the cross-domain sequence recommendation model training method as described in claim 9.

13. A computer-readable storage medium, characterized in that, It stores a computer program / instruction that, when executed by a processor, implements the steps of the cross-domain sequence recommendation method based on a hypergraph convolutional neural network as described in any one of claims 1 to 8, and / or the cross-domain sequence recommendation model training method as described in claim 9.

14. A computer program product, characterized in that, Includes a computer program / instruction that, when executed by a processor, implements the steps of the cross-domain sequence recommendation method based on a hypergraph convolutional neural network as described in any one of claims 1 to 8, and / or the cross-domain sequence recommendation model training method as described in claim 9.

Images