Hypergraph-based collaborative filtering recommendation

Abstract

An electronic device and method for performing hypergraph-based collaborative filtering recommendations are provided. 【Solution means】The electronic device receives a collaborative filtering graph corresponding to a user set and an item set. The electronic device determines a first user embedding set and a first item embedding set. The electronic device applies a semantic clustering model to determine a second user embedding set and a second item embedding set. The electronic device constructs a hypergraph to determine a third user embedding set and a third item embedding set. The electronic device determines a first contrast loss and a second contrast loss to determine a collaborative filtering score. The electronic device determines item recommendations for a user based on the determined collaborative filtering score. The electronic device renders the determined recommended items on a display device.

Description

Technical Field

[0001] 〔Cross - Reference to Related Applications / Incorporation by Reference〕 This application claims the benefit of priority of U.S. Patent Application No. 18 / 319,096, filed on May 17, 2023, which claims the priority of U.S. Provisional Patent Application Serial No. 63 / 365,540, filed on May 31, 2022, and the contents of these documents are hereby incorporated by reference in their entirety.

[0002] Various embodiments of the present disclosure relate to a recommendation system. Specifically, various embodiments of the present disclosure relate to an electronic device and method for hypergraph - based collaborative filtering recommendations.

Background Art

[0003] Advances in the field of recommendation systems have led to the development of various types of recommendation models that can provide personalized recommendations to users. Recommendation systems can be used in a variety of fields such as media and entertainment, finance, e - commerce, retail, banking, telecommunications, etc. Typically, a recommendation system can recommend items (such as movies) related to a certain domain (such as the movie domain for an over - the - top platform) to a user based on parameters such as the user's personal details / profile, the user's viewing history, movie consumption patterns (e.g., the time spent watching each movie), and the genre of movies in the viewing history. Conventional recommendation systems may ignore the higher - order relationships between users and items. Therefore, conventional recommendation systems may not be optimal and often make inaccurate recommendations.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Those skilled in the art will appreciate the limitations and disadvantages of conventional and customary techniques by comparing the described system with some aspects of the present disclosure shown with reference to the drawings in the remainder of this application. **Means for Solving the Problem**

[0005] Provided are an electronic device and method for hypergraph-based collaborative filtering recommendation that are illustrated and / or described in relation to at least substantially one figure and are further shown more fully in the claims.

[0006] These and other features and advantages of the present disclosure can be understood by considering the following detailed description of the present disclosure with reference to the accompanying drawings, in which like elements are denoted by like reference numerals throughout. **Brief Description of the Drawings**

[0007]

Figure 1

Figure 2

Figure 3

Figure 4A

Figure 4B

Figure 5

Figure 6

Figure 7

Figure 8

DETAILED DESCRIPTION OF THE INVENTION

[0008] In an electronic device and method for hypergraph-based collaborative filtering recommendation, embodiments described below can be found. Exemplary aspects of the present disclosure can provide an electronic device capable of receiving a collaborative filtering graph corresponding to a user set and an item set related to the user set. The collaborative filtering graph can correspond to user-item interaction data. The electronic device can determine a first user embedding set and a first item embedding set based on the received collaborative filtering graph. The electronic device can apply a semantic clustering model to each of the determined first user embedding set and the determined first item embedding set. The electronic device can determine a second user embedding set and a second item embedding set based on the application of the semantic clustering model. The electronic device can construct a hypergraph from the received collaborative filtering graph. The electronic device can determine a third user embedding set and a third item embedding set based on the constructed hypergraph. The electronic device can determine a first contrastive loss based on the determined second user embedding set and the determined third user embedding set. The electronic device can determine a second contrastive loss based on the determined second item embedding set and the determined third item embedding set. Further, the electronic device can determine a collaborative filtering score based on the determined first contrastive loss and the determined second contrastive loss. Thereafter, the electronic device can determine an item recommendation for the user based on the determined collaborative filtering score. The electronic device can render the determined recommended items on a display device.

[0009] Typically, a recommendation system can recommend items related to a certain domain based on one or more parameters such as personal details related to the target user (e.g., age, gender, and demographic information, etc.), item consumption history, item consumption patterns, and the similarity between the items to be recommended and the items consumed by the target user. In some other typical recommendation systems, user embeddings can be generated based on features extracted from one or more parameters. The recommendation system can generate embeddings related to items (e.g., movies) based on the features of the domain (e.g., genre, length, cast, and studio, etc.). The recommendation system can compare the embeddings of items in the item consumption history of the target user with the items in the domain. The recommendation system can recommend items in the domain related to embeddings similar to the embeddings of items in the item consumption history.

[0010] Furthermore, in some conventional recommendation systems for content recommendation on OTT platforms or streaming services, regular bipartite graphs can be provided as input. Such a bipartite graph can include a series of edges connecting pairs of nodes. Furthermore, the bipartite graph can also provide only inter-domain correlations (e.g., user-item correlations). Conventional recommendation systems can also learn intra-domain similarities (e.g., user-user correlations or item-item correlations) at the same time. Generalization of such intra-domain similarities can be difficult in some cases. Furthermore, since most users may not interact with all items in the item set, the data related to intra-domain similarities can be sparse. Therefore, the distribution of edge types can become highly unbalanced. Therefore, the recommendation system may not be optimal.

[0011] To address the above problems, the disclosed electronic device can adopt a hypergraph-based collaborative filtering framework for item recommendation. Here, the electronic device can apply a semantic clustering model to each of the determined first user embedding set and the first item embedding set to determine a second user embedding set and a second item embedding set. The electronic device can obtain positive samples and negative samples based on the application of the semantic clustering model. Further, the electronic device can construct a hypergraph from the received collaborative filtering graph. Using the constructed hypergraph, higher-order relationships between the user set and the item set can be explored. Further, the electronic device can determine a third user embedding set and a third item embedding set based on the constructed hypergraph. The determined third user embedding set and the determined third item embedding set can include features related to the potential relationships between the user set and the item set captured in the constructed hypergraph. The electronic device can adopt a contrastive framework to determine a first contrastive loss and a second contrastive loss for determining recommendations. In some embodiments, the electronic device can determine the final user embedding and the final item embedding that can take into account the higher-order relationships captured in the constructed hypergraph such that non-structural but similar nodes (e.g., user set and item set) are placed close to each other and dissimilar nodes are placed far apart. The final user embedding and the final item embedding can maintain a balance between the higher-order view and the collaborative view of the interaction data inferred from the collaborative filtering graph. The balance between the final user embedding and the final item embedding can help achieve optimal results in downstream tasks such as recommendation systems, user clustering, community clustering, classification tasks, etc.

[0012] FIG. 1 is a block diagram showing an exemplary network environment for hypergraph-based collaborative filtering recommendation according to an embodiment of the present disclosure. FIG. 1 shows a network environment 100. The network environment 100 can include an electronic device 102, a server 104, a database 106, and a communication network 108. The electronic device 102 can include a semantic clustering model 110, a recommendation model 112, a graph neural network (GNN) model 114, a first hypergraph convolutional network (HGCN) model set 116A, and a second HGCN model set 116B. FIG. 1 further shows a collaborative filtering graph 118 that can be stored in the database 106. Further shown is a user 120 who can be associated with or operate the electronic device 102.

[0013] The electronic device 102 can include suitable logic, circuitry, interfaces, and / or code configured to receive a collaborative filtering graph 118 corresponding to a user set and an item set related to the user set. The electronic device 102 can receive the collaborative filtering graph 118 from the server 104 via the database 106 (which can store the collaborative filtering graph 118). The electronic device 102 can determine a first user embedding set and a first item embedding set based on the received collaborative filtering graph 118. The electronic device 102 can apply a semantic clustering model 110 to each of the determined first user embedding set data and the determined first item embedding set data. The electronic device 102 can determine a second user embedding set and a second item embedding set based on the application of the semantic clustering model 110. The electronic device 102 can construct a hypergraph from the received collaborative filtering graph 118. The electronic device 102 can determine a third user embedding set and a third item embedding set based on the constructed hypergraph. The electronic device 102 can determine a first contrastive loss based on the determined second user embedding set and the determined third user embedding set. The electronic device 102 can determine a second contrastive loss based on the determined second item embedding set and the determined third item embedding set. For example, the electronic device 102 can determine the first contrastive loss based on the determined first user embedding set having spectral similarity and a local collaborative graph, the second user embedding set from the hypergraph, and the determined third item embedding set. The electronic device 102 can determine the second contrastive loss based on the determined first item embedding set from the local collaborative graph grouped by semantic similarity and the determined second item embedding set from the hypergraph.Furthermore, the electronic device 102 can determine a collaborative filtering score based on the determined first control loss and the determined second control loss. Thereafter, the electronic device 102 can determine a recommendation of an item for a user (e.g., user 120) based on the determined collaborative filtering score. The electronic device 102 can render the determined recommended item on a display device.

[0014] Examples of the electronic device 102 include, but are not limited to, a computer device, a smartphone, a cellular phone, a mobile phone, a game device, a mainframe machine, a server, a computer workstation, a machine learning device (e.g., corresponding to or hosting computing resources, memory resources, and networking resources), and / or a consumer electronics (CE) device.

[0015] Server 104 can include suitable logic, circuitry, interfaces, and / or code configured to receive from database 106 a collaborative filtering graph 118 corresponding to a user set and an item set associated with the user set. Server 104 can determine a first user embedding set and a first item embedding set based on the received collaborative filtering graph 118. Server 104 can apply semantic clustering model 110 to each of the determined first user embedding set data and the determined first item embedding set data. Server 104 can determine a second user embedding set and a second item embedding set based on the application of semantic clustering model 110. Server 104 can construct a hypergraph from the received collaborative filtering graph 118. Server 104 can determine a third user embedding set and a third item embedding set based on the constructed hypergraph. Server 104 can determine a first contrastive loss based on the determined second user embedding set and the determined third user embedding set. Server 104 can determine a second contrastive loss based on the determined second item embedding set and the determined third item embedding set. Server 104 can determine a collaborative filtering score based on the determined first contrastive loss and the determined second contrastive loss. Server 104 can determine recommendations for items for a user such as user 120 based on the determined collaborative filtering score. Server 104 can render the determined recommended items on a display device.

[0016] Server 104 can be implemented as a cloud server and can execute operations via, for example, web applications, cloud applications, HTTP requests, repository operations, and file transfers. Other implementation examples of server 104 include, but are not limited to, database servers, file servers, web servers, media servers, application servers, mainframe servers, machine learning servers (e.g., corresponding to or hosting computing resources, memory resources, and networking resources), or cloud computing servers.

[0017] In at least one embodiment, server 104 can be implemented as a plurality of distributed cloud-based resources by using a plurality of techniques well known to those skilled in the art. Those skilled in the art will understand that the scope of the present disclosure can also not be limited to the implementation of server 104 and electronic device 102 as two independent entities. In some embodiments, without departing from the scope of the present disclosure, the functions of server 104 can be wholly or at least partially incorporated into electronic device 102. In some embodiments, server 104 can host database 106. Alternatively, server 104 can be communicably coupled to database 106 separately from database 106.

[0018] Database 106 can include suitable logic, interfaces, and / or code configured to store collaborative filtering graph 118. Database 106 can also store information related to user sets and item sets. Database 106 can be derived from data in a relational or non-relational database, or from a set of comma-separated values (csv) files in conventional storage or big data storage. Database 106 can be stored or cached in a device such as a server (e.g., server 104) or electronic device 102. The device storing database 106 can be configured to receive queries for collaborative filtering graph 118 from electronic device 102 or server 104. In response, the device of database 106 can be configured to search collaborative filtering graph 118 that received the query based on the received query and provide it to electronic device 102 or server 104.

[0019] In some embodiments, database 106 can be hosted on multiple servers stored in the same or different locations. The operation of database 106 can be executed using hardware including a processor, a microprocessor (e.g., one or more that execute or control execution), a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC). In some other cases, software can also be used to implement database 106.

[0020] The communication network 108 can include a communication medium that enables the electronic device 102 and the server 104 to communicate with each other. The communication network 108 can be either a wired connection or a wireless connection. Examples of the communication network 108 include, but are not limited to, the Internet, a cloud network, a cellular or wireless mobile network (such as Long-Term Evolution and 5th Generation (5G) New Radio (NR)), a satellite communication system (such as using low-earth orbit satellites), a Wireless Fidelity (Wi-Fi) network, a Personal Area Network (PAN), a Local Area Network (LAN), or a Metropolitan Area Network (MAN). Various devices within the network environment 100 can be configured to connect to the communication network 108 according to various wired and wireless communication protocols. Examples of such wired and wireless communication protocols include, but are not limited to, Transmission Control Protocol and Internet Protocol (TCP / IP), User Datagram Protocol (UDP), Hypertext Transfer Protocol (HTTP), File Transfer Protocol (FTP), ZigBee, EDGE, IEEE802.11, Light Fidelity (Li-Fi), 802.16, IEEE802.11s, IEEE802.11g, multi-hop communication, wireless access point (AP), device-to-device communication, cellular communication protocols, and at least one of the Bluetooth (BT) communication protocol.

[0021] The semantic clustering model 110 can be a machine learning (ML) model that clusters an input data set into a set of clusters. Here, each cluster can include a subset of a similar data set. The semantic clustering model 110 of the present disclosure can be applied to each of the determined first user embedding set and the determined first item embedding set. The semantic clustering model 110 can determine a second user embedding set and a second item embedding set from the determined first user embedding set and the determined first item embedding set, respectively. In some embodiments, the semantic clustering model 110 can correspond to a spectral clustering model configured for dimensionality reduction. That is, here, the dimensions of the determined second user embedding set and the determined second item embedding set (determined based on the application of the semantic clustering model 110) can be smaller than the dimensions of the determined first user embedding set and the dimensions of the determined first item embedding set, respectively.

[0022] The recommendation model 112 can be an ML model that can determine recommendations based on various criteria. For example, the recommendation model 112 can recommend one or more products to a customer based on the customer's purchase history, the customer's geographical location, and the customer's needs, etc. The recommendation model 112 of the present disclosure can determine recommendations for items for user 120 based on the determined collaborative filtering scores.

[0023] The GNN model 114 can be a deep learning model that can construct a graph based on the received dataset. Subsequently, the GNN model 114 can process the constructed graph and make inferences based on the constructed graph. The GNN model 114 of the present disclosure can be applied to the received collaborative filtering graph 118. The GNN model 114 can process the applied collaborative filtering graph 118 to determine each of the first user embedding set and the first item embedding set.

[0024] The first HGCN model set 116A can be an ML model that can process information related to a hypergraph and determine an inference based on the processing. The first HGCN model set 116A can be applied to the fourth user embedding set. Here, the fourth user embedding set can be determined based on a series of user-item correlations and a series of user-user correlations, and these correlations can be determined based on the constructed hypergraph. The first HGCN model set 116A can determine the third user embedding set based on the determined fourth user embedding set. The second HGCN model set 116B can be applied to the fourth user embedding set. Here, the fourth item embedding set can be determined based on a series of item-user correlations, and these correlations can be determined based on the constructed hypergraph. The second HGCN model set 116B can determine the third item embedding set based on the determined fourth item embedding set.

[0025] The GNN model 114, the first HGCN model set 116A, and the second HGCN model set 116B can be graph neural network (GNN) models. The GNN model can include suitable logic, circuitry, interfaces, and / or code configured to classify or analyze input graph data to generate an output result for a particular real-time application. For example, a trained GNN model such as the GNN model 114 can recognize different nodes within the input graph data and edges between each node within the input graph data. The edges can correspond to different connections or relationships between each node within the input graph data. The trained GNN model 114 can classify different nodes within the input graph data into different labels or classes based on the recognized nodes and edges. In one example, a particular node of the input graph data can include an associated feature set. The feature set can include, but is not limited to, media content type, length of media content, genre of media content, and geographical location of the user 120. Further, each edge can connect different nodes having a similar feature set. The electronic device 102 can be configured to use the GNN model to encode the feature set to generate a feature vector. After encoding, information can be passed between a particular node connected via an edge and neighboring nodes. Based on the information passed to the neighboring nodes, a final vector can be generated for each node. Such a final vector can provide reliable and accurate information related to a particular node by including information related to the feature sets of the particular node and neighboring nodes. As a result, the GNN model can analyze the information represented as the input graph data. The GNN model can be implemented using hardware including a processor, a microprocessor (e.g., one or more that execute or control execution of operations), a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC). In some other instances, the GNN model can be code, a program, or a software instruction set.The GNN model can also be implemented using a combination of hardware and software.

[0026] In some embodiments, the GNN model can correspond to a plurality of classification layers where each successive layer can use the output of the previous layer as input, and which classify different nodes within the input graph data. Each classification layer can be associated with a plurality of edges, and each edge can in turn be associated with a plurality of weights. During training, the GNN model can be configured to filter or remove edges or nodes based on the input graph data and further provide an output result of the GNN model (i.e., a graph representation). Examples of GNN models can include, but are not limited to, graph convolutional networks (GCNs), hypergraph convolutional networks (HGCNs), graph spatio-temporal networks with GCNs, recurrent neural networks (RNNs), deep Bayesian neural networks, and / or combinations of such networks.

[0027] In one embodiment, the semantic clustering model 110, the recommendation model 112, the GNN model 114, the first set of HGCN models 116A, and the second set of HGCN models 116B can be machine learning (ML) models. Each ML model can be trained to identify relationships between inputs such as features within a training data set and output labels. Each ML model can be defined by hyperparameters such as, for example, the number of weights, cost function, input size, and number of layers. The parameters of each ML model can be adjusted to update the weights towards a global minimum of the cost function of the corresponding ML model. Each ML model can be trained to output recommendations, predictions, information related to a cluster set, or classification results for a series of inputs after being trained for a plurality of epochs on the feature information within the training data set. For example, the ML model associated with the recommendation model 112 can recommend items for the user 120.

[0028] Each ML model can include electronic data that can be implemented as a software component of an application executable, for example, on the electronic device 102. Each ML model can rely on libraries, external scripts, or other logic / instructions for execution by a processing device. Each ML model can include code and routines configured to enable a computer device, such as the electronic device 102, to perform one or more operations, such as making a recommendation decision. In addition to or instead of this, each ML model can also be implemented using hardware including a processor, a microprocessor, a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC). Alternatively, in some embodiments, a combination of hardware and software can be used to implement the ML model.

[0029] The collaborative filtering graph 118 can provide a compact representation of the interactions between a user set and an item set. The user set and the item set can be represented by a user node set and an item node set, respectively. Each edge of the collaborative filtering graph 118 can provide an interaction between a pair of nodes. Thus, the collaborative filtering graph 118 can be a bipartite graph. Details regarding the collaborative filtering graph 118 are further shown in FIG. 3.

[0030] During operation, the electronic device 102 can receive a collaborative filtering graph 118 corresponding to a user set and an item set related to the user set. For example, the database 106 can store the collaborative filtering graph 118. The electronic device 102 can request the collaborative filtering graph 118 from the database 106 and receive the requested collaborative filtering graph 118 from the database 106 via the server 104. The collaborative filtering graph 118 can be a bipartite graph that can represent various interactions between the user set and the item set. The user set and the item set can be represented as nodes of the collaborative filtering graph 118. Each edge of the collaborative filtering graph 118 can represent an interaction between a pair of nodes of the collaborative filtering graph 118. For example, it can be assumed that the user "A" has bookmarked the item "B". Therefore, the collaborative filtering graph 118 can include an edge between the user "A" and the item "B" that represents that the user "A" has bookmarked the item "B". Details regarding the collaborative filtering graph 118 will be further described, for example, in FIG. 3.

[0031] Based on the received collaborative filtering graph 118, the electronic device 102 can determine a first user embedding set and a first item embedding set. It can be understood that the embedding can correspond to a vector representation of features related to an entity. Each user embedding in the first user embedding set can provide features related to an item subset from the item set that the user associated with the corresponding user embedding is considered to have viewed or selected. Each item embedding in the first item embedding set can correspond to features related to a user subset from the user set that the item associated with the corresponding item embedding is considered to have been viewed or selected by.

[0032] In the neighborhood aggregation phase, the collaborative filtering graph 118 can be used to generate a first set of user embeddings and a first set of item embeddings having a plurality of "k" hops. Further, local collaborative signals can be a technique for dealing with user-item interactions in a way that can make hypergraph signals appear like global signals. In certain embodiments, the process of generating the first set of user embeddings and the first set of item embeddings having a plurality of "k" hops described above can be iteratively executed separately for odd hops and even hops. For example, in the first hop, the user embedding associated with user "U1" can be represented by vectors such as items "I1" and "I2". In the third hop, additional items can be added from the collaborative filtering graph 118 in the user embedding associated with user "U1". Similarly, in the even hops, each item can be associated with a plurality of users (users who are thought to have interacted with the item in some way). Thus, the first hop aggregation can include user "U1" represented as a vector from the perspective of directly connected items such as "I1" and "I2". The second hop can help represent an item as a vector from the perspective of users who can be directly or indirectly connected to an item. For example, user "U1" can be represented by an aggregation of items such as "I1" and "I2" with which user "U1" can directly interact on the first hop. However, if item "I1" is also connected to user "U2" and user "U2" is connected to item "I5", there can also be an indirect connection between user "U1" and item "I5". The relationships described above can be aggregated on the third hop. Details regarding the determination of the first set of user embeddings and the first set of item embeddings are further described, for example, in FIG. 4A.

[0033] The electronic device 102 can apply the semantic clustering model 110 to each of the determined first user embedding set data and the determined first item embedding set data. Based on the application of the semantic clustering model 110, a semantic view of the user set and the item set can be determined. User subsets and item subsets that can be directly connected to each other are considered similar and can be grouped together to form clusters. Details regarding the application of the semantic clustering model are further described, for example, in FIG. 4A.

[0034] The electronic device 102 can determine a second user embedding set and a second item embedding set based on the application of the semantic clustering model 110. The second user embedding set and the second item embedding set can be extracted from the semantic view of the user set and the item set. Details regarding the determination of the second user embedding set and the second item embedding set are further described, for example, in FIG. 4A.

[0035] The electronic device 102 can construct a hypergraph from the received collaborative filtering graph 118. The hypergraph can be a graph that represents the higher-order relationship between the user set and the item set related to the collaborative filtering graph 118 by hyperedges. Note that in an OTT platform, users are not always directly or indirectly connected to each other through item nodes. The collaborative filtering graph may be prone to information loss. To mitigate the above-described problems, a third user embedding set and a third item embedding set can be determined from the constructed hypergraph. Details regarding the construction of the hypergraph are further described, for example, in FIG. 5.

[0036] The electronic device 102 can determine a third user embedding set and a third item embedding set based on the constructed hypergraph. The third user embedding set and the third item embedding set determined in this way can include information related to the higher-order relationship between the user set and the item set. Furthermore, the third user embedding set and the third item embedding set can also include features related to the potential relationship between the user set and the item set captured in the constructed hypergraph. Details regarding the determination of the third user embedding set and the third item embedding set are further shown, for example, in FIG. 5.

[0037] The electronic device 102 can determine a first contrastive loss based on the determined second user embedding set and the determined third user embedding set. The first contrastive loss can be a variant form of nearest-neighbor contrastive learning (NNCLR) that can be determined based on the determined second user embedding set and the determined third user embedding set. Details regarding the determination of the first contrastive loss are further described, for example, in FIG. 4B.

[0038] The electronic device 102 can determine a second contrastive loss based on the determined second item embedding set and the determined third item embedding set. The second contrastive loss can be a variant form of NNCLR that can be determined based on the determined second item embedding set and the determined third item embedding set. Details regarding the determination of the second contrastive loss are further described, for example, in FIG. 4B.

[0039] The electronic device 102 can determine a collaborative filtering score based on the determined first control loss and the determined second control loss. The collaborative filtering score can provide a score set for each user's item set in the user set. The score set can be used as a basis for determining recommendations for the user set. Details regarding the determination of the collaborative filtering score will be further described, for example, in FIG. 4B.

[0040] The electronic device 102 can determine recommendations for items for user 120 based on the determined collaborative filtering score. For each user, the item that can be associated with the highest score can be selected as a recommendation. For example, for user 120, the score set can be assumed to be "0.78", "0.67", and "0.82". Therefore, the item associated with the score of "0.82" can be determined as a recommendation for user 120. Details regarding the determination of item recommendations will be further described, for example, in FIG. 4B.

[0041] The electronic device 102 can render the determined recommended items on a display device. In one example, the determined recommended item can be an action movie that can be displayed as a recommendation on the display device. As a result, user 120 can select and then play this action movie. Details regarding the rendering of the determined recommended items will be further described, for example, in FIG. 4B.

[0042] The electronic device 102 can adopt contrastive learning with positive and negative pair formation from hypergraph embedding, GCN collaborative structural embedding, and spectral cluster-based semantic embedding. The formation of positive pairs with a third user embedding set and a third item embedding set using the semantic clustering model 110 can help retain similar information for better learning. The electronic device 102 can be used to perform personalized recommendations on an over-the-top (OTT) platform and an e-commerce platform, etc. Here, the electronic device 102 can further process the recommendation task as a link prediction task or an edge prediction task for each item in the item set.

[0043] FIG. 2 is a block diagram showing an exemplary electronic device of FIG. 1 according to an embodiment of the present disclosure. The description of FIG. 2 is made in relation to the elements of FIG. 1. FIG. 2 shows an exemplary electronic device 102. The electronic device 102 can include a circuit 202, a memory 204, an input / output (I / O) device 206, a network interface 208, a semantic clustering model 110, a recommendation model 112, a GNN model 114, a first HGCN model set 116A, and a second HGCN model set 116B. The memory 204 can store a collaborative filtering graph 118. The input / output (I / O) device 206 can include a display device 210.

[0044] Circuit 202 can include suitable logic, circuitry, and / or interfaces configured to execute program instructions related to different operations performed by electronic device 102. These operations can include receiving a collaborative filtering graph, applying a GNN model, determining a first embedding, applying a semantic clustering model, determining a second embedding, constructing a hypergraph, determining a third embedding, determining a first contrastive loss, determining a second contrastive loss, determining a collaborative filtering score, determining a recommendation, and rendering the recommendation. Circuit 202 can include one or more processing units that can be implemented as a stand-alone processor. In some embodiments, one or more processing units can be implemented as an integrated processor or a group of processors that collectively execute the functions of the one or more processing units. Circuit 202 can be implemented based on a plurality of processor technologies well known in the art. Examples of implementations of circuit 202 can be an X86-based processor, a graphics processing unit (GPU), a reduced instruction set computing (RISC) processor, an application specific integrated circuit (ASIC) processor, a complex instruction set computing (CISC) processor, a microcontroller, a central processing unit (CPU), and / or other control circuitry.

[0045] Memory 204 can include suitable logic, circuitry, interfaces, and / or code configured to store one or more instructions executed by circuitry 202. The one or more instructions stored in memory 204 can be configured to perform different operations of circuitry 202 (and / or electronic device 102). Memory 204 can be further configured to store collaborative filtering graph 118. In certain embodiments, memory 204 can also store user embeddings and item embeddings. Examples of implementations of memory 204 include, but are not limited to, random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), hard disk drive (HDD), solid state drive (SSD), CPU cache, and / or secure digital (SD) card, among others.

[0046] I / O device 206 can include suitable logic, circuitry, interfaces, and / or code configured to receive an input and provide an output based on the received input. For example, I / O device 206 can receive a first user input indicating a request to generate recommendations for items for user 120. I / O device 206 can be further configured to display or render the recommended items. I / O device 206 can include display device 210. Examples of I / O device 206 include, but are not limited to, a display (e.g., touch screen), keyboard, mouse, joystick, microphone, or speaker. Examples of I / O device 206 can further include braille I / O devices such as a braille keyboard and a braille reader.

[0047] The network interface 208 can include suitable logic, circuitry, interfaces, and / or code configured to facilitate communication between the electronic device 102 and the server 104 via the communication network 108. The network interface 208 can be implemented to support wired or wireless communication between the electronic device 102 and the communication network 108 using various known techniques. The network interface 208 can include, but is not limited to, an antenna, a radio frequency (RF) transceiver, one or more amplifiers, a tuner, one or more oscillators, a digital signal processor, a coder-decoder (CODEC) chipset, a subscriber identity module (SIM) card, or a local buffer circuit.

[0048] The network interface 208 can be configured to communicate wirelessly with a network such as the Internet, an intranet, a wireless network, a cellular telephone network, a wireless local area network (LAN), or a metropolitan area network (MAN). The wireless communication can use one or more of a plurality of communication standards, protocols, and technologies such as Global System for Mobile Communications (GSM), Enhanced Data GSM Environment (EDGE), Wideband Code Division Multiple Access (W-CDMA), Long Term Evolution (LTE), 5th Generation (5G) New Radio (NR), Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Bluetooth, Wireless Fidelity (WiFi) (such as IEEE802.11a, IEEE802.11b, IEEE802.11g, or IEEE802.11n), Voice over Internet Protocol (VoIP), Light Fidelity (Li-Fi), Worldwide Interoperability for Microwave Access (Wi-MAX), protocols for email, instant messaging, and Short Message Service (SMS).

[0049] The display device 210 can include suitable logic, circuitry, and interfaces configured to display or render the determined recommended items. The display device 210 can be a touch screen that enables a user (e.g., user 120) to provide user input via the display device 210. The touch screen can be at least one of a resistive touch screen, a capacitive touch screen, or a thermal touch screen. The display device 210 can be implemented through a plurality of known technologies such as, but not limited to, a liquid crystal display (LCD) display, a light emitting diode (LED) display, a plasma display, or an organic LED (OLED) display technology, or at least one of other display devices. According to an embodiment, the display device 210 can refer to the display screen of a head-mounted display (HMD), smart glasses, a see-through display, a projection-based display, an electrochromic display, or a transparent display. Various operations of the circuit 202 for performing hypergraph-based collaborative filtering recommendations will be further described, for example, in FIGS. 4A and 4B.

[0050] FIG. 3 is a diagram showing an exemplary scenario of a collaborative filtering graph according to an embodiment of the present disclosure. The description of FIG. 3 will be made in relation to the elements of FIGS. 1 and 2. An exemplary scenario 300 is shown in FIG. 3. The scenario 300 can include a user set and an item set. The user set can include a first user 302A, a second user 302B, and a third user 302C. The item set can include a first item 304A, a second item 304B, and a third item 304C. Here, a series of operations related to the scenario 300 will be described.

[0051] In scenario 300 of FIG. 3, an item set such as a first item 304A, a second item 304B, and a third item 304C can be different multimedia contents such as sitcoms, news reports, and digital games. First, a first user 302A, a second user 302B, and a third user 302C can be registered on an OTT platform. Each user in the user set can view one or more items in the item set and evaluate each of the one or more viewed items on a scale of "1" to "5". An evaluation of "1" can mean that the user considers the evaluated item to be not liked at all, and an evaluation of "5" can mean that the user considers the evaluated item to be very liked.

[0052] Referring to FIG. 3, the first user 302A can interact with the first item 304A and give an evaluation of "5" as indicated by edge 306A. The second user 302B can interact with the first item 304A and the second item 304B as indicated by edges 306B and 306C, respectively. Further, the second user 302B can evaluate the first item 304A as "5" and the second item 304B as "2". That is, the second user 302B can consider that the first item 304A is liked more than the second item 304B. The third user 302C can interact with the first item 304A, the second item 304B, and the third item 304C as indicated by edges 306D, 306E, and 306F, respectively. Further, the third user 302C can evaluate the first item 304A, the second item 304B, and the third item 304C as "5", "5", and "5", respectively. That is, the third user 302C can consider that the first item 304A, the second item 304B, and the third item 304C are liked equally.

[0053] Note that scenario 300 in FIG. 3 is for illustrative purposes and should not be construed as limiting the scope of the present disclosure.

[0054] FIGS. 4A and 4B are diagrams showing an exemplary processing pipeline for hypergraph-based collaborative filtering recommendation according to an embodiment of the present disclosure. The descriptions of FIGS. 4A and 4B are made in relation to the elements of FIGS. 1, 2, and 3. FIGS. 4A and 4B show an exemplary processing pipeline 400 showing exemplary operations 402 to 424 for performing hypergraph-based collaborative filtering recommendation. The exemplary operations 402 to 424 can be executed by any computer system such as, for example, the electronic device 102 of FIG. 1 or the circuit 202 of FIG. 2. FIGS. 4A and 4B further include a collaborative filtering graph 118, a GNN model 114, a first user embedding set 406A, a first item embedding set 406B, a second user embedding set 410A, a second item embedding set 410B, a third user embedding set 414A, and a third item embedding set 414B.

[0055] At 402, a collaborative filtering graph receiving operation can be executed. Circuit 202 can be configured to receive a collaborative filtering graph 118 corresponding to a user set and an item set related to the user set. Here, the item set can include different multimedia contents such as sitcoms, news reports, and digital games that can be related to the user set. The item set can also include various items such as clothes, electronic devices, game devices, and books that can be sold on an e-commerce application or website. It can be understood that there can be different types of interactions between the user set (such as the first user 302A, the second user 302B, and the third user 302C in FIG. 3) and the item set (such as the first item 304A, the second item 304B, and the third item 304C in FIG. 3). For example, different types of interactions can be selecting an item, adding an item to a digital cart, registering an item on a wish list on an e-commerce application, watching a video, bookmarking a video, or liking a video, or evaluating a video on an OTT platform. If there is only one type of interaction between the user set and the item set, the interaction between the user set and the item set can be represented as a graph called a bipartite graph. On the other hand, if the interactions between the user set and the item set are of different types, the graph formed in this way can be qualitatively heterogeneous and can form a multiplex bipartite graph. The collaborative filtering graph 118 can be a bipartite graph or a multiplex bipartite graph formed based on the interactions between the user set and the item set. Details regarding the collaborative filtering graph are further shown, for example, in FIG. 3.

[0056] At 404, an application operation of the GNN model 114 with respect to the received collaborative filtering graph 118 can be executed. The circuit 202 can be configured to apply the GNN model 114 to the received collaborative filtering graph 118. The GNN model 114 can process the received collaborative filtering graph 118 to derive information related to each user and each item. In some embodiments, the GNN model 114 can be a graph convolutional network (GCN) model.

[0057] At 406, a determination operation of the first user embedding set 406A and the first item embedding set 406B can be executed. The circuit 202 can be configured to determine the first user embedding set 406A and the first item embedding set 406B. Here, each of the first user embedding set 406A and the first item embedding set 406B can be determined based on the application of the GNN model 114. An embedding can correspond to a vector representation of features related to an entity. For example, each of the first user embedding set 406A can correspond to features related to an item subset from a set of items that the corresponding user is considered to have viewed or selected. Each item embedding of the first item embedding set 406B can correspond to features related to a user subset from a set of users that are considered to have viewed or selected the item.

[0058] Referring to FIG. 3, it can be assumed that the third user 302C evaluated the first item 304A, the second item 304B, and the third item 304C as "5", "5", and "5", respectively. Therefore, the third user 302C user embedding can include the identification numbers of the items that the third user 302C could evaluate as "5". That is, the user embedding of the third user 302C can include the identification numbers of the first item 304A, the second item 304B, and the third item 304C. Further, the user embedding of the third user 302C can include identification numbers such as item type, genre, video length, and language related to the first item 304A, the second item 304B, and the third item 304C. Similarly, for each evaluation provided by each of the first user 302A and the second user 302B, user embeddings related to the first user 302A and the second user 302B can be determined. Further, referring to FIG. 3, it can be recognized that the third item 304C is evaluated as "5" only by the third user 302C. Therefore, the item embedding of the third item 304C can include information such as the name, identification, and geographical location of the third user 302C. Similarly, for each evaluation provided by each of the first user 302A, the second user 302B, and the third user 302C, item embeddings related to the first item 304A and the second item 304B can be determined. In this way, the first user embedding set 406A and the first item embedding set 406B can be determined.

[0059] Referring again to FIG. 4, at 408, a semantic clustering model application operation can be executed. The circuit 202 can be configured to apply the semantic clustering model 110 to each of the determined first user embedding set 406A and the determined first item embedding set 406B.

[0060] In one embodiment, the semantic clustering model 110 can correspond to a spectral clustering model configured for dimensionality reduction of each of the first user embedding set 406A and the first item embedding set 406B. The spectral clustering model can be understood to be a clustering mechanism that can perform dimensionality reduction of the input data set by leveraging the spectrum, such as the eigenvalues of the similarity matrix of the input data set, before clustering the input data set into a lower dimension. The input data set of the present disclosure can include each of the first user embedding set 406A and the first item embedding set 406B.

[0061] Spectral clustering algorithms related to the spectral clustering model can project the input data set into the "Rn" matrix that may be required for clustering into "k" clusters. To construct an affinity matrix based on the projected input data set, a Gaussian kernel matrix "K" or an adjacency matrix "A" can be created. Note that in the spectral clustering algorithm, it can be understood that a Gaussian kernel function can be used to measure similarity. The adjacency matrix "A" can be a representation of the projected input data set. A series of rows related to the adjacency matrix "A" can represent the first user set, and a series of columns related to the adjacency matrix "A" can represent the first item set. Each entry in the adjacency matrix "A" can provide information about the interaction between a user and an item. In an example, the entry in the first row and first column of the adjacency matrix "A" can be "1". Therefore, it can be assumed that the first user related to the first row has viewed or selected the first item related to the first column of the adjacency matrix "A". Further, in an example, the entry in the first row and second column of the adjacency matrix "A" can be "0". Therefore, it can be assumed that the first user related to the first row has not viewed or selected the second item related to the second column of the adjacency matrix "A". An affinity matrix can be constructed based on the created Gaussian kernel matrix "K" or adjacency matrix "A". The affinity matrix can also be called a similarity matrix and provides information related to how similar pairs of entities are to each other. If the entry related to a pair of entities in the affinity matrix is "0", the corresponding pair of entities can be dissimilar. If the entry related to a pair of entities is "1", the corresponding pair of entities can be similar. In other words, each entry in the affinity matrix can correspond to the weight of the edge related to a pair of entities. Based on the constructed affinity matrix, a graph Laplacian matrix "L" can be created.It can be understood that the graph Laplacian matrix "L" can be obtained based on the difference between the adjacency matrix "A" and the degree matrix. Once the graph Laplacian matrix "L" is determined, the eigenvalue problem can be corrected. The advantage of using the graph Laplacian matrix "L" is that it can determine how well the clusters are connected to each other based on the minimum eigenvalue of the graph Laplacian matrix "L". A low value can mean that the connection of the clusters is weak, and since different clusters can have weak connections, this is particularly useful. A k-dimensional subspace can be established based on the selection of "k" eigenvectors corresponding to the "k" lowest (or highest) eigenvalues. Then, the "k-means" clustering algorithm can be used to create clusters within the k-dimensional subspace. Details regarding spectral clustering are further shown, for example, in FIG. 6.

[0062] In 410, a second embedding determination operation can be performed. Circuit 202 can be configured to determine a second user embedding set 410A and a second item embedding set 410B based on the application of the semantic clustering model 110. A cluster set can be determined based on the application of the semantic clustering model 110. The second user embedding set 410A and the second item embedding set 410B can be extracted from this cluster set. The determination of the second user embedding set and the second item embedding set is further described, for example, in FIG. 6.

[0063] At 412, a hypergraph construction operation can be performed. Circuit 202 can be configured to construct a hypergraph from the received collaborative filtering graph 118. The hypergraph can be a graph that can represent a higher-order relationship between a set of users and a set of items related to the collaborative filtering graph 118 by using hyperedges. It can be understood that the regular edges of the graph can indicate the interaction between pairs of nodes, and thus can ignore the information between one node type and the potential representation of the node type having other node types. In one example, the received collaborative filtering graph 118 can indicate that user "A" is considered to like movie "X". Such information can be incorporated into the embedding space, for example, using the first user embedding set 406A and the first item embedding set 406B. However, due to the nature of the received collaborative filtering graph 118 and the sparsity of the information, the embedding space may not include information related to other items that user "A" is considered not to have interacted with. For example, user "A" is considered to have interacted with movie "X" and not with other movies. Therefore, in the hypergraph, a special type of edge called a hyperedge that can connect multiple nodes within the "n dimensions" can be used. Details regarding the hypergraph are further shown, for example, in FIG. 5.

[0064] At 414, a third embedding determination operation can be performed. Circuit 202 can be configured to determine a third user embedding set 414A and a third item embedding set 414B based on the constructed hypergraph. Details regarding the determination of the third user embedding set 414A and the third item embedding set 414B are further shown, for example, in FIG. 5.

[0065] At 416, a first contrastive loss determination operation can be performed. Circuit 202 can be configured to determine a first contrastive loss based on the determined second user embedding set 410A and the determined third user embedding set 414A. The first contrastive loss can be a variant form of NNCLR. Here, the nearest neighbor operator can be replaced with a cluster of similar nodes of each type, and a hypergraph embedding of similar users can be used instead of the augmented view. NNCLR can be obtained according to the following equation (1), TIFF2025519073000002.tif17150(1) where "τ" can be the SoftMax temperature, "X ui " can be the third user embedding related to user "i", "Z ui*,j " can be the second embedding of the most similar user "i * " of user "i" obtained from cluster "j".

[0066] At 418, a second contrastive loss determination operation can be performed. Circuit 202 can be configured to determine a second contrastive loss based on the determined second item embedding set 410B and the determined third item embedding set 414B. The second contrastive loss can be similar to the first contrastive loss and can be determined according to the following equation (2), TIFF2025519073000003.tif17150(2) where "τ" can be the SoftMax temperature, "X ui " can be the third item embedding related to item "i", "Z vi*,j " can be the second embedding of the most similar item "i * " of item "i" obtained from cluster "j".

[0067] At 420, a collaborative filtering score determination operation can be executed. Circuit 202 can be configured to determine a collaborative filtering score based on the determined first contrast loss and the determined second contrast loss. The collaborative filtering score can provide a set of scores for the item set of each user in the user set. The set of scores can follow the preferences, past interactions, and selections of the user set.

[0068] In an embodiment, circuit 202 can be further configured to determine a fifth user embedding set based on the first contrast loss and a third user embedding set. Circuit 202 can be further configured to determine a fifth item embedding set based on the second contrast loss and a third item embedding set. The fifth user embedding set can provide a vector representation of features related to the user set. The fifth item embedding set can provide a vector representation of features related to the item set.

[0069] In an embodiment, circuit 202 can be further configured to determine a final user embedding based on the determined fifth user embedding set. Circuit 202 can be further configured to determine a final item embedding based on the determined fifth item embedding set. Here, the determination of the collaborative filtering score can be based on the determined final user embedding and the determined final item embedding.

[0070] In one example, a user set can interact with an item set by bookmarking an item, partially viewing an item, and fully viewing an item. Accordingly, the determined fifth user embedding set can include a fifth user embedding related to bookmarking of an item subset from each user's item set, a fifth user embedding related to partial viewing of an item subset from the item set, and a fifth user embedding related to full viewing of an item subset from the item set. Similarly, the determined fifth item embedding set can include, for each item, a fifth item embedding related to bookmarking of a corresponding item by a user subset from the user set, a fifth user embedding related to partial viewing of the corresponding item by a user subset from the user set, and a fifth user embedding related to full viewing of the corresponding item by a user subset from the user set. The final user embedding of a user, such as User 120, can be determined based on a combination of the determined fifth user embeddings of the corresponding user. That is, a fifth user embedding related to bookmarking of a user such as User 120, a fifth user embedding related to partial viewing, and a fifth user embedding related to full viewing can be combined to determine the final user embedding of the corresponding user. Similarly, the final item embedding of an item can be determined based on a combination of the determined fifth item embeddings of the corresponding item. That is, a fifth item embedding related to bookmarking of the corresponding item, a fifth item embedding related to partial viewing, and a fifth item embedding related to full viewing can be combined to determine the final item embedding. In one embodiment, the final user embedding and the final item embedding can be applied to a graph neural network (GNN) model or a natural language processing (NLP) model to generate a recommendation probability for the item set.

[0071] In one embodiment, each of the determined final user embedding and the determined final item embedding can correspond to a connection of at least one of a collaborative view, a hypergraph view, or a semantic view. Note that the collaborative view of each of the determined final user embedding and the determined final item embedding can be related to a first user embedding set 406A and a first item embedding set 406B, respectively. The hypergraph view can also be called a higher-order view. The hypergraph view of each of the determined final user embedding and the determined final item embedding can be related to a second user embedding set 410A and a second item embedding set 410B, respectively. The semantic view of each of the determined final user embedding and the determined final item embedding can be related to a third user embedding set 414A and a third item embedding set 414B, respectively. The determined final user embedding can be related to the first user embedding set 406A, the second user embedding set 410A, and the third user embedding set 414A. Similarly, the determined final item embedding can be related to the first item embedding set 406B, the second item embedding set 410B, and the third item embedding set 414B. Therefore, each of the determined final user embedding and the determined final item embedding can correspond to a connection of at least one of a collaborative view, a hypergraph view, or a semantic view.

[0072] At 422, a recommended decision operation can be executed. Circuit 202 can be configured to determine recommendations for items for user 120 based on the determined collaborative filtering scores. In one embodiment, the collaborative filtering scores can provide a set of scores for each user in a set of users with respect to a set of items. For each user, an item that may be associated with the highest score can be selected as a recommendation. In one example, the set of users includes user "A", user "B", and user "C", and the set of items includes item "X", item "Y", and item "Z". For user "A", the set of scores can include "0.1", "0.5", and "0.7" associated with item "X", item "Y", and item "Z", respectively. In such a case, since item "Z" has the highest score for user "A", item "Z" can be determined as a recommendation for user "A".

[0073] At 424, a rendering operation of the recommended item can be executed. Circuit 202 can be configured to render the determined recommended item on display device 210. In one example, the determined recommended item can be movie "X". The recommended movie "X" can be displayed on display device 210 to notify user 120 associated with electronic device 102. Thereafter, movie "X" can be played based on a user input related to the selection of movie "X" from user 120.

[0074] FIG. 5 is a diagram showing an exemplary scenario of an architecture for hypergraph embedding according to an embodiment of the present disclosure. The description of FIG. 5 is made in relation to the elements of FIGS. 1, 2, 3, 4A, and 4B. FIG. 5 shows an exemplary scenario 500. The scenario 500 can include a hypergraph 502, a fourth user embedding 504A, a fourth user embedding 504B, a first hypergraph convolutional network (HGCN) model 506A, a first HGCN model 506B, a third user embedding 508A, a third user embedding 508B, a fourth item embedding 510A, a fourth item embedding 510B, a second HGCN model 512A, a second HGCN model 512B, a third item embedding 514A, and a third item embedding 514B. Here, a series of operations related to the scenario 500 will be described.

[0075] In the scenario 500, the hypergraph 502 can be constructed based on the received collaborative filtering graph (e.g., the collaborative filtering graph 118 in FIG. 4A). In some embodiments, the constructed hypergraph 502 can correspond to a multi-bipartite graph having homogenous edges. The constructed hypergraph 502 can be a multi-bipartite graph because it can show multiple types of interactions between a set of users and a set of items. Further, the constructed hypergraph 502 can be formed such that one hyperedge represents one type of interaction.

[0076] In one embodiment, the first edge type in hypergraph 502 can correspond to an interaction between a first user and a first subset of items associated with the first user. The second edge type in the hypergraph can correspond to an interaction between a second subset of users and a second item associated with each of the second subset of users. For example, the first hyper-edge type can be formed to indicate a subset of items that are considered to have been rated "1" by the first user. Another first hyper-edge type can be formed to indicate a subset of items that are considered to have been rated "2" by the first user. The second hyper-edge can be formed to indicate a subset of users that are considered to have rated the first item "1". Another second hyper-edge can be formed to indicate a subset of users that are considered to have rated the first item "2".

[0077] Note that the homogeneous hypergraph constructed based on the first hyper-edge type can be defined according to the following equation (3), TIFF2025519073000004.tif11157 and TIFF2025519073000005.tif6150(3) where, "G U,base " can be a homogeneous graph, "U" can be a set of users, and "E U,i " can be a set of the first hyper-edge types.

[0078] Note that the homogeneous hypergraph constructed based on the second hyper-edge type can be defined according to the following equation (4), TIFF2025519073000006.tif16157 and TIFF2025519073000007.tif6150(4) where, "G U,base " can be a homogeneous graph, "I" can be a set of items, and "EI,j " can be the second hyper-edge type set.

[0079] Note that for the hypergraph 502, an incidence matrix "H" can be used for the user set "U". The connection matrix for the user set "U" can be defined according to the following equation (5), TIFF2025519073000008.tif9150 and TIFF2025519073000009.tif6150(5) Here, "E U,i " can be the first hyper-edge type set, and "i" can indicate the constructed hypergraph. Similarly, the connection matrix of the item set "I" can be defined as "H I,j (i,e)".

[0080] In one embodiment, circuit 202 can be further configured to determine a user-item correlation set, an item-user correlation set, and a user-user correlation set based on the constructed hypergraph 502. The user-item correlation set can be determined based on a first edge type and can represent the relationship between a user and an item. For example, a first user-item correlation can provide information related to a set of items that a first user is considered to have fully watched. A second user-item correlation can provide information related to a set of items that are considered to be selected based on a first user. The item-user correlation set can be determined based on a second edge type and can provide information related to the relationship between an item and a user. For example, a first item-user correlation can represent a set of users who are considered to have fully watched a first item. A second item-user correlation can provide information related to a set of users who are considered to be selected based on a first item. The user-user correlation set can be determined based on the first edge type and the second edge type and can provide information related to the potential relationship between users. For example, a first user can fully watch movie "X". Similarly, a second user can also fully watch movie "X". Here, a relationship can exist between the first user and the second user. A user-user correlation can be determined to capture the relationship described above.

[0081] Circuit 202 can be further configured to determine a fourth user embedding set (e.g., fourth user embedding 504A) based on the determined user-item correlation set and user-user correlation set. The fourth user embedding set can include one or more user embeddings for each user. Each of the one or more user embeddings associated with a user can correspond to one interaction type. For example, referring to FIG. 5, the first interaction type can be related to fully viewing one or more items, and the second interaction type can be related to partially viewing one or more items. The fourth user embedding 504A can be formed based on the user-item correlation and user-user correlation sets corresponding to the first interaction type associated with the first user. The fourth user embedding 504B can be formed based on the user-item correlation and user-user correlation sets corresponding to the second interaction type associated with the first user.

[0082] When the fourth user embedding set is determined, circuit 202 can be further configured to apply a first HGCN model set (e.g., the first HGCN model set 116A in FIG. 1) to the determined fourth user embedding set (e.g., fourth user embedding 504A). An HGCN model from the first HGCN model set can be applied to each of the fourth user embedding sets. The first HGCN model set (e.g., the first HGCN model set 116A in FIG. 1) can be an ML model that processes information related to hypergraph 502 and determines an inference based on the processing.

[0083] The convolutional operator associated with the first HGCN model set (e.g., the first HGCN model set 116A in FIG. 1) for the constructed hypergraph 502 can be defined according to the following equation (6), TIFF2025519073000010.tif6150(6) Here, "σ" can be a non-linear activation function, "X" can be a feature matrix, and "P" can be a learnable weight matrix. Further, "HWH T " can be used to measure the pairwise relationships between nodes within the same homogeneous hypergraph, and "W" can be a weight matrix that can assign weights to all hyperedges.

[0084] The normalized versions of the symmetric and asymmetric convolution operators can be defined according to the following equations (7) and (8). TIFF2025519073000011.tif9150(7) And TIFF2025519073000012.tif6150(8) Referring to equation (7), "I" can be an identity matrix, and "D" can be a node degree matrix of a simple graph. Referring to equation (8), "σ" can be a non-linear activation function, TIFF2025519073000013.tif6150 can be the features of layer "l", "W U " ∈ "R|V|×|V|" can be an identity matrix, "P" can represent a learnable filter matrix, "D l " and "D l+1 " can be the dimensions of layer "l" and layer "l + 1", respectively.

[0085] For example, referring to FIG. 5, the first HGCN model 506A can be applied to the fourth user embedding 504A, and the first HGCN model 506B can be applied to the fourth user embedding 504B. A third set of user embeddings can be determined based on the application of the first set of HGCN models (e.g., the first set of HGCN models 116A in FIG. 1). For example, referring to FIG. 5, the third user embedding 508A of the first user can be determined based on the application of the first HGCN model 506A. Based on the application of the first HGCN model 506B, the third user embedding 508B of the first user can be determined. Similarly, for each interaction type, the fourth user embedding of each user in the user set can be determined.

[0086] The circuit 202 can be further configured to determine a fourth set of item embeddings (e.g., the fourth item embedding 510A) based on the determined set of item-user correlations. The fourth set of item embeddings can include one or more item embeddings for each item. One or more item embeddings associated with an item can each correspond to one interaction type. For example, referring to FIG. 5, the fourth item embedding 510A can be formed based on the item-user correlation corresponding to the first interaction type associated with the first item. The fourth item embedding 510B can be formed based on the item-user correlation corresponding to the second interaction type associated with the first item.

[0087] When the fourth item embedding set is determined, circuit 202 can be further configured to apply a second HGCN model set (e.g., second HGCN model set 116B) to the determined fourth item embedding set. An HGCN model can be applied to each fourth item embedding. The second HGCN model set (e.g., second HGCN model set 116B in FIG. 1) can be an ML model that processes information related to hypergraph 502 and can determine inferences based on the processing. For example, referring to FIG. 5, a second HGCN model 512A can be applied to the fourth item embedding 510A, and a second HGCN model 512B can be applied to the fourth item embedding 510B. A third item embedding set can be determined based on the application of the second HGCN model set (e.g., second HGCN model set 116B in FIG. 1). For example, referring to FIG. 5, based on the application of the second HGCN model 512A, a third item embedding 514A of the first item related to completely viewing the first item can be determined. Based on the application of the second HGCN model 512B, a third item embedding 514B of the first item related to selection based on the first item can be determined. Similarly, for each interaction type, a fourth item embedding of each item in the item set can be determined.

[0088] Note that scenario 500 in FIG. 5 is for illustrative purposes and should not be construed as limiting the scope of the present disclosure.

[0089] FIG. 6 is a diagram showing an exemplary scenario of contrastive learning according to an embodiment of the present disclosure. The description of FIG. 6 is made in relation to the elements of FIGS. 1, 2, 3, 4A, 4B, and 5. FIG. 6 shows an exemplary scenario 600. The scenario 600 can include a collaborative filtering graph 602, a graph convolutional network (GCN) 604, a semantic cluster 606 of user nodes and item nodes, a second user embedding 608, a hypergraph embedding block 610, and a third user embedding 612. Here, a series of operations related to the scenario 600 will be described. FIG. 6 describes contrastive learning for user embedding. However, the scenario 600 of FIG. 6 can be similarly applied to contrastive learning for item embedding without departing from the scope of the present disclosure.

[0090] It can be understood that self-supervised approaches commonly used in the field of computer vision can involve a process of determining the most discriminative representation of embeddings. In one example, a discriminative representation of embeddings can be determined by augmentation for a given set of different views of the same object within an image. In another example, a discriminative representation of embeddings can be obtained by the use of similar objects and comparison with other dissimilar objects. The contrastive learning approach described above can be extended to a recommendation system. Here, different augmentations of user-item interactions can be used. Different augmentations can be obtained based on, for example, node deletion, edge deletion, and node replication. The augmented node embeddings views in a mini-batch of interactions can form positive pairs, and the remaining embeddings from the mini-batch can form negative pairs.

[0091] For example, referring to FIG. 6, GCN 604 can be applied to the collaborative filtering graph 602. GCN 604 can be a generalized convolutional neural network that can adopt a semi-supervised based learning approach for the graph. Based on the application of GCN 604 to the collaborative filtering graph 602, a first user embedding set and a first item embedding set can be obtained. Further, based on the application of GCN 604 to the collaborative filtering graph 602, a semantic cluster of user nodes and item nodes 606 can be obtained. Thereafter, a second user embedding 608 can be obtained based on the semantic cluster of user nodes and item nodes 606. The second user embedding 608 can be related to similar users determined from the semantic cluster of user nodes and item nodes 606. Further, the collaborative filtering graph 602 can be applied to the hypergraph embedding block 610. The hypergraph embedding block 610 can include a first set of HGCN models (e.g., the first HGCN model 506A and the first HGCN model 506B of FIG. 5) and a second set of HGCN models (e.g., the second HGCN model 512A and the second HGCN model 512B of FIG. 5). Based on the application of the collaborative filtering graph 602 to the hypergraph embedding block 610, a third user embedding 612 can be obtained. The second user embedding 608 and the third user embedding 612 can be a positive pair of embeddings and can be used for contrastive learning purposes. The positive pair of embeddings can be used for contrastive learning purposes. Further, the negative samples can be samples that are not considered to be part of the cluster to which the user "U1" belongs.

[0092] Note that the scenario 600 of FIG. 6 is for illustrative purposes and should not be construed as limiting the scope of the present disclosure.

[0093] FIG. 7 is a diagram showing an exemplary scenario for recommending an item set to a user set according to an embodiment of the present disclosure. The description of FIG. 7 is made in relation to the elements of FIGS. 1, 2, 3, 4A, 4B, 5, and 6. FIG. 7 shows an exemplary scenario 700. The scenario 700 can include a hyper edge 702, a first user 704A, a second user 704B, a third user 704C, a first news channel 706, a final user embedding 708, a final item embedding 710, and a recommended item set 712. The recommended item set 712 can include a second news channel 712A, a third news channel 712B, and a fourth news channel 712C. Here, a series of operations related to the scenario 700 will be described.

[0094] In scenario 700 of FIG. 7, the second user 704B and the third user 704C can have a positive interest in the first news channel 706. For example, the second user 704B and the third user 704C can be assumed to have watched the first news channel 706. The first user 704A can be assumed not to have watched the first news channel 706. On the other hand, the first user 704A and the third user 704C can be assumed to have also watched a news channel (not shown) similar to the first news channel 706. Therefore, a potential relationship can exist between the first user 704A and the first news channel 706. Furthermore, a potential relationship can also exist between the first user 704A and the second user 704B. Therefore, the first user 704A, the second user 704B, and the third user 704C can form a hyperedge such as hyperedge 702 together with the first news channel 706. A plurality of hyperedges similar to hyperedge 702 can be formed to construct a hypergraph. Based on the constructed hypergraph, a third user embedding set can be determined. The third user embedding set (not shown) can include a third user embedding related to the first user 704A, a third user embedding related to the second user 704B, and a third user embedding related to the third user 704C. The third user embedding related to the first user 704A, the third user embedding related to the second user 704B, and the third user embedding related to the third user 704C can be similar to each other. Based on the determined third user embedding set, the final user embedding 708 and the final item embedding 710 can be obtained. For example, as shown in FIG. 7, the final user embedding 708 can be "0.87", "0.79", and "0.77" for the first user 704A, the second user 704B, and the third user 704C, respectively. The final user embedding 708 can correspond to the collaborative filtering scores related to the users 704A, 704B, and 704C.Furthermore, the final item embedding 710 can be "0.95" for the first item, "0.89" for the second item, and "0.87" for the third item, and these can be recommended to the first user 704A, the second user 704B, and the third user 704C, respectively. The final item embedding 710 can correspond to the collaborative filtering scores associated with the first item, the second item, and the third item. For example, based on the final user embedding 708 and the final item embedding 710, the second news channel 712A can be recommended to the first user 704A, the third news channel 712B can be recommended to the second user 704B, and the fourth news channel 712C can be recommended to the third user 704C. Note that the second news channel 712A, the third news channel 712B, and the fourth news channel 712C can be similar to each other.

[0095] Note that the scenario 700 in FIG. 7 is for illustrative purposes only and should not be construed as limiting the scope of the present disclosure.

[0096] FIG. 8 is a flowchart showing the operations of an exemplary method for hypergraph-based collaborative filtering recommendation according to an embodiment of the present disclosure. The description of FIG. 8 is made in relation to the elements of FIGS. 1, 2, 3, 4A, 4B, 5, 6, and 7. FIG. 8 shows a flowchart 800. The flowchart 800 can include operations 802 to 824 and can be implemented by the electronic device 102 of FIG. 1 or the circuit 202 of FIG. 2. The flowchart 800 can start from 802 and proceed to 804.

[0097] At 804, a collaborative filtering graph 118 corresponding to a user set and an item set related to the user set can be received. The circuit 202 can be configured to receive the collaborative filtering graph 118 corresponding to the user set and the item set related to the user set. Details regarding the collaborative filtering graph 118 are further described, for example, in FIG. 3.

[0098] At 806, based on the received collaborative filtering graph 118, the first user embedding set 406A and the first item embedding set 406B can be determined. Circuit 202 can be configured to determine the first user embedding set 406A and the first item embedding set 406B based on the received collaborative filtering graph 118. Details regarding the first user embedding set 406A and the first item embedding set 406B are further described, for example, in FIG. 4A.

[0099] At 808, the semantic clustering model 110 can be applied to each of the determined first user embedding set 406A and the determined first item embedding set 406B. Circuit 202 can be configured to apply the semantic clustering model 110 to each of the determined first user embedding set 406A and the determined first item embedding set 406B. Details regarding the application of the semantic clustering model 110 are further described, for example, in FIG. 4A.

[0100] At 810, based on the application of the semantic clustering model 110, the second user embedding set 410A and the second item embedding set 410B can be determined. Circuit 202 can be configured to determine the second user embedding set 410A and the second item embedding set 410B based on the application of the semantic clustering model 110. Details regarding the second user embedding set 410A and the second item embedding set 410B are further described, for example, in FIG. 4A.

[0101] At 812, a hypergraph (such as the hypergraph 502 in FIG. 5) can be constructed from the received collaborative filtering graph 118. Circuit 202 can be configured to construct a hypergraph (such as the hypergraph 502 in FIG. 5) from the received collaborative filtering graph 118. Details regarding the hypergraph 502 are further described, for example, in FIG. 5.

[0102] At 814, a third user embedding set 414A and a third item embedding set 414B can be determined based on the constructed hypergraph. Circuit 202 can be configured to determine a third user embedding set 414A and a third item embedding set 414B based on the constructed hypergraph. Details regarding the third user embedding set 414A and the third item embedding set 414B are further described, for example, in FIG. 4B.

[0103] At 816, a first contrastive loss can be determined based on the determined second user embedding set 410A and the determined third user embedding set 414A. Circuit 202 can be configured to determine a first contrastive loss based on the determined second user embedding set 410A and the determined third user embedding set 414A. Details regarding the first contrastive loss are further described, for example, in FIG. 4B.

[0104] At 818, a second contrastive loss can be determined based on the determined second item embedding set 410B and the determined third item embedding set 414B. Circuit 202 can be configured to determine a second contrastive loss based on the determined second item embedding set 410B and the determined third item embedding set 414B. Details regarding the second contrastive loss are further described, for example, in FIG. 4B.

[0105] At 820, a collaborative filtering score can be determined based at least on the determined first control loss and the determined second control loss. Circuit 202 can be configured to determine a collaborative filtering score based at least on the determined first control loss and the determined second control loss. Details regarding the collaborative filtering score are further described, for example, in FIG. 4B.

[0106] At 822, recommendations for items for user 120 can be determined based on the determined collaborative filtering score. Circuit 202 can be configured to determine recommendations for items for user 120 based on the determined collaborative filtering score. Details regarding the item recommendations are further described, for example, in FIG. 4B.

[0107] At 824, the determined recommended items can be rendered on display device 210. Circuit 202 can be configured to render the determined recommended items on display device 210. Details regarding the rendering of the determined recommended items are further described, for example, in FIG. 4B. The control can proceed to end.

[0108] Flowchart 800 is shown as discrete operations such as 804, 806, 808, 810, 812, 814, 816, 818, 820, 822, and 824, but the present disclosure is not so limited. Thus, in some embodiments, such discrete operations can be further divided into additional operations, combined into fewer operations, or removed, depending on the implementation, without detracting from the essence of the disclosed embodiments.

[0109] Various embodiments of the present disclosure can provide a non-transitory computer-readable medium and / or storage medium storing computer-executable instructions executable by a machine and / or a computer to operate an electronic device (e.g., the electronic device 102 of FIG. 1). Such instructions can cause the electronic device 102 to perform operations including receiving a collaborative filtering graph (e.g., the collaborative filtering graph 118) corresponding to a user set and a set of items related to the user set. The operations can further include determining a first user embedding set (e.g., the first user embedding set 406A) and a first item embedding set (e.g., the first item embedding set 406B) based on the received collaborative filtering graph 118. The operations can further include applying a semantic clustering model (e.g., the semantic clustering model 110) to each of the determined first user embedding set 406A and the determined first item embedding set 406B. The operations can further include determining a second user embedding set (e.g., the second user embedding set 410A) and a second item embedding set (e.g., the second item embedding set 410B) based on the application of the semantic clustering model 110. The operations can further include constructing a hypergraph (such as the hypergraph 502 of FIG. 5) from the received collaborative filtering graph 118. The operations can further include determining a third user embedding set (e.g., the third user embedding set 414A) and a third item embedding set (e.g., the third item embedding set 414B) based on the constructed hypergraph 502. The operations can further include determining a first contrastive loss based on the determined second user embedding set 410A and the determined third user embedding set 414A. The operations can further include determining a second contrastive loss based on the determined second item embedding set 410B and the determined third item embedding set 414B.The operation can further include determining a collaborative filtering score based on the determined first contrast loss and the determined second contrast loss. The operation can further include determining recommendations for items for a user, such as user 120, based on the determined collaborative filtering score. The operation can further include rendering the determined recommended items on a display device (such as display device 210).

[0110] Exemplary aspects of the present disclosure can provide an electronic device (such as the electronic device 102 in FIG. 1) that includes a circuit (such as the circuit 202). The circuit 202 can be configured to receive a collaborative filtering graph 118 corresponding to a user set and an item set related to the user set. The circuit 202 can be configured to determine a first user embedding set 406A and a first item embedding set 406B based on the received collaborative filtering graph 118. The circuit 202 can be configured to apply a semantic clustering model 110 to each of the determined first user embedding set 406A and the determined first item embedding set 406B. The circuit 202 can be configured to determine a second user embedding set 410A and a second item embedding set 410B based on the application of the semantic clustering model 110. The circuit 202 can be configured to construct a hypergraph (such as the hypergraph 502 in FIG. 5) from the received collaborative filtering graph 118. The circuit 202 can be configured to determine a third user embedding set 414A and a third item embedding set 414B based on the constructed hypergraph. The circuit 202 can be configured to determine a first contrastive loss based on the determined second user embedding set 410A and the determined third user embedding set 414A. The circuit 202 can be configured to determine a second contrastive loss based on the determined second item embedding set 410B and the determined third item embedding set 414B. The circuit 202 can be configured to determine a collaborative filtering score based on the determined first contrastive loss and the determined second contrastive loss. The circuit 202 can be configured to determine item recommendations for the user 120 based on the determined collaborative filtering score. The circuit 202 can be configured to render the determined recommended items on the display device 210.

[0111] In one embodiment, circuit 202 can be further configured to apply a GNN model (e.g., GNN model 114) to the received collaborative filtering graph 118, and can further determine each of the first user embedding set 406A and the first item embedding set 406B based on the application of the GNN model 114.

[0112] In one embodiment, circuit 202 can be further configured to determine a user-item correlation set, an item-user correlation set, and a user-user correlation set based on the constructed hypergraph 502. Circuit 202 can be further configured to determine a fourth user embedding set based on the determined user-item correlation set and user-user correlation set. Circuit 202 can be further configured to apply a first set of HGCN models (e.g., the first set of HGCN models 116A) to the determined fourth user embedding set. Circuit 202 can be further configured to determine a third user embedding set 414A based on the application of the first set of HGCN models 116A. Circuit 202 can be further configured to determine a fourth item embedding set based on the determined item-user correlation set. Circuit 202 can be further configured to apply a second set of HGCN models (e.g., the second set of HGCN models 116B) to the determined fourth item embedding set. Circuit 202 can be further configured to determine a third item embedding set 414B based on the application of the second set of HGCN models 116B.

[0113] In one embodiment, the semantic clustering model 110 can correspond to a spectral clustering model configured for dimensionality reduction.

[0114] In one embodiment, circuit 202 can be further configured to determine a fifth user embedding set based on the first contrastive loss and the third user embedding set 414A. Circuit 202 can be further configured to determine a fifth item embedding set based on the second contrastive loss and the third item embedding set 414B.

[0115] In one embodiment, circuit 202 can be further configured to determine a final user embedding based on the determined fifth user embedding set. Circuit 202 can be further configured to determine a final item embedding based on the determined fifth item embedding set, and the determination of the collaborative filtering score can be further based on the determined final user embedding and the determined final item embedding.

[0116] In one embodiment, each of the determined final user embedding and the determined final item embedding can correspond to a connection of at least one of a collaborative view, a hypergraph view, or a semantic view.

[0117] In one embodiment, the constructed hypergraph 502 can correspond to a multiplex bipartite graph having homogeneous edges.

[0118] In one embodiment, the first edge type in hypergraph 502 corresponds to an interaction between a first user and a first subset of items related to the first user, and the second edge type in hypergraph 502 can correspond to an interaction between a second subset of users and a second item related to each of the second subset of users.

[0119] The present disclosure includes all features enabling the implementation of the methods described herein and can also be positioned within a computer program product that can execute these methods when loaded into a computer system. A computer program in this context means any representation in any language, code, or notation of a set of instructions intended to cause a system with information processing capabilities to execute a particular function either directly or after performing any one or both of a) conversion to another language, code, or notation, and b) reproduction in a different form of content.

[0120] Although the present disclosure has been described with reference to some embodiments, those skilled in the art will understand that various changes can be made without departing from the scope of the present disclosure and equivalents can be substituted. Also, many modifications can be made to adapt a particular situation or content to the teachings of the present disclosure without departing from the scope of the present disclosure. Accordingly, the present disclosure is not intended to be limited to the specific embodiments disclosed, but is intended to include all embodiments falling within the scope of the appended claims.

Description of Reference Numerals

[0121] 100 Network environment 102 Electronic device 104 Server 106 Database 108 Communication network 110 Semantic clustering model 112 Recommendation model 114 Graph neural network (GNN) model 116A First hypergraph convolutional network (HGCN) model set 116B Second HGCN model set 118 Collaborative filtering graph 120 User

Claims

1. An electronic device, receiving a collaborative filtering graph corresponding to a user set and an item set related to the user set, determining a first user embedding set and a first item embedding set based on the received collaborative filtering graph, applying a semantic clustering model to each of the determined first user embedding set and the determined first item embedding set, determining a second user embedding set and a second item embedding set based on the application of the semantic clustering model, constructing a hypergraph from the received collaborative filtering graph, determining a third user embedding set and a third item embedding set based on the constructed hypergraph, determining a first contrastive loss based on the determined second user embedding set and the determined third user embedding set, determining a second contrastive loss based on the determined second item embedding set and the determined third item embedding set, determining a collaborative filtering score based on the determined first contrastive loss and the determined second contrastive loss, determining item recommendations for a user based on the determined collaborative filtering score, rendering the determined recommended items on a display device, An electronic device comprising a circuit configured as such.

2. The circuit is further configured to apply a graph neural network model to the received collaborative filtering graph, wherein each of the first user embedding set and the first item embedding set is further determined based on the application of the graph neural network model. The electronic device according to claim 1.

3. The circuit is further configured to determine a user-item correlation set, an item-user correlation set, and a user-user correlation set based on the constructed hypergraph, determining a fourth user embedding set based on the determined user-item correlation set and the user-user correlation set, applying a first hypergraph convolutional network (HGCN) model set to the determined fourth user embedding set, Determine the third user embedding set based on the application of the first HGCN model set, Determine a fourth item embedding set based on the determined item-user correlation set, Apply a second HGCN model set to the determined fourth item embedding set, Determine the third item embedding set based on the application of the second HGCN model set, The electronic device according to claim 1, further configured as such.

4. The semantic clustering model corresponds to a spectral clustering model configured for dimensionality reduction, The electronic device according to claim 1.

5. The circuit, Determine a fifth user embedding set based on the first contrast loss and the third user embedding set, Determine a fifth item embedding set based on the second contrast loss and the third item embedding set, The electronic device according to claim 1, further configured as such.

6. The circuit, Determine the final user embedding based on the determined fifth user embedding set, Determine the final item embedding based on the determined fifth item embedding set, Further configured as such, The determination of the collaborative filtering score is further based on the determined final user embedding and the determined final item embedding, The electronic device according to claim 5.

7. Each of the determined final user embedding and the determined final item embedding corresponds to the connection of at least one of a collaborative view, a hypergraph view, or a semantic view, The electronic device according to claim 6.

8. The constructed hypergraph corresponds to a multiplex bipartite graph having homogeneous edges, The electronic device according to claim 1.

9. The first edge type in the hypergraph corresponds to the interaction between a first user and a first item subset related to the first user, The second edge type in the hypergraph corresponds to the interaction between a second user subset and a second item related to each of the second user subsets, The electronic device according to claim 1.

10. In an electronic device, Receiving a collaborative filtering graph corresponding to a user set and an item set related to the user set, Determining a first set of user embeddings and a first set of item embeddings based on the received collaborative filtering graph; Applying a semantic clustering model to each of the determined first set of user embeddings and the determined first set of item embeddings; Determining a second set of user embeddings and a second set of item embeddings based on the application of the semantic clustering model; Constructing a hypergraph from the received collaborative filtering graph; Determining a third set of user embeddings and a third set of item embeddings based on the constructed hypergraph; Determining a first contrastive loss based on the determined second set of user embeddings and the determined third set of user embeddings; Determining a second contrastive loss based on the determined second set of item embeddings and the determined third set of item embeddings; Determining a collaborative filtering score based on the determined first contrastive loss and the determined second contrastive loss; Determining item recommendations for a user based on the determined collaborative filtering score; Rendering the determined recommended items on a display device; A method characterized by including the above.

11. Further including applying a graph neural network model to the received collaborative filtering graph, wherein each of the first set of user embeddings and the first set of item embeddings is further determined based on the application of the graph neural network model. The method according to claim 10.

12. Based on the constructed hypergraph, determining a user-item correlation set, an item-user correlation set, and a user-user correlation set; Determining a fourth set of user embeddings based on the determined user-item correlation set and the user-user correlation set; Applying a first set of hypergraph convolutional network (HGCN) models to the determined fourth set of user embeddings; Determining the third set of user embeddings based on the application of the first set of HGCN models; Determining a fourth item embedding set based on the determined item-user correlation set; Applying a second HGCN model set to the determined fourth item embedding set; Determining the third item embedding set based on the application of the second HGCN model set; The method according to claim 10, further comprising.

13. The semantic clustering model corresponds to a spectral clustering model configured for dimensionality reduction. The method according to claim 10.

14. Determining a fifth user embedding set based on the first contrast loss and the third user embedding set; Determining a fifth item embedding set based on the second contrast loss and the third item embedding set; The method according to claim 10, further comprising.

15. Determining a final user embedding based on the determined fifth user embedding set; Determining a final item embedding based on the determined fifth item embedding set; Further comprising, The determination of the collaborative filtering score is further based on the determined final user embedding and the determined final item embedding. The method according to claim 14.

16. Each of the determined final user embedding and the determined final item embedding corresponds to a connection of at least one of a collaborative view, a hypergraph view, or a semantic view. The method according to claim 15.

17. The constructed hypergraph corresponds to a multi-bipartite graph having homogeneous edges. The method according to claim 10.

18. The first edge type in the hypergraph corresponds to an interaction between a first user and a first item subset related to the first user. The second edge type in the hypergraph corresponds to an interaction between a second user subset and a second item related to each of the second user subsets. The method according to claim 10.

19. A non-transitory computer-readable medium storing computer-executable instructions, wherein when the computer-executable instructions are executed by an electronic device, Receiving a collaborative filtering graph corresponding to a user set and an item set related to the user set; Determining a first user embedding set and a first item embedding set based on the received collaborative filtering graph; Applying a semantic clustering model to each of the determined first user embedding set and the determined first item embedding set; Determining a second user embedding set and a second item embedding set based on the application of the semantic clustering model; Constructing a hypergraph from the received collaborative filtering graph; Determining a third user embedding set and a third item embedding set based on the constructed hypergraph; Determining a first contrastive loss based on the determined second user embedding set and the determined third user embedding set; Determining a second contrastive loss based on the determined second item embedding set and the determined third item embedding set; Determining a collaborative filtering score based on the determined first contrastive loss and the determined second contrastive loss; Determining item recommendations for a user based on the determined collaborative filtering score; Rendering the determined recommended items on a display device; Causing the electronic device to perform operations including the above, a non-transitory computer-readable medium characterized by this.

20. The constructed hypergraph corresponds to a multiplex bipartite graph having homogeneous edges. The non-transitory computer-readable medium according to claim 19.

Images