Recommender systems (RSs) have revolutionized the way users discover and engage with content. These systems play a crucial role in numerous domains for providing personalized recommendations to individual users, such as product recommendations in e-commerce platforms (e.g., Amazon, Walmart, and Target), curated playlists in streaming platforms (e.g., YouTube and Spotify), targeted ads in online advertising, and many others. The primary purpose of recommender systems is to recommend a product or service to a user, and recommender systems do this by consuming the users’ historical data to find patterns, learn users’ preferences, and predict the likelihood of the user liking the product or service. Moving beyond the traditional collaborative filtering (CF) modeling, where the main input to such systems is the past (explicit or implicit) preferences of users, most modern RSs leverage additional sources of information, such as context and social network data. In particular, social RSs, where the user’s social network is also used as input to the recommendation process on the premise that these users are more similar and thus better predictors in terms of user preferences, have been shown to be very effective and computationally efficient ^[1][2][1,2]. Social networks are naturally represented as graphs, and this representation has allowed researchers to model concepts such as influence propagation, trust, etc. Social network analysis tools [3] can enhance a recommender system by considering the local and global characteristics of users in the corresponding social network of users, such as centrality scores, discovery of weak/strong ties, community detection, neighborhood overlap, positive and negative edges, behavior analysis, etc.

In recent years, deep learning (DL) models have emerged as the dominant underlying architecture for RSs, garnering substantial interest in both academic research and industrial applications ^{[4][5][6][7][8][9]}[4,5,6,7,8,9]. The allure of deep learning lies in its ability to capture complex and non-linear relationships between users and items. By leveraging sophisticated neural architectures, DL models excel at capturing intricate user–item dynamics, thus enhancing the accuracy and relevance of recommendations. Additionally, these models offer the flexibility to seamlessly integrate diverse data sources, such as contextual information, textual data, and visual cues, thereby enriching the recommendation process with a wealth of information.

Within the realm of DL algorithms, there exists a distinct category known as graph learning-based methods, which offer a unique perspective in RSs. In these methods, RS data are represented and analyzed through the lens of graphs. Specifically, the interactions between users and items can be depicted as interconnected nodes in a graph, where the links reflect the relationships between them. By leveraging graph-based representations, RSs gain the advantage of incorporating structured external information, such as social relationships among users, into the recommendation process. This integration of graph learning provides a unified framework for modeling the diverse and abundant data present in RSs.

2. Social Recommender Systems

With the widespread adoption of online social platforms, social RSs have emerged as a highly promising approach that leverages social networks among users to enhance recommendation performance ^[10][11][12][14,15,16]. Grounded in sociological concepts, such as homophily and social influence ^[13][17], this field of study operates under the premise that users’ preferences are more profoundly shaped by those of their interconnected peers than by those of unfamiliar users ^[14][18]. Tang et al. ^[11][15] give a narrow definition of social recommendation as “any recommendation with online social relations as an additional input, i.e., augmenting an existing recommendation engine with additional social signals”. Researchers have long recognized the influence of social connections on individuals, highlighting the phenomenon of similar preferences among social neighbors as information diffuses within social networks ^{[1][2][15][16][17][18][19][20]}[1,2,11,19,20,21,22,23]. Social regularization ^[1][2][1,2] has been shown to be effective in social recommendation scenarios, operating under the assumption that users with similar preferences exhibit shared latent preferences within popular latent factor-based models ^[21][24]. In the realm of social recommendations, GNNs have emerged as a powerful tool for capturing the intricate relationships between users, items, and other contextual features, such as time and location ^{[15][19][22][23]}[11,12,13,22]. The incorporation of GNNs into social recommendation models allows for a comprehensive understanding of the complex dynamics present in social networks, resulting in more accurate and contextually relevant recommendations. Recent studies investigate various aspects, such as modeling user–item interactions, capturing social influence, incorporating contextual information, and addressing scalability challenges. Through a deeper understanding of these developments, researchers and practitioners can gain insights into the potential benefits and challenges associated with integrating GNNs into social recommendation frameworks, thereby fostering further innovation and advancements in this exciting research area. Ying et al. ^[19][22] introduced PinSage, a novel framework based on GNNs, designed specifically for personalized feed recommendations. The motivation behind PinSage stems from the scalability limitations observed in traditional CF methods. PinSage revolutionizes the landscape of personalized feeds (news) recommendations by constructing a graph representation that encompasses both items and users. Leveraging the expressive power of GNNs, PinSage efficiently learns personalized feed representations for each user. This graph-based approach enables the model to capture the complex relationships between items and users, thereby facilitating accurate and relevant pin recommendations. PinSage innovatively combines content-based and collaborative filtering approaches. In their paper, Fan et al. ^[15][11] presented GraphRec, an innovative recommendation algorithm that leverages the power of graphs. They highlight the limitations of traditional recommendation techniques, particularly in dealing with the cold-start problem and effectively capturing intricate user–item connections. GraphRec aims to overcome these challenges by using graphs to model user–item interactions in the form of a diverse graph, rating scores, and differentiating the ties strengths by considering the heterogeneous strengths of social relations. Wu et al. ^[22][12] proposed Diffnet, a neural influence diffusion model for social recommendations. Please note that this term refers to information diffusion in a social graph in this context (not to be confused with Generative AI diffusion). Diffnet utilizes a user’s social network data to provide personalized recommendations. Its neural architecture comprises four main components: the embedding layer, the fusion layer, the layer-wise influence diffusion layers, and the prediction layer. Once the influence diffusion process reaches stability, the output layer predicts the final preference for a user–item pair. Compared to other existing social recommendation models, the Diffnet architecture leverages both user–item interaction data and social network information to enhance the recommendation accuracy. In subsequent research work, Wu et al. ^[23][13] introduced an enhanced version of the Diffnet model, called Diffnet++. This enhanced model builds upon the neural influence diffusion framework for social recommendations. In addition to learning user embeddings through influence diffusion from their social network, Diffnet++ incorporates user interest embeddings acquired through interest diffusion from user–item interactions. As each user is connected to their social connections, they form a user–user graph used to learn the user influence embeddings. Similarly, the user is connected to items, enabling the learning of user interest embeddings from item interactions, represented as user–consumed items graphs. Diffnet++ incorporates a neural architecture consisting of four essential components: the embedding layer, the fusion layer, the layer-wise influence diffusion layers, and the prediction layer. To ensure the efficacy of the user embeddings in both influence and interest diffusion graphs, a node attention layer is employed to selectively emphasize the most relevant information from the surrounding connections. Subsequently, after training the user–influence embeddings and user–interest embeddings separately, they are aggregated in a graph attention layer to generate the final user embeddings from the influence and interest perspectives. The model predicts the final preference for a user–item pair once the influence and interest diffusion processes have reached a stable state. This comprehensive architecture empowers the model to effectively capture both influence and interest dynamics. By incorporating item relations, the RelationalNet algorithm emphasizes both user and item influences and interests by adding the layer-wise item diffusion layer. This extension enables the algorithm to capture and leverage the complex relationships between users and items. Consequently, the algorithm aims to provide enhanced user recommendations by considering the individual interests of users and the intricate interplay between users and items. The inclusion of the item–item graph and the item–consumed users graph in RelationalNet further enriches the modeling capabilities, allowing for a more comprehensive representation of the user–item ecosystem. Through the integration of these additional relational graphs, the RelationalNet model aspires to deliver more accurate and contextually relevant recommendations to users, considering a broader spectrum of influences and connections.