Article Preview
TopIntroduction
In the last decades, the internet has explosively expanded. In a vast sea of information, technologies of big data engineering, machine learning, and deep learning, which solve problems in several areas such as fake news detection, healthcare, computer vision, and recommendation, prosper but following computational load remains a burden (Hao et al., 2022; Li et al., 2022; Tembhurne et al., 2022; Wang et al., 2021; Yu & Reiff-Marganiec, 2022). Meanwhile, internet surfers find it difficult to choose what they want from a large amount of varied information. Therefore, offering content lists of what they might want becomes more important.
Recommender systems are widely used by various web sites, such as Amazon, YouTube, and Netflix to help users find contents they might wish to interact with. Top-K personalized recommendation is conventionally performed by recommender systems to satisfy each user’s individual preference for various items (Li et al., 2022; Wang et al., 2021; Wu et al., 2016; Xue et al., 2019).
Social recommendations use additional information on social relations encoded as a network to improve recommendation performance (Gao et al., 2012; Tang et al., 2013a). Trust networks are social networks involving both a user-item network and user-user relations (Ma et al., 2009; Tang et al., 2013a). Because trust networks have the property that people with similar tastes naturally tend to gather or associate, called homophily (Gao et al., 2012; McPherson et al., 2001; Tang et al., 2013a), enhancement of item recommendation performance in trust networks has been verified with real-world data, such as Ciao, Epinions, Yelp, and Gowalla (Ardissono & Mauro, 2020; Fan et al., 2019; Tang et al., 2012a, 2012b, 2013a; Wang et al., 2020b).
Existing recommendation methods in trust networks for top-K recommendations are based on two approaches: rating prediction and direct ranking (Aggarwal, 2016, pp. 345–361). Rating prediction approaches offer predicted ratings for unrated items of a user and recommend top-K items to users based on their ratings (Ardissono & Mauro, 2020; Chanyoung et al., 2016; Hao et al., 2020; Wang et al., 2020b; Wang et al., 2021; Zhang et al., 2020). However, the best K-rated items do not always match with the best top-K recommendation; direct ranking methods are therefore preferred (Cremonesi et al., 2010; Yang et al., 2012). Moreover, rating prediction is generally based on explicit data representing specific and explicit preferences of users (e.g., from 1- to 5-star ratings on movies) and cannot be applied to implicit data, which includes a wider range of data, such as clicks, purchases, and views. In the process of learning user-user interactions, several models utilize a user-user cosine similarity matrix which causes a computational burden (Ardissono & Mauro, 2020; Tang et al., 2013a; Zhang et al., 2020).