Article Preview
Top1. Introduction
A recommender system can be defined as a form of information filtering system that offers products and services that interest users (Deshpande and Karypis, 2004). Recommender systems have emerged as an independent field of research since the mid-1990s (Ricci et al., 2011) and played an important role in helping to provide information matching users’ preferences in areas, such as online shopping (e.g., Amazon), e-news (e.g., Yahoo! News Today), music (e.g., Last.fm), travel (e.g., TripAdvisor), movies (e.g., Netflix), and social networks (e.g., Facebook). Recommender systems often exploit the characteristics of items (in content-based approaches (Pazzani and Billsus, 2007; Mooney and Roy, 2000; Ahn et al., 2007; Balabanovic and Shoham, 1997; Middleton et al., 2004)) or user preferences in the past and the similarity in the tastes of users (in collaborative filtering approaches (Nguyen and Huynh, 2017; Xu and Dang, 2014; Deshpande and Karypis, 2004; Konstan et al., 1997; Koren, 2008; Mnih and Salakhutdinov, 2008; Hofmann, 2003)) to give personalized recommendations. However, these systems usually fail to consider evolving user preferences in different contextual situations (Kulkarni and Rodd, 2020).
Many research works (Adomavicius et al., 2005; Palmisano et al., 2008; Baltrunas et al., 2012; Zheng et al., 2015) have shown that contextual information, such as time, location, weather conditions, and mood, plays an important role and affects the user’s product experience. In other words, user preferences may vary depending on the context. For example, users may choose a different restaurant when they go with their children instead of with friends. Users can choose to visit Dalat1 city with a cool climate in summer rather than winter. The role of context has been recognized in enhancing recommendation results and retrieval performance (Adomavicius et al., 2019). In many application domains, a context-independent representation may lose predictive power because of aggregating useful contextual information (Adomavicius et al., 2011). Contextual information has been integrated into recommendation systems in several practical applications, such as Sourcetone2, which allows selecting songs based on the listener’s mood (Adomavicius and Tuzhilin, 2011). Hydra is a recommender system that offers multimodal transportation planning and is adaptive to various situational contexts (e.g., nearby point-of-interest distribution and weather) (Liu et al., 2019). It has been deployed in Baidu Maps3 since August 2018.
Context-aware recommender systems (CARSs) have become a very interesting topic. Workshops on CARSs have been organized for many years, attracting researchers to discuss issues related to considering and integrating contextual information in traditional recommender systems, including the CARS workshop series (2009-2012, 2019) and CARR (context-aware retrieval and recommendation) workshop series (2011-2014). Recently, the 2019 CARS workshop was held in Denmark; and there were discussions on a new generation of CARS 2.0, such as latent CARS, exploiting the sequential actions of users (sequence-based recommender systems), and using contextual information from different data sources, such as text, images, videos, and speech (Adomavicius et al., 2019).