Article Preview
TopIntroduction
Recommender system (RS) has emanated as the most popular application of big data (Mayer-Schönberger & Cukieralker, 2014; Livinus et al., 2016) that helps the user to find relevant information in the era of data deluge. As a result of the expansion in the e-commerce business, the customers (buyers) have to process a large amount of information before they decide to buy an item. RS provides the solution to this information overload problem as it is used to suggest products to the customers. It further enhances the e-commerce business since it helps to find the interesting things for the users which they may wish to purchase, however, did not come across things, due to information overload. RS is often used to improve customer’s trust and loyalty as it provides unique personalized service to each customer. User interest may be determined in the following ways:
A formal definition of RS is as follows: Let U be the set of users and I be the set of items. We define a function F that quantifies the utility of item to a user by the mapping where R is defined as the set of ratings (Sarwar et al. 2000). The goal is to learn and use F to determine the rating of a previously unseen item i by user u in order to generate a possible set of recommended items.
Several traditional recommendation techniques include collaborative filtering (CF) (Ekstrand et al., 2011), Content-Based Recommender System (CBS) (Lops et al., 2011) and Hybrid Recommender System (Varshavsky et al., 2014). Each technique has its advantages and limitations; for example, domain free nature of CF has averted most research work from CBS (Ekstrand et al., 2011). However, CF suffers from cold start and sparsity problem (Hedge et al., 2015) while CBS suffers from overspecialization (Lops et al., 2011). Consider a case where a user is new to the system and the system does not have any information about the user. In such a case, drawing inference about user’s preferences becomes difficult and a quality recommendation cannot be produced. Such a scenario occurs both in the case of novel users and new items and is a perfect example of a cold start problem which indirectly leads to sparsity problem. Many advanced techniques have been suggested for mitigating these limitations such as, social, interactive, deep learning, context-aware, and fuzzy logic-based RS which also increase coverage and accuracy. For recommending items to a group of users, group recommender system (GRS) came into picture which exploits individual contextual information, tastes and demographic information for developing suggestions (Boratto, 2016). Various nature-inspired algorithms such as Cuckoo search (Tosun, 2014), swarm optimization (Chen et al., 2012) and genetic algorithm (Whitley, 1994) have been investigated by the researchers for accuracy and optimization of the system.
Several evaluation methods used for RS are Recall, Precision, F-measure, Mean absolute error (MAE), Coverage, Root Mean Square (RMSE), Sensitivity, Specificity and Receiver Operating Characteristics (ROC) (Franzen, 2011). In order to analyze the system developed, an appropriate metric should be chosen taking into consideration the user perspective and purpose that needs to be fulfilled by the system. Therefore, to help the researchers understand the RS development from various aspects including evaluation metrics and datasets and to assist them, this paper reviews the RS from all aspects that need to be considered for its development.