Page 172 - The-5th-MCAIT2021-eProceeding

2. Implementation Approaches
2.1. Early approaches

Content-based (CB) or collaborative filtering (CF) based recommendation algorithms are pioneer algorithms
in the early approaches to personalize recommender systems (Ricci et al., 2011).

2.1.1. Content-based
Content-based algorithms use various available content information such as the utility, price, color, brand,
size, etc. of the goods in e-commerce (Schafer, 1999), or the genre, director, duration, language, casting, etc. of
the movies in review systems (Balabanovic, 1997) to model recommender systems. It is intuitive to explain the
recommendation to users through various available item content information. Ferwerda, 2012 provided a
complete study for elucidation of CB recommendations. CB algorithms have the main demerit of collecting
various content information in different application domains, and it is a time-consuming effort.

2.1.2. Collaborative Filtering-based
Collaborative filtering-based approaches (Ekstrand, 2011) have addressed the issue associated with the CB
approach. User-based CF (UBCF) represents each user as a vector of ratings and predicts the missing rating
based on the weighted average of other users' ratings (Resnick, 1994). Item-based CF (IBCF), represents each
item as a vector of ratings and predicts the missing rating based on the weighted average ratings from similar
items (Sarwar, 2001). The explainability of CF lies in its design. UBCF can explain the behavior as “users who
are similar to you loved this item”, and IBCF can explain as “the item is similar to your previously loved items”.
CF approaches have improved the accuracy of the predictions but have less explainability (Herlocker, 2000a)
and an issue of cold start (J. Gope, 2017), where it can’t recommend the contents to new users.

2.2. Advanced approaches

The RS have been further enhanced through advanced techniques like latent factors (LFM), deep learning
(DL), knowledge graph (KG), reinforcement learning (RL), and language models (LM) (Shaoxiong, 2021).
2.2.1. Latent Factor Models
The CF approach discussed in the previous “Early approaches” section has been further enhanced through
dimensionality reduction methods like latent factor models (LFM). The most well-known LFM is matrix
factorization (MF). MF methods have many specific techniques like singular value decomposition (Koren,
2009), non-negative MF (DD Lee, 2001), and probabilistic MF (A Mnih, 2008). These approaches create a
latent factor representation to calculate the matching score of the user-item pairs. LFM especially MF and its
variants were very successful in rating prediction tasks. Latent factors in LFMs do not possess any intuitive
meanings, which restrict the explanatory power of recommendations. It necessitates the need for explainable
recommendations (Zhang, 2018) efforts. Zhang et al., 2014, defined the explainable recommendation problem
by aligning the latent dimensions with explicit features and proposed an Explicit Factor Model.

2.2.2. Deep Learning Models
CF methods have been further improved by the usage of deep learning (DL). Similarity learning and
representation learning are the two main approaches in DL methods of CF approaches. In similarity learning, it

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [159]
Artificial Intelligence in the 4th Industrial Revolution