1 The Impact of Basic Matrix Factorization Refinements on Recommendation Accuracy Parisa Lak 1 Bora Caglayan 2 Ayse Basar Bener 3 Data Science Lab 1,2,3 Department of Mechanical and Industrial Engineering 1,3 , Department of Mathematics 2 Ryerson University 1,2,3 , Toronto, Canada parisa.lak@ryerson.ca; bora.caglayan@ryerson.ca; ayse.bener@ryerson.ca Abstract- Consumers are commonly overloaded with various choices when it comes to the selection of a product or service. Many e-tailers have adopted built-in recommenders to help consumers make more informed decisions. While Accuracy of the recommender agents has high impact on customer satisfaction, achieving high accuracy in these systems is challenging. Various models and techniques were proposed in the literature to improve accuracy of these systems. Matrix factorization (MF) has been widely used in previous studies mostly to overcome cold start problem. In this study, we show that fine-tuning the parameters used in the basic MF model plays a significant role in achieving higher prediction accuracy. Our evaluations are performed on a basic model with and without simple user and item biases on two datasets. Keywords- Matrix Factorization; Collaborative Filtering; Recommender Systems; I. Introduction Recommender systems as described by Dietmar et al. are “software tools and techniques providing users with suggestions for items a user may wish to utilize” [6]. This topic emerged as an independent research area in the mid-1990s [13,14,15]. Collaborative filtering is one of the methods that have widely been used in recommender systems. Matrix Factorization (MF), as a latent factor model, has been demonstrated to be one of the most competitive forms of collaborative filtering strategy [2]. MF models have been successful mainly because it scales well with large datasets [27]. In this technique, both items and users are characterized by factor vectors that are inferred from previous rating patterns. An item recommendation would be presented to a user, when there is a high correspondence between the item and user factors. In MF models the rating values in the matrix are assumed to be generated uniformly [2]. One user’s rating is calculated using the same factor vector without taking any specific information about the user or item into account. For instance, a user’s mood, her general rating behavior, or the time of the day might have an effect on how she rates an item. Also, the popularity of an item or the fact that two items in the same category might be rated similarly is not taken into account in the basic MF method. Adding biases is suggested to mitigate this problem to increase prediction accuracy [10]. The biases include but are not limited to simple biases and temporal biases. Simple biases are related to users’ rating behavior and the popularity of an item according to the observed rating pattern, while temporal biases reflects the deviation of rating pattern according to the time the rating occurs. Another solution to MF problems is to add constructs derived from implicit information to the model [16,17,25] or introduce context dependence rating predictions [26]. While these solutions improve the accuracy of the system, they may be computationally expensive. In some cases, a fine-tuned simple model with relatively lower accuracy but less computational cost might be preferred to a complex model, which needs more computational power to perform. Therefore, techniques that yield higher accuracy while avoiding increase in computational cost will be of value. In MF models, there are some parameters that should be set initially in the system. These parameters include factor vector size, learning rate, number of iterations (epochs), regularization rate, and initial feature value. Previous studies report their accuracy results along with the parameters that were used for their analysis [27, 3, 31], without referring to why those values were selected in the beginning. There are also other studies that do not report on these variables and only report on the final evaluation of the system [33, 34]. In both cases, it is not clear whether there were any primary steps beforehand to optimize these values to generate optimal system accuracy. In this study, the question that we would like to address is: -What is the effect of initial MF parameters on the accuracy of the model?