Synthetic dataset generation methodology for Recommender Systems using statistical sampling methods, a Multinomial Logit model, and a Fuzzy Inference System

It is said that we live in the age of data, and that data is ubiquitous and readily available if one has the tools to harness it. That may well be true, but so is the opposite. It is ever more common to try to start a data science project only to find oneself without quality data. Be it due to just not having collected the needed features, or due to insufficient data, or even legality issues, the list goes on. When this happens, either the project is prematurely abandoned, or similar datasets are searched for and used. However, finding a dataset that answers your needs in terms of features, type of ratings, etc., may not be an easy task, this is particularly the case for recommender systems. In this work, a methodology for the generation of synthetic datasets for recommender systems is presented, thus allowing to overcome the obstacle of not having quality data in sufficient amount readily available. With this methodology, one can generate a synthetic dataset for recommendation composed by numerical/ordinal and nominal features. The dataset is built with Gaussian copulas, Dirichlet and Gaussian distributions, a Multinomial Logit model and a Fuzzy Logic Inference System that generates the ratings according to different user behavioural profiles and perceived item quality.