Συστήματα συστάσεων στην επιστήμη των δεδομένων : τεχνικές και εφαρμογές
View/ Open
Abstract
Recommendation Systems are extremely popular in the modern era. It may not be an exaggeration to characterize them as one of the most powerful tools of Machine Learning, as they are applied today on a broad scale and in a plethora of domains. One of the most characteristic areas is e-commerce, where Recommendation Systems are used to promote and increase the sales of online stores. Recommendation Systems aim to predict the interests of users and create personalized recommendations based on individual’s preferences. The data required for the operation of such a system usually come from search engine queries, as well as the purchase history or other information about the users themselves and the products they have chosen in the past. Websites such as Google, Netflix, Spotify, Amazon, etc., use such data to create recommendations that are not common to all users but personalized, based on each user's preferences.
The goal of this Thesis is the theoretical description of Recommendation Systems and then the application of the aforementioned methodology on a set of real movie data. We first catty over statistical analysis of the available data and then, a Recommendation System (RS) is constructed based on the Content-Based Recommender Systems category, using a logistic regression model. Then, two RSs belonging to the Collaborative Filtering category are constructed. The first one uses the Singular Value Decomposition-SVD technique, and the second one the Alternating Least Squares-ALS technique. A comparison of the three RSs is then performed using evaluation metrics like Precision, Recall and F1-Score. The best-performing RS was the one based on Singular Value Decomposition (SVD) technique; therefore, it was exploited to generate recommendations.