An Empirical Evaluation of Collaborative Filtering Methods with Implicit Feedback

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Date

2018-06-18

Department

Major/Subject

Computer Science

Mcode

SCI3042

Degree programme

Master’s Programme in Computer, Communication and Information Sciences

Language

en

Pages

70

Series

Abstract

There are two primary ways of collecting preferences of users towards items. In the first method, users express their interest explicitly using features such as star ratings. On the other, their interest is inferred from their implicit interaction such as frequency of visit and purchase. Data collected on preferences of users is used by recommender systems to produce suggestions to the users. A user receives recommendations of items when other users who have similar preference express interest towards those items. This way of producing recommendations is known as collaborative filtering. In this thesis, we set out to examine if implicit feedback datasets can be comparable to explicit feedback datasets in being the source of user information to generate recommendations. To achieve this, we use a real-life dataset that contains product visit and purchase information of users. We examine how open-source tools perform in processing our dataset that contains implicit feedback. We study the availability, reliability, and usability of these tools. Using these tools and our dataset, we demonstrate how to build models that are generalizable and have predictive power that extends beyond observed user preferences of the dataset. We demonstrate the use of evaluation metrics to measure the accuracy and relevancy of recommendations generated using the models we built. We discover from the experiments, our dataset can be used to generate accurate and relevant recommendation to the users. However, we report the inadequacy of an output of such valuable qualities as accuracy and relevancy without elements of novelty, serendipity, and diversity. A recommender engine has to have the capacity to provide useful suggestions that are outside of the obvious interest range of users.

Description

Supervisor

Kaski, Petteri

Thesis advisor

Salakka, Aki

Keywords

recommender systems, collaborative filtering, implicit feedback, matrix factorization, open source tools, KNN

Other note

Citation