Explainability on recommender systems using contextual data
Tutor / directorSalamó, Maria
Document typeMaster thesis
Rights accessOpen Access
All rights reserved. This work is protected by the corresponding intellectual and industrial property rights. Without prejudice to any existing legal exemptions, reproduction, distribution, public communication or transformation of this work are prohibited without permission of the copyright holder
Recommender Systems (RS) are a fundamental part of any relevant e-commerce website, and the inclusion, improvement and optimization of this kind of systems is a growing trend. As these systems evolve, people are more demanding with the accuracy and explainability of their recommendations. The goal of this work is to create a RS that provides simple explanations of the recommendations that it provides. These explanations are obtained by nding the di erent relations that can be interpreted from the data of any dataset. This work shows that recommendations can be explained in an easy way with Machine Learning (ML) algorithms that let contexts be understood from di erent perspectives. In this master thesis, every ML algorithm provides recommendations. Since algorithms can be explained, recommendations provided by them can also be explained in natural language, and, in some cases, the explanations are even provided with a helper picture. ML algorithms used in this master thesis provide recommendations with some features that can be queried by the user in order to meet user's current recommendation expectations or ltering settings. Users, therefore, obtain not only explainability, but a very open recommender system where dark patterns do not exist and there are neither boring nor repetitive recommendations. Users can easily explore explainable recommendation spaces that are prepared speci cally for each one of them, with parameters that help them decide implicitly the ratio of exploration and exploitation. All this work is presented in a framework that allows to encode as much context information as wanted by capturing as best recommendation space as possible at the time explainability results evolve even richer. This framework comprehends all the di erent representation, aggregation, computing and visualization techniques coded in a modular and extendable way so that it can be applied to any recommendation scenario. A variety of datasets have been tested in many ways showing surprisingly positive results of recommendation explanation capabilities in all the datasets.
SubjectsRecommender systems (Information filtering), Machine learning, Sistemes recomanadors (Filtratge d'informació), Aprenentatge automàtic
DegreeMÀSTER UNIVERSITARI EN INTEL·LIGÈNCIA ARTIFICIAL (Pla 2017)