Designing a Recommender System based on Generative Adversarial Networks

Master’s thesis by Jonas Falkner, April 2018

1. Introduction

1.1 Motivation

Since the beginning of the World Wide Web an increasing amount of companies successfully employs Decision Support Systems (DSS), precisely Recommender Systems (RS) to improve user satisfaction and to engage users in long-term relationships driving additional sales opportunities. This development has led to RS being an essential Information System (IS) providing strong competitive advantage. Therefore the field of RS has received a lot of scientific attention what resulted in enormous progress with respect to the employed methods and algorithms. This development is the motivation for a work on advanced technologies for RS.

More precisely the employment of RS leads to several advantages on the customer side as well as on the business side. For the customer the advantage lies in an easy way to find new and interesting things. Therefore the RS supports the customer in effectively exploring the usually huge product space, narrowing down a particular set of choices while fostering the discovery of new, yet unseen products. The business in turn can benefit from the conversion of mere visitors to profitable buyers since RS help them to find products they are willing to purchase. Furthermore RS yield many possibilities for the cross- and up-selling of products via automated and personalized recommendation targeting. An additional advantage is provided by improving customer experience and thereby driving customer loyalty. This is achieved by the RS inducing a ”value-added relationship” between the customer and the site. Customers are much more likely to return to a site that offers a positive user experience by providing an easy to use interface and additional value-adding services. Finally the RS and its data can even be leveraged to predict future sales volume.

This thesis is written in cooperation with inovex. The inovex GmbH is a German IT consultancy and project house with around 250 employees that is focusing on digital transformation. As a company providing consultancy services to small and large business partners alike one of its main contributions is the allocation of technical knowhow. This concept is very successful w.r.t. the deployment of RS since even large companies often do not have the knowhow and resources to design, develop and implement advanced recommendation technologies. This is due to the mathematical and computational complexity of RS as well as practical challenges involving the efficient processing of huge amounts of data, the requirement of recommendation in real-time and some RS specific problems which are explained in detail in 2.1.5. To this side the acquisition of new and innovative ideas and techniques plays an important role which motivates the cooperative work on this thesis.

The state-of-the-art in RS is indeed still mostly represented by traditional methods like matrix factorization models for Collaborative Filtering (CF). However these methods have problems to produce high quality recommendations when there is no or only little data about user-item interactions and when this lack cannot be compensated by side information either about the user or about the items. One of the most immense technical developments in computer science and adjacent domains in recent times is the rise of Artificial Neural Networks (ANN). Respective models show state-of-the-art performance on several Natural Language Processing (NLP) and image recognition tasks. This success led to the application of deep learning in other domains like RS. Recent advances in ANN (especially Recurrent Neural Networks (RNN) ) for RS are producing state-of-the-art results on session-only data. These results were further improved by context and time sensitivity. ANN have also disrupted the domain of image and text generation as well as machine translation with the deployment of Generative Adversarial Nets (GAN). Furthermore recent improvements on the architecture and the training of GAN have rendered them applicable on a greater variety of problems, e.g sequential or discrete data.

Beholding the recent attention to the respective research topics it is not surprising that there is already ongoing research on using GAN in conjunction with RNN to produce recommendations. However these are only emerging ideas that still present several problems and lack implementations and comparative studies to evaluate their real potential and value.

I intend to address these deficiencies in this master’s thesis.


Get in touch!

Christian Meder

Chief Technology Officer