Scaling the scales - A suggested improvement to IBM's Intelligent Recommendation Algorithm

Abstract

Recommender systems appear in a large variety of applications, and their use has become very common in recent years. As a lot of money can be made by companies having a better recommender system than their competitors, much of the research behind the best recommendation algorithms is proprietary and has not been published. We suggest an improvement to a graph-based collaborative filtering recommendation algorithm developed at IBM Research and published in "KDD '99 Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery data mining" in 1999: the Intelligent Recommendation Algorithm (IRA). We start by giving an overview of the field of recommender systems, how they work, and how they can be utilized. Next we give a detailed description of the graph-theoretical ideas IRA is built upon, and take an in-depth look at how the algorithm works. We then present a suggested improvement to IRA, called Scaling the scales. In Scaling the scales, customers using differently sized ranges of the rating scale can still be used to predict each other. We give a short overview of the design behind the implementation of our recommender system, which is available at GitHub (https://github.com/Stadly/Scaling-the-scales), as well as BORA (https://bora-uib-no.pva.uib.no), where this thesis is published. The implementation uses a modular software design to allow for developing extensions to the recommendation algorithms we have implemented, as well as deploying other recommendation algorithms in our system. Next we compare the recommendation quality of IRA and Scaling the scales using leave-one-out cross-validation. This method works by hiding a known rating, predicting what this rating should be, and comparing the correct and predicted ratings. Experiments on four different datasets were run. Our results indicate that Scaling the scales give slightly smaller errors than IRA when predicting ratings. A bigger improvement is that Scaling the scales calculates predictions in many cases where IRA is unsuccessful. Lastly, ideas for further improvement of Scaling the scales are presented.