Recommendations based on collaborative filtering are a popular feature on many websites. Picking the right collaborative filtering algorithm and configuring it correctly for your specific use case is not straightforward.
In this talk I will show how to perform offline experiments with MyMediaLite. I will discuss useful evaluation protocols and metrics for information retrieval and recommendation systems, and demonstrate how to measure them with MyMediaLite.
MyMediaLite is a CLI/.NET cross-platform recommender system toolkit. MyMediaLite not only contains many recommendation algorithms, but also an extensive evaluation framework.
WhileI will use MyMediaLite to demonstrate how to perform offline experiments, I will concentrate on general principles, so that it is interesting for a wider audience, i.e. developers and product managers that rely on Free Software/Open Source recommender system implementations like Apache Mahout or GraphLab, and people interested in ranking. About the speaker: Zeno Gantner is the main author of the MyMediaLite recommender system package. While working on his PhD, he was leader of the workpackage "Recommender Algorithm Development" in the European research project Dynamic Personalization of Multimedia" (MyMedia), with partners like BBC, British Telecom, and Microsoft. In 2012 he moved to Berlin to co-found the Berlin Recommender-Stammtisch and to work at Nokia on HERE Maps search and places discovery.