Lars Schmidt-Thieme jointly with Steffen Rendle, will have a tutorial in the KDD-2012 conference. The abstract of this tutorial is as follows:
In many problems nominal variables with many levels occur, e.g., in recommender systems that try to predict how users will rate items based on past rating behavior, whereby users and items are treated as mere ID. For such problems, factorization models that associate every level of these variables with a hidden feature vector and model their influence on a target variable indirectly through the hidden features alone, have proven to work very well. In this tutorial we will provide a broad introduction to factorization models, starting from the very beginning with matrix factorization and then proceed to generalizations such as tensor factorization models, multi-relational factorization models and factorization machines. We will provide theoretical insight in the assumptions behind the respective modelling approaches as well as present the state-of-the-art learning algorithms. As time permits, we will cover further aspects of this model family such as factorization models involving time, strong generalization vs. semi-supervised learning with factorization models, factorization models to handle missing data, and Bayesian approaches to factorization models.
You can download the slides of this tutorial with the following links:
More information about the tutorial can be found here.