Tuning of Learning Algorithms for Use in Automated Product Recommendations

Abstract: In this thesis, we study the problem of predicting users' media preferences based on their, as well as other users' historical rating data. The model used is that a user's rating for an item can be explained by a sum of bias terms and an inner product between two vectors in some multidimensional feature space that is specic to the product domain. This model is tted using a stochastic descent type method, the stochastic diagonal Levenberg-Marquardt method. The method is studied with respect to its stability and how fast it adapts its parameter estimates and it is found that these characteristics can be accurately predicted for linear regression problems, but that the dynamics become much more complex when training the nonlinear features. The concept of regularization and the tuning of regularization parameters with the Nelder-Mead simplex method is also discussed, as well as some methods for making the Python implementation faster by using Cython.