Two algorithms under stochastic gradient descent framework for recommender systems

Recommender systems is a subfield of machine learning that aims in creating algorithms to predict user preferences based on known user ratings or user behavior in selecting or purchasing items. Such a system has great importance in applications to sport, marketing and education. In the later case, we are interested to improve state of the art in student evaluation by predicting whether a student will answer the next question correctly. This prediction will help student to get right orientation which learning area should to be given greater attention. Note that the available data are given in the form of list, but not in the traditional form of matrix. Consequently, standard factorisation technique is not applicable here. However, stochastic gradient methods work well with the lists of data, where the most of relations are missing, and maybe required to be predicted. In this paper we shall consider optimisation of the most important regulation parameters, such as numbers of factors, learning and regularisation rates, numbers of global iterations. Our study is based on the Grockit and Chess data, which were used online during popular data mining contests on the platform Kaggle.