Active Learning

Almost everyone has turned to the internet to vet a product before purchase. Some online retailers – like amazon.com – even provide you with recommendations for products you might like. In this age of online retailers and multiple anonymous raters, it is a worthwhile goal to use the pool of collective knowledge and opinions to guide product recommendations. The objective of recommender systems is to automate this process. Recommender systems aim to extract, the often implicit, information about a user or an item to make product recommendations. This benefits both the retailer and the enduser. One of the goals of machine learning is to make predictions on previously unseen data by learning from past experience. Therefore, recommender systems provide a natural context for applying machine learning techniques. We can think of a recommender system as learning about the preferences of a user to model the user’s preferences and to use this model to provide recommendations tailored to that user. Active Learning is a sub-class of supervised learning problems where the labels on the data points are not available a priori. The learning algorithm can actively query the user to obtain labels on a subset of data points in order to improve its model of the user’s preferences. However, there is a cost associated with obtaining each label from the user. The challenge in active learning is to minimize the number of active queries while maximizing the accuracy of predictions. The ultimate goal of a recommender system is to present items to the user that the user will like (i.e. rate highly). However, in order to build an accurate model of the user’s preferences we also need to query the user to learn about his/her preferences. Similar to the multi-armed bandit problem in learning, this presents an exploration vs. exploitation dilemma. In the exploration phase, the goal is to obtain labels (aka ratings) so as to improve the overall quality of predictions on the unlabeled data points (aka items). In the exploitation phase, the goal is to present the item with the highest predicted rating with the information obtained so far. In this work we employ active learning techniques to address this problem. We present a modification to existing techniques in recommender systems that show an improvement in prediction accuracy relative to published results. We also present a recommender system that makes an adaptive exploration/exploitation tradeoff by shifting the emphasis from exploration to exploitation as we learn more about a user’s preferences. Offline experiments on the MovieLens dataset [mle] are used to study the performance of our contributions. We also include some work spectral clustering and a probabilistic framework used for clustering.