A Context Ultra-Sensitive Approach to High Quality Web Recommendations based on Web Usage Mining and Neural Network Committees

Personalization tailors a user’s interaction with the Web information space based on information gathered about them. Declarative user information such as manually entered profiles continue to raise privacy concerns and are neither scalable nor flexible in the face of very active dynamic Web sites and changing user trends and interests. One way to deal with this problem is through a completely automated Web personalization system. Such a system can be based on Web usage mining to discover Web usage profiles, followed by a recommendation system that can respond to the users’ individual interests. We present several architectures that rely on pre-discovered user profiles: Context Sensitive Approaches based on single-step Recommender systems (CSA-1-step-Rec), and Context UltraSensitive Approaches based on two-step Recommender systems (CUSA-2-step-Rec). In particular, the two-step recommendation strategy based on a committee of profile-specific URL-Predictor models, is more accurate and faster to train because only the URLs that are relevant to a specific profile are used to define the relevant attributes for this profile’s specialized URL-Predictor model. Hence, the model complexity, such as the neural network architecture, can be significantly reduced compared to a single global model that could involve hundreds of thousands of URLs/items. The two-step approach can also be expected to handle overlap in user interests, and even to mend the effects of a profile dichotomy that is too coarse. Finally, we note that all the mass-profile based recommendation strategies investigated are intuitive, and are low in recommendation-time cost compared to collaborative filtering, (no need to store or compare to a large number of instances). In our simulations on real Web activity data, the proposed context ultra-sensitive two-step recommendation strategy achieves unprecedented high coverage and precision compared to other approaches such as K-NN collaborative filtering and single-step recommenders such as the Nearest-Profile recommender.