Long-Tail Recommendation Based on Reflective Indexing

We propose a collaborative filtering data processing method based on reflective vector-space retraining, referred to as Progressive Reflective Indexing (PRI). We evaluate the method's ability to provide recommendations of items from a long tail. In order to reflect 'real-world' demands, in particular those regarding non-triviality of recommendations, our evaluation is novelty-oriented. We compare PRI with a few widely-referenced collaborative filtering methods based on SVD and with Reflective Random Indexing (RRI) - a reflective data processing method established in the area of Information Retrieval. To demonstrate the superiority of PRI over other methods in long tail recommendation scenarios, we use the probabilistically interpretable AUROC measure. To show the relation between the structural properties of the user-item matrix and the optimal number of reflections we model the analyzed data sets as bipartite graphs.