kBrowse: kNN Graph Browser

The construction of k-nearest Neighbor Graph (kNNG) in several applications, such as a recommender system, similarity search, and data exploration is heavily based on the distance function which is usually unweighted and considered constant for all users. However, attributes are not all equally important and using different attribute weight gives different kNNGs. We present kBrowse, which allows users to explore, modify and understand kNNG computed from a weighted Manhattan distance function on loosely-defined weight space. It samples possible weight vectors, and computes their corresponding kNNGs. The system summarizes all the kNNGs into one graph by keeping all the edges with high edge certainty, a probabilistic measurement on how likely an edge is going to appear in the weight space. To make the weight space more defined, users can directly adjust the weight space or gives kNN examples. Sample weight vectors failing to satisfy the given conditions are then removed and the graph is summarized again. Finally, kBrowse also gives a user better understanding of kNN by showing which attribute is important in connecting nodes.