HGraph: A Connected-Partition Approach to Proximity Graphs for Similarity Search

Similarity search is a common approach to support new applications that deal with complex data (e.g., images, videos, georeferenced data, etc.). As a consequence, appropriate indexing structures to support this task have been proposed in the literature. Recently, graph-based methods have shown to be very efficient for approximate similarity search. However, some of the main types of graphs used still suffer from two main drawbacks: (i) slow construction, and (ii) inaccurate retrieval. To reduce these drawbacks, in this paper, we propose the HGraph method. HGraph is a divide-and-conquer method for building graphs for similarity search that recursively partitions the input dataset and connect vertices across partitions at different levels. The method can be used with different types of graphs proposed in the literature to speed up the graph construction time as well as to increase the approximate search results quality through long-range edges connecting pivots of different partitions. We present experimental results using real datasets that show that HGraph applied to the k-NNG graph was able to decrease the construction time while increasing the approximate search recall when compared to the k-NNG. Regarding the application of HGraph to the NSW graph, the query recall also increased, however with a higher computational cost. An analysis of different combinations of the tested methods demonstrated HGraph query times given a recall rate were always among the top results regarding different setups.