A Lightweight Indexing Approach for Efficient Batch Similarity Processing with MapReduce

Similarity search is a principle operation in different fields of study. However, the cost for that operation is expensive due to several reasons, mainly by redundancy and big data load. There are many approaches that concentrate on how to speed up similarity search, especially with massive datasets, so that we can employ it for plenty of recent applications. In this paper, we study an efficient way for either single or batch similarity processing with MapReduce while minimizing redundant data by building lightweight indexes from the data and query sources. More specifically, we propose a general query processing scheme that not only handles a single query but also deals with sets of query in an incremental manner. In addition, we build the indexes in an ordered fashion, the so-called sorted inverted indexes, so that we can perform our quick pruning strategy that discards unrelated objects. Moreover, we embed metadata inside the indexes to reduce inessential duplicates. Last but not least, we measure our proposed solution by conducting empirical experiments on real datasets. The results verify the efficiency of our method when we do similarity search with query batches, especially when both query sets and datasets are large.