Optimizing Query Performance with Inverted Cache in Metric Spaces

Popularity-Based Ranking for Fast Approximate kNN Search

“…The Inverted Cache Index was recently proposed in Antol and Dohnal (2016). It takes into account previously processed queries too, which is similar to some approaches presented above.…”

“…It does not record the queries processed so far, but rather the number of times a given partition/bucket (or even data object) contributed to the final result of such queries. This technique was originally proposed in Antol and Dohnal (2016). In this article, we further study its properties and parameter settings and, more importantly, apply it to approximate kNN queries.…”

“…If an indexing structure is hierarchical, this is applied also in internal "nodes". In particular, we apply the variant called object ratio (Antol and Dohnal, 2016). This ICI counter is then used to stretch/extend the original metric distance between the bucket's or node's representative (p -pivot) and a query (q) as follows:…”

“…where d ICI denotes the modified distance and d norm is the normalization distance. The parameter base is fixed to 10 and the parameter pwr is set to 2 for M-tree and 5 for M-index, which is recommended in Antol and Dohnal (2016). The normalization distance controls the point where the "attraction" force turns into the "repulsive" force to push irrelevant data partitions away.…”

“…This technique exploits queries executed previously to gather necessary statistics, so the so-called popularity of data partitions can be established and applied to prioritize access to the partitions. We have proposed this technique in Antol and Dohnal (2016) and applied it to precise kNN queries. Here, we elaborate on this concept more and reveal a connection between algorithm parameters and data-set properties.…”