Processing top-N relational queries by learning

Abstract

Abstract.A top-N selection query against a relation is to find the N tuples that satisfy the query condition the best but not necessarily completely. In this paper, we propose a new method for evaluating top-N queries against a relation. This method employs a learning-based strategy.Initially, this method finds and saves the optimal search spaces for a small number of random top-N queries. The learned knowledge is then used to evaluate new queries. Extensive experiments are carried out to measure the performance of this strategy and the results indicate that it is highly competitive with existing techniques for both low-dimensional and high-dimensional data.Furthermore, the knowledge base can be updated based on new user queries to reflect new query patterns so that frequently submitted queries can be processed most efficiently. The maintenance and stability of the knowledge base are also addressed in the paper.

Evaluating Top-N queries in n-dimensional normed spaces

“…The query model in [5,9,41] is most related to the model in this paper. In [5,9,41], the ranking functions are p-norm distances for p = 1, 2, and , and the queries may be arbitrary points in the real vector space  n .…”

“…In [5,9,41], the ranking functions are p-norm distances for p = 1, 2, and , and the queries may be arbitrary points in the real vector space  n . Given a top-N query, the basic idea of the strategies in [5,9,41] is to find a small ndimensional square centered at the query point with side length 2r such that all of the top N tuples but very few undesired ones are contained in the n-dimensional square, and then the strategies are used to map the top-N query over a relational database into a traditional range selection query. The key of techniques in [5,9,41] is how to estimate the search distance r, i.e., half of the side length 2r.…”

“…Given a top-N query, the basic idea of the strategies in [5,9,41] is to find a small ndimensional square centered at the query point with side length 2r such that all of the top N tuples but very few undesired ones are contained in the n-dimensional square, and then the strategies are used to map the top-N query over a relational database into a traditional range selection query. The key of techniques in [5,9,41] is how to estimate the search distance r, i.e., half of the side length 2r. The main weakness of the methods in [5,9,41] is that an incorrect estimation of the search distance r may lead to the number of retrieved tuples to be either less than or much larger than N.…”

“…The key of techniques in [5,9,41] is how to estimate the search distance r, i.e., half of the side length 2r. The main weakness of the methods in [5,9,41] is that an incorrect estimation of the search distance r may lead to the number of retrieved tuples to be either less than or much larger than N.…”

“…In many applications, there are opportunities for evaluating top-N queries with arbitrary query points, and the ranking function is not F-monotone, as demonstrated by the following example [39,41].…”