Tracking Where Events Take Place: Reverse Spatial Term Queries on Streaming Data

TL;DR

This paper introduces efficient probabilistic methods for reverse spatial term queries on streaming geo-tagged data, enabling fast, accurate location-based content analysis with limited memory.

Contribution

It develops a novel probabilistic model and algorithms for reverse spatial term queries in streaming data, improving efficiency and accuracy over existing methods.

Findings

Achieves 2-3 orders of magnitude faster query times

Maintains high accuracy with limited memory

Significantly reduces update times under certain conditions

Abstract

A large volume of content generated by online users is geo-tagged and this provides a rich source for querying in various location-based services. An important class of queries within such services involves the association between content and locations. In this paper, we study two types of queries on streaming geo-tagged data: 1) "Top-k reverse frequent spatial queries", where given a term, the goal is to find top K locations where the term is frequent, and 2) "Term frequency spatial queries", which is finding the expected frequency of a term in a given location. To efficiently support these queries in a streaming setting, we model terms as events and explore a probabilistic model of geographical distribution that allows us to estimate the frequency of terms in locations that are not kept in a stream sketch or summary. We study the back-and-forth relationship between the efficiency of queries, the efficiency of updates and the accuracy of the results and identify some sweet spots where both efficient and effective algorithms can be developed. We demonstrate that our method can be extended to support multi-term queries. To evaluate the efficiency of our algorithms, we conduct experiments on a relatively large collection of both geo-tagged tweets and geo-tagged Flickr photos. The evaluation reveals that our proposed method achieves a high accuracy when only a limited amount of memory is given. Also the query time is improved, compared to a recent baseline, by 2-3 orders of magnitude without much loss in accuracy and that the update time can further be improved by at least an order of magnitude under some term distributions or update strategies.