Analysing Temporal Evolution of Complex Data Using Similarity Queries

Abstract

Regardless of the data domain, there are applications that must track the temporal evolution of data elements. Based on the instances present in the database, the goal is to estimate the state of a given element at a different time instant from those available in the database. This kind of task is common in many database application domains, such as medicine, meteorology, agriculture, financial, and others. In content-based retrieval with complex data (such as images, sounds and videos), data are usually represented in metric spaces, where only the distances between elements are available. Without dimensional coordinates, it is not possible simply to add a time dimension for trajectory estimation in these spaces, as is the case in multidimensional spaces. In this article we propose to map the metric data to a multidimensional space so that we can estimate the element’s status at a given time instant, based on known states of the same element. As it is not possible to create the complex data equivalent to its estimated position in mapped space, we propose to apply similarity queries using this position as query center. Then, we estimate how this element would be, retrieving the real data elements present in the database that are close to the estimate. In this article, in addition to the nearest neighbor query (k-NN), we propose to use two other queries: kAndRange and kAndRev. With both methods, we aim to prune non-relevant elements from the query results, retrieving only the elements that are really close to the estimates. We present experiments with different query scenarios, evaluating the effects of varying input parameters of the proposed queries.