Time-Aware Similarity Search: A Metric-Temporal Representation for Complex Data

Abstract

Recent advances in information technology demand handling complex data types, such as images, video, audio, time series and genetic sequences. Distinctly from traditional data (such as numbers, short strings and dates), complex data do not possess the total ordering property, yielding relational comparison operators useless. Even equality comparisons are of little help, as it is very unlikely to have two complex elements exactly equal. Therefore, the similarity among elements has emerged as the most important property for comparisons in such domains, leading to the growing relevance of metric spaces to data search. Regardless of the data domain properties, the systems need to track evolution of data over time. When handling multidimensional data, temporal information is commonly treated as just one or more dimensions. However, metric data do not have the concept of dimensions, thus adding a plain “temporal dimension” does not make sense. In this paper we propose a novel metric-temporal data representation and exploit its properties to compare elements by similarity taking into account time-related evolution. We also present experimental evaluation, which confirms that our technique effectively takes into account the contributions of both the metric and temporal data components. Moreover, the experiments showed that the temporal information always improves the precision of the answer.