Time Series Search Using Traits

Abstract: Time series data occurs in many real world applications. For examplea system might have a database with a large number of time series, and a user could have a query like Find all stocks tha tbehave ”similarly” to stock A. The meaning of ”similarly” can vary between different users, use cases and domains. The goal of this thesis is to develop a method for time series search that can search based on domain specific patterns. We call these domain specific patterns traits. We have chosen to apply a trait based approach on top of a interest point based search method. First the search is conducted using a interest point method and then the results are ranked using the traits. The traits are extracted from sections of the time series and converted to a string representing its structure. The strings are then compared using Levenshtein distance to rank the search results. We have developed two types of traits. The new time series search method can be useful in many applications where a user is not looking for point-wise similarity, but rather looks at the general structure and some specific patterns. Using a trait based approach can better translate to how a user perceives time series search. The method can also yield more relevant results, since this new method can find results that a classic point-wise based search would rule out.