Leveraging Meta-Data for Multi-Variate Time Series Analysis

Thursday, October 31, 2019

2.30 p.m.

ISI seminar room 1st floor

Dr. Silvestro Roberto Poccia - Università degli Studi di Torino

We are living in a dynamic world where, in more and more domains, scientists  are starting to record observations from global phenomena (i.e.: epidemic, disaster, weather changes) and sensory networks (i.e.: smart buildings, motion recognition). Large scale phenomena are usually characterized by geographically or/and semantically inter-related events. Complex phenomena, such as the epidemic spread of a disease or other disaster events (such as hurricane or floods), are characterized by multiple aspects that can be recorded over time.

Because of the largely different circumstances characterizing different phenomena domains, researchers can rely on simulations to analyze, understand and predict their trends.

In these domains, we have observations with relations between themselves and their temporal evolution. To store and represent such observations, scientists use a particular data structure in which it is possible to record the temporal evolution of multiple aspects or observations (i.e. an observation can refer to a sensor for motion recognition or smart building or to a region for epidemic or disasters), the Multi- Variate Time-Series. Multi-Variate Time-Series are able to encode the temporal trend of a complex phenomenon, but to use this data structure for analytical purposes there are some key questions we have to consider:

❼  Can we discover and classify key events in multi-variate time-series?

❼ Can we compare multi-variate time-series? (i.e.: in the epidemiological domain, to predict future trends as well as suggest possible interventions).

❼ Can we search and retrieve multi-variate time-series based on the underlying key events or the overall trace similarities? (i.e.: in motion recognition, to understand the specific gesture executed, we would search for series with similar key events).

 A common characteristic that complex domains (building energy, epidemics, and gesture recognition) share is the presence of a relationship between observations (i.e. metadata a sensor network of a building as well as the mobility graph of an epidemic). My research start from the intuition that by leveraging this additional information, it is possible to perform better analysis of multi-variate time-series.

In this presentation I will focus on the following key challenges:

❼ Challenge 1: Can we outperform the standard strategy for comparing multi- variate time-series by leveraging metadata? During my PhD, me and my coau- thors propose some extensions to Dynamic Time Warping. By taking into account the metadata, obtain a better precision in a classification task.

❼ Challenge 2: Is it possible to identify multi-variate features to identify key events and support indexing? Me and my coauthors introduce the local features extraction strategy for identifying key-points in multi-variate time-series. We evaluate the power of this features in a classification task.

❼ Challenge 3: Can we extract salient multi-variate motifs (i.e., repeating patterns of the key events over time)? I and my coauthors propose new strategies for identifying motifs (repeated patterns) over multi-variate time-series. Our ap- proach outperforms the recent status of the art approach, MStamp, and opens the street for the identification of a more general kind of motifs.

Silvestro Roberto Poccia received the master’s  degree in computer science and engineering from the University of Naples “Federico II,” Italy, in 2013.

In 2013 he was a research fellow at the same university.

In the following, he was a research assistant technologist at Arizona State University in 2015-16.

In 2018 he completed the PhD in computer science at the University of Turin.

He is currently a postdoctoral research fellow at the same University.
His current research interests include the field of multivariate time series  applications, multimedia databases, and knowledge representation and management.