Mining Quantitative Temporal Dependencies Between Interval-Based Streams

Abstract

Data streams gathered from sensor systems can contain a significant amount of noise and are challenging to sequential pattern mining algorithms. A majority of existing approaches deals with such data as time point events to find before/after relations that induces loss of information when dealing with events lasting in time, i.e intervals. Other interval-based approaches focus on qualitative patterns and are sensitive to temporal variability. We consider that quantitative patterns maintaining lag information permit to deal better with this problem. In this work, we propose an efficient algorithm devised to extract quantitative patterns from interval streams: Interval Time Lag Discovery (ITLD). It is based on an intersection-based confidence that is automatically assessed with a statistical χ2 test. Experimental results, on both synthetic and real-life data, show that our method is more suitable for sensor interval streams and provides more precise information in comparison with existing approaches.