Event Detection
In the context of time series data analysis, it is often possible to observe the occurrence of a significant change in the behavior of a time series at a certain point or time interval. Such a change in behavior usually characterizes the occurrence of an event. An event detected in time series data may often represent the occurrence of a phenomenon with a specific and well-defined meaning in a given domain of knowledge.
Events detected in time series commonly appear as anomalies or change points. Anomalies are observations that do not conform to the expected behavior pattern within the dataset. In turn, change points separate different states in the process that generates the time series. The problem of change point detection is related to the problem of concept drift detection (i.e., concept drift) in time series. In this case, change point detection aims to find the specific moment (or interval) in time that marks the occurrence of the concept drift.
The complexity of this task is aggravated considering that the nature of the events observed in a time series is often unknown. Furthermore, the challenges of detecting events become even more critical in real-time monitoring systems (commonly associated with streaming data), where the demand for detection of these events is pressured by the need for speed in computational processing.
Pattern and Motif Discovery
Research on methods for pattern discovery in time series and spatio-temporal series is divided into two main activities: (i) pattern and sequence mining, and (ii) motif identification. The first focuses on supporting the knowledge extraction process using transactional techniques in the observation of frequent item patterns and frequent item sequences. The second seeks to explore techniques created for pattern discovery in time series of continuous observations.
Pattern mining is a broad research field encompassing various approaches. Among them, frequent pattern mining plays an important role in discovering associations and correlations between data. Patterns that are frequent in a dataset can be expressed by association rules (ARs). ARs highlight sets of frequent items in the antecedent, leading to those in the consequent.
During frequent pattern mining, it is common to produce thousands of ARs, making the study of each one burdensome. This problem weakens the process of discovering useful information. There is scientific effort to develop approaches capable of filtering interesting patterns, balancing the quantity of ARs produced with the goal of avoiding trivial ones already known by experts. Among the approaches to filter interesting patterns, some use interestingness measures, others list them based on properties, and others use subjective analysis.
There are some challenges associated with the non-stationarity of data that this project intends to explore. The first consists of discovering patterns that diverge from the general distribution of the data. The second consists of studying emerging patterns. The third is within the context of establishing dynamic relations in space-time; this project makes use of frequent sequence mining techniques. In this scenario, the goal is to discover sequences of related events in space and time. Finally, the identification of previously unknown patterns in continuous time series is known as motif identification. This project is particularly interested in finding spatio-temporally constrained motifs, i.e., patterns that may not be frequent in the entire dataset but are frequent within a given time and space interval (spatio-temporal blocks).
Prediction
Research on prediction methods is divided into three main activities: (i) regression methods, (ii) classification methods, and (iii) methods for dealing with concept drift. The first is aimed at predicting continuous values. The second addresses the general problem of classification. Finally, in the third, concept drift has an influence, especially in the context of imbalanced datasets and data stream influence. In particular, there are several possibilities for exploring and developing models capable of addressing different aspects of concept drift. Models are expected to adjust dynamically to achieve greater robustness and stability in their use. However, adaptability may not lead to robustness. An adaptive model that reacts to short-duration phenomena changes quickly and, therefore, tends to make models respond to spurious disturbances. Adaptations should occur when phenomena last long enough to characterize them as significant changes. Such a relationship between robustness and adaptability is not trivial, again referring to the stability-plasticity dilemma, and is the subject of study in this project.
For more information, watch the following video:
The presentation is also available at: slides.pdf