Time Series Mining

In this chapter, we will first give the background and review existing works in time series mining. The background material will include commonly used similarity measures and techniques for dimension reduction and data discretization. Then we will examine techniques to discover periodic and sequential patterns. This will lay the groundwork for the subsequent three chapters on mining dense periodic patterns, incremental sequence mining, and mining progressive patterns.

Conclusions and Future Work

Association rule mining in spatial databases and temporal databases have been studied extensively in data mining research. Most of the research studies have found interesting patterns in either spatial information or temporal information, however, few studies have handled both efficiently. Meanwhile, developments in spatio-temporal databases and spatio-temporal applications have prompted data analysts to turn their focus to spatio-temporal patterns that explore both spatial and temporal information.

Mining Generalized Flow Patterns

In this chapter, we investigate an efficient method to discover this class of relative-location sensitive flow patterns. These generalized flow patterns aim to summarize the sequential relationships between events that are prevalent in sharing the same topological structures. We adopt the pattern growth approach and develop an algorithm called GenSTMiner to discover these patterns. In order to increase the efficiency of the mining process, we also present two optimization techniques. The first is the use of conditional projected databases to prune infeasible events and sequences, and the second is pseudo projection to reduce memory requirement.

Mining Flow Patterns in Spatio-Temporal Data

In this chapter, we describe flow patterns and the design of the algorithm called FlowMiner to find such flow patterns. FlowMiner incorporates a new candidate generation algorithm and employs various optimization techniques for better efficiency. The discovery of generalized spatio-temporal patterns will be described in the next chapter.

Mining Sequence Patterns in Evolving Databases

In this chapter, we analyze and improve the I/O performance of the GSP algorithm (Agrawal & Srikant, 1996). We also study the problem of incremental maintenance of frequent sequences.

Mining Dense Periodic Patterns in Time Series Databases

In this chapter, we describe a new periodicity detection algorithm to efficiently discover short period patterns that may exist in only a limited range of the time series. We refer to these patterns as the dense periodic patterns, where the periodicity is focused on part of the time series. We present a dense periodic pattern mining algorithm called DPMiner to find dense periodic patterns, and design a pruning strategy to limit the search space to the feasible periods. Experimental results on both real-life and synthetic datasets indicate that DPMiner is both scalable and efficient.

Introduction

Temporal databases capture time-related attributes whose values change with time, for example, stock exchange data. Temporal data mining is an important extension of data mining as it can be used to mine the activity rather than just states, and thus, infer relationships of contextual and temporal proximity, some of which may also indicate a cause-effect association.

Mining Spatio-Temporal Trees

We observe that many spatio-temporal trees patterns are both unordered and embedded. Unordered refers to the condition that the sequence between children of the same parent is not important (for example, the tree with bedroom1 as the left son and bedroom2 as the right son is the same as the tree with bedroom2 as the left son and bedroom 1 as the right son) while embedded suggests that it is not necessary to strictly keep the parent-child relations among nodes.

Mining Progressive Confident Rules in Sequence Databases

Real-life objects can be described by its attribute values. For example, a person has attributes such as gender, date of birth, education level, and job, and so forth. While the gender and date of birth of a person do not change, the education level and job may change with time. If we denote the set of attribute values of an object as its state, then the state of an object changes as the attribute values change with time. The states of an object at different time stamps form a state sequence.

Mining Spatio-Temporal Graph Patterns

Data mining in graph databases has received much attention. We have witnessed many algorithms proposed for mining frequent graphs. Inokuchi, Washio, and Nishimura (2002) and Karpis and Kumar (1998) introduce the Apriori-like algorithms, AGM and FSG, to mine the complete set of frequent graphs. However, both algorithms are not scalable as they require multiple scans of databases and tend to generate many candidates during the mining process. Subsequently, Yan and Han (2002) and Nijssen and Kok (2004) propose depth-first graph mining approaches called gSpan and Gaston, respectively. These approaches are essentially memory-based and their efficiencies decrease dramatically if the graph database is too large to fit into the main memory.