Mining Temporal Relationships with Multiple Granularities in Time Sequences

Discovering sequential relationships in a time sequence is often important to many application domains, including financial, manufacturing, etc. Indeed, a lot of work has been do e in this area (see [1, 6] for an overview). An important aspect for such a discovery process is, however, l argely missing from the literature: namely discovering temporal patterns or relationships that involve multiple t ime granularities. This paper reports the progress in this front. A more detailed study can be found in [4]. In this paper, we focus on algorithms for discovering sequen tial relationships when a rough pattern of relationships is given. The rough pattern (which we term “event s tructure”) specifies what sort of relationships a user is interested in. For example, a user may be interested in “wh ich pairs of events occur frequently one week after another”. The algorithms will find the instances that fit the e v nt structure. They can also be used when more information is filled in the structure, as in in the following example from the stock market domain: “what kind of events frequently follow within 5 trading-days from the o ccurrence of a rise in IBM stocks, when this is preceded, one week before, by the fall of another hi-tech compan y stock”. These algorithms form a core component for a data mining environment. We view the actual data mining process as being done interact ively through a user interface. Data mining requests (in terms of event structures) are issued through t he user interface and processed using the algorithms. Initially, the user will issue a request with a simple struct ure. Complicated structures may be given by the user