Association Rules

One of the most invetigated topics in data mining is the problem of discovering association rules. Most existing association rule mining algorithms have a batch behavior: given the user specified parameters, i.e., minimum support and confidence, the database is scanned, and eventually all large itemsets, then association rules, are produced. An inappropriate parameter choice yields either too many or too few association rules, which is annoying and wastes time.

Carma[Hidber 99] first brought the computation of large itemsets online. It generates candidates on the fly based on the currently large itemsets, in contrast to the tradional way, i.e., generating candidates a priori. It is this new way of generating candidates that brings Carma to online.

However, this new way of generating candidiates may suffer an exponential effect, that is, an exponentially increasing number of candidates may be generated. This exponential effect will not only slow down the algorithm, but also waste a lot of main memory.

Online and Adaptive Association Rule Mining:

To overcome the exponential effect of Carma and bring main memory in consideration, we propose to bring association rule mining both online and adaptive. The expoential effect is overcome by introducing a hybrid way of generating candidates. The main memory adaptability of our previous work is also incorporated.