A graph based approach to inferring item weights for pattern mining

In this paper we present a novel approach to semi supervised classification based on a weight transmission model. Our research is motivated by rule extraction from transactional data, where a transaction consists of a collection of items, each of which is assigned a weight denoting its importance in relation to other items in the collection. The assignment of weight to items enables the end user to guide the rule extraction process to generate rules involving high impact items, thus enhancing the knowledge discovery process. Most previous research to weight assignment has used domain specific information to assign weights to items. We propose a model, XWeightTransmitter, that relaxes the assumption that domain information is available for all items. XWeightTransmitter interpolates the unknown weights from a known subset of weights and is an extension of the WeightTransmitter approach. Our experimentation shows that XWeightTransmitter outperforms a previously established weight transmission model known as WeightTransmitter, producing higher Recall and Precision in inferring unknown weights while incurring lower execution overheads. Although the research setting is weighted association rule mining the methods developed are equally applicable to the supervised classification context where class labels are not known for all instances, a typical scenario in many data mining applications.