A landmark model for assigning item weight for pattern mining
MetadataShow full metadata
In weighted association rule mining, items are typically weighted based on background domain knowledge. However, it may not be feasible to gather domain information on every item in high dimensional datasets especially in a dynamically changing environment. Thus, it is more practical to exploit domain information to set weights for only a small subset of items and then estimate the weights of the rest through the use of a suitable interpolation mechanism. In the recent study (Koh et al., 2012), weight transmitter model was proposed. The weight transmitter model uses a subset of items, termed landmark items, whose weights are known in advance to propagate known weights to the rest of the items with unknown weights. In this study, we seek to extend this approach by improving performance of the weight transmitter model while seeking to lower the percentage of landmark items employed in the weight estimation process. Firstly, we propose a new interestingness measure called Proportional Confidence, which is derived from the standard confidence measure, to use as a measure for quantifying interactions between items. Secondly, we propose a novel method to partition a global graph into a number of smaller sub-graphs called Sub-graph generation algorithm by utilizing divide-and-conquer approach. Thirdly, we propose a new method used in allocating landmark items by utilizing stratified random sampling approach. The results of our experiments show that our proposed landmark items assignment produces higher performance in terms of Precision, Recall, Accuracy, Lift and Execution Time compared to the original simple random sampling while our proposed sub-graph approach substantially reduces time complexity in the weight fitting process. We also investigate the impact of our proposed weight transmitter approach compared to weighting with the domain based approach in relation to cases where sharp differences arose in the assignment of weight values to the same item. The results from the in depth study show that our proposed weight transmitter approach is in a better position to assign item weight as it takes into account interactions between items.