以商品驅動之推薦方法

Abstract

Recommendation research has achieved successful results in many application areas. However, for supermarkets, since the transaction data is extremely skewed in the sense that a large portion of sales is concentrated in a small number of best selling items, collaborative filtering based customer-triggered recommenders usually recommend hot sellers while rarely recommend cold sellers. But recommenders are supposed to provide better campaigns for cold sellers to increase sales. In this thesis, we propose an alternative ``item-triggered' recommendation to identify potential customers for cold sellers. In item-triggered recommendation, the recommender system will return a ranked list of customers who are willing to buy a given item. This problem can be formulated as a problem of classifier learning, but due to the skewed distribution of the transaction data, we need to solve the rare class problem, where the number of negative examples is much larger than the positive ones. We present a boosting algorithm to train an ensemble of SVM classifiers to solve the rare class problem and compare the algorithm with its variants. We apply our algorithm to a real-world supermarket database and use the area under the ROC curve (AUC) metric to evaluate the quality of the output ranked lists. Experimental results show that our algorithm can improve from a baseline approach by about twenty-three percent in terms of the AUC metric for cold sellers which is as low as 0.64\% of customers have ever purchased.