A multi-valued and sequential-labeled decision tree method for recommending sequential patterns in cold-start situations

We plan to recommend some initial suitable single-itemed sequences like a flight itinerary based on a preference pattern in the form of personalized sequential pattern to each cold-start user. However, sequential pattern mining has never treated a conventional sequential pattern as a personalized pattern. Besides, as a cold-start user lacks the personalized sequential pattern, collaborative filtering cannot recommend one any single-itemed sequences. Thus, we first design such a preference pattern, namely representative sequential pattern, which reflects one’s main frequently recurring buying behavior mined from the item-sequences during a time period. After sampling a training-set from non-cold-start users who prefer similar items, we propose an auxiliary algorithm to mine the representative sequential pattern as the sequential class labels of each training instance. A multi-label classifier seems therefore be trained to predict the sequential-label for each cold-start user based on one’s features. However, most multi-label classification methods are designed to classify data whose class labels are non-sequential. Besides, some of the predictor attributes would be multi-valued in the real world. Aiming to handle such data, we have developed a novel algorithm, named MSDT (Multi-valued and Sequential-labeled Decision Tree). Experimental results indicate it outperforms all the baseline multi-label algorithms in accuracy even if three of them are deep learning algorithms.