Ad Hoc Carpooling Prediction Based on Improved kNN

At present, taxi applications have been popularized in our lives. Carpooling has become a new trend in saving money and protecting environment. Two kinds of carpool applications in today's market are mainly focused either on ad hoc carpooling service for express cars or reservation service for shuttles. Normally in a real-world carpool scenario, a prior estimated fare is the mere reference available for a rider to make further decision, which is actually insufficient in terms of user experiences, e.g. whether to keep waiting as in a period of taxicab shortage during rush hours or after major events. This paper presents a novel algorithm to ad hoc carpooling prediction based on improved kNN for practical use. First, based on the idea of density reachability, a set of reasonable determined neighbors can be effectively identified and making the subsequent voting process more accurate. Second, with awareness of data distribution both across and within categories, the accuracy of noise neighbor detection may less be influenced by the uneven distribution of data. It is proved that the algorithm does have positive effects on unbalanced data through experiments with an industrial carpool dataset.