Inability to identify where a user is based at a given time affects taxi services from providing a personalized customer experience. Such information would improve customer experience, provide a personalized user experience and provide a targeted customer base for the geo-based promotions/campaigns. Predicting human mobility is not an easy task. But taxi services do have a history of trips taken by each user with pick-up, dropoff locations with time. It's known that humans tend to have cyclic behavior with similar patterns, which makes it possible to predict the mobility of a user from his/her previous trip information. Using these information, it is possible to predict passenger mobility. Once mobility is predicted, it is possible to derive the user location at a given time. Existing techniques presented using Markov models are mainly tested against continuously tracked user trajectories. But none had been tested in the domain of taxi. In a domain of taxi, data set does not contain continuously tracked user locations for each individual passenger. It will only contain locations if user decides to travel in a taxi. In addition, the traditional Markov model approaches can only predict the next location but it is unable to predict the duration or the arrival time. Non-Markov model approaches such as non-linear time series analysis require a large amount of individual user locations for an accurate prediction. Our proposed methodology attempts to combine both these concepts used in Markov and non-linear time series analysis. Proposed methodology predicts the user location on a given time using a Markov inspired model. This prediction will be based on a given individuals' past pick-up and drop-off history while giving priority to the last drop off location to create a sequence of predicted events. This predicts the time; a specific passenger will leave his/her current location and arrive at the next location. Our proposed methodology proved to be more accurate towards frequent users. Frequent users have more trip information to derive location transition probabilities and to accurately calculate the duration of stay at each location which is critical for proposed methodology.
