The Research About Sequential Recommendation Technics In Mobile Environment

In recent years, with the explosion of Internet data, recommendation techniqueshavebeen proved to be an effective approach to solve the issue of information overload,andhave become a hot research topic in artificial intelligence, data mining, machine learningand other related areas. On the other hand, with the rise of mobilecommunicationtechnology and the popularity of mobile equipment, the combination of traditionalrecommendation techniquesand mobile environment has gradually become a newresearch direction. However, the specialcharacteristics of the mobile environmentpresentnew challenges to the direct application of traditional recommendation techniques. Inparticular, people’s behavior and activities are always with sequential propertiesin themobile environment, and sequence will become an important factor in decision-makingin many application and scenarios. So in this paper, weconduct the deep research onsequential recommendation problem, and takethe scenarios of passenger-finding locationin an urban taxi service as our research subject. Weinvestigate this problem in thefollowing three aspects:First, we make a research on multi-point sequential recommendation problem.Specifically, we are based on taxi GPS trajectory data sets, and establish the multi-pointsequential recommendation model based on Markov decision process, as well as thealgorithm to solve the model. Simulation results show that, compared with the traditionalTopK methods, thismodel could give better recommendation results, which bring moreexpected profit for taxi drivers.Secondly, we focus on preferences and context based sequential recommendationinorder to give more personalized results for drivers. Specifically, we will analyze the dataset of high-profitdrivers (which we regard them as experts), and applyinversereinforcement learning and apprenticeship learning to mine these experts’ knowledge,and then learn their corresponding reward function R. To test the accuracy of therecommendation result, we conductthe experiments in three different prediction tasks andmake a comparison with other approaches. The results demonstrate that the proposedmethod improve the prediction accuracy significantly.Finally, we do the sequential recommendation under the multi-agent environment.Specifically, we are based on taxi GPS trajectory data sets to estimate the differenttransfer rate, and establish a driver and passengers’ model based oncontinuous-timeMarkov chain. Meanwhile, we makestochastic gamesto model the taxidrivers’passenger-finding behavior, and proposealgorithms tocalculate a Nashequilibrium strategy, which reflect the competitive nature of the passenger-findingbehavior betweendifferent drivers. Simulation results show that, compared with the original strategies which do not consider the multi-agent competitive environment, theNash equilibrium strategy presented in this paper can better reduce the waiting time forpassengersin a real environment, and thereby increasing the expected profit for taxidrivers.