Research On Point-of-interest-level Crowd Flow Prediction Based On Incremental Learning

Point-of-Interest-level(POI-level)crowd flow prediction is a sub-task of urban traffic prediction,and is of great significance to POI managers and visitors.Based on POI-level crowd flow prediction,POI managers can make more reasonable business arrangements,and visitors can also make more approval travel plans.POI-level crowd flow prediction is currently less researched and has many challenges.Compared with region-level crowd flow,the area of POI is smaller and the fluctuation of POI-level crowd flow is greater.Besides,there are diverse temporal correlations of different POIs and varies over time.In the practical application of POIlevel crowd flow prediction,the distribution of crowd flow would change over time,and the trained model cannot achieve the best performance on the real-time data stream.To address these challenges,the main work of this paper is as follows:We propose a multi-step POI-level crowd flow prediction model,MsPLCFP.To address the large fluctuation of POI crowd flow,a multi-scale temporal attention mechanism which contains multiple different scales of temporal attention is employed to smooth input crowd flow at lower level and capture global dependencies of input crowd flow at higher level.To capture the diversified temporal attention,meta knowledge learner is used to obtain meta knowledge from POI category,POI popularity,etc.Then meta-RNN+is applied to model the relations between temporal correlations and meta knowledge so as to capture diverse temporal correlations.We have conducted experiments on two datasets,and compared the performance of the model with several mainstream time series prediction models to verify the effectiveness of our model.We evaluated Ms-PLCFP on two real-world datasets and Ms-PLCFP achieved significant improvements over the baselines,which shows the effectiveness of our model.We propose an online POI-level crowd flow prediction framework based on incremental learning to adapt to changes in data stream.The framework contains two part:a concept drift detection part and a model update part.The concept drift detection part uses UDD for detection,and the model update part uses AFEC for model update.We design a experiment to verify the performance of our framework,and compare the fitting error,prediction error,and warm-up error with the baseline model.The experimental results show that the proposed framework has excellent performance in alleviating the problem of catastrophic forgetting and also achieve the best performance on future prediction.