Prediction Of Urban Commercial Activeness And Commercial Entity Trend Based On Online Review Software

In recent years,more and more consumers get used to using online review software to search and compare stores before consumption,check in during consumption,and publish their experience and feelings after consumption.It is believed that making good use of the massive online review software data that closely related to commercial activities to predict the trend of urban commercial activeness and commercial entities has important practical significance for urban planning and investment analysis.However,on the one hand,there are still many problems in the existing prediction methods of urban commercial activeness,including how to capture the nonlinear spatial and temporal relationship,how to extract the local spatial relationship,and how to model short-term temporal relationship and periodic temporal relationship.On the other hand,the prediction methods of commercial entity trend are also facing problems such as how to extract review semantic,how to model user importance and review habit,how to simulate cluster effect and so on.The main contributions of this article are as follows:(1)As for the prediction of urban commercial activeness,a deep spatio-temporal residual neural network model(DSTRN)based on review and check-in data is proposed.About the spatial dimension,this proposed model uses local CNN to capture the close spatial relationship of adjacent regions and avoid the influence of distant unrelated regions.About the time dimension,this proposed model extracts the short-term and periodic temporal trend respectively by using 3D convolutional neural network and uses LSTM neural network to integrate the two temporal trends.In addition,in order to extract the high-dimensional spatio-temporal relations,ResNet is integrated into 3D convolutional neural network to avoid vanishing gradient and degeneracy caused by the increase of network depth.(2)As for the prediction of commercial entity trend,taking restaurant for example,a multi-view semantic business cluster effect model(SBCM)based on online review software data is proposed.About the semantic dimension,this proposed model extracts the rich semantics of reviews by the multi-granularity sentiment classification neural network,and a weight is added to reflect the importance of review,user influence and user habit.About the business attribute dimension,all business attributes are sorted by calculating their importance,and the important ones are screened out.About the dimension of cluster effect,commercial entity clusters are identified by density-based clustering method,and the commercial activeness of each cluster is calculated to reflect cluster effect.Finally,LightGBM is introduced to integrate various features,and the business trend prediction can be realized after training.(3)The open dataset from Yelp is used to verify the accuracy of the two proposed models.Experiments demonstrate that DSTRN vastly outperformed other approaches and reduce the mean square error by 51.2%,57.5%and 8.5%compared to HA,ARIMA and XGBoost.On the other hand,experiments demonstrate that SBCM outperforms SVM and XGBoost by 14.0%and 3.4%,respectively in terms of AUC.In addition,the actual effect of DSTRN model to predict the trend of urban commercial activeness and the business attribute importance ranking formed by SBCM are given.