Location Prediction Algorithm Based On LSTM-CNN

With the rapid development of Internet technology and continuously increasing customers' needs,various information-based products have launched.People enjoy the convenience and efficiency brought by these applications,and meanwhile the amount of user data is collected.Location data is one of the basic information among the user data,which is easy to collect and able to reflect users' habits and daily routines.While location prediction models,as the most direct method of using location data,providing the essential help for user analysis,as well as the products' update and iteration.Most traditional location prediction models can only deal with the short time series data by establishing markov transition probability matrix or frequent pattern,so how to mine the information of long time series data is a main difficulty.In addition,the granularity of the original location data is too fine,it must be processed before using because it can't be used as the prediction target directly.How to use the appropriate method to determine the location prediction target is also one of the problems to be solved.In this thesis,a position prediction model(MBi-LSTM-CNN)based on the fusion of Multilayer Bidirectional Long Short-Term Memory(LSTM)and Convolutional Neural Networks(CNN)is built to solve the above two problems.The main work of research is as follows:(1)A new position prediction model MBi-LSTM-CNN is proposed.Based on the analysis of the traditional position prediction technology,this model can mine the hidden information of long time series through multilayer bidirectional LSTM structure,and then extract the local spatial features through CNN.(2)According to the Geo Hash,a spatial indexing technology,this thesis proposes a vector coding scheme,while a new vector editing distance is put forward based on this scheme as well.The vector coding scheme can simplify the location prediction target and save more memory space.Specifically,it can transform the original fine-grained longitude and latitude data into six-bit length vector which can be predicted directly.(3)A new method based on time entropy is proposed to obtain the segmented sequence of position data.To be specific,by introducing the time entropy formula,calculating the frequency of each user's location data and dividing different time intervals,the segmented location sequences can be recorded.Meanwhile,this method is able to determine the objective threshold of time division,avoiding the subjective difference caused by manual operation.(4)Seven different LSTM structures are built,considering that different LSTM structures and model parameters have influence on each experiment.Also,the model structure is adjusted with the combination of different optimizer and learning rate.The verification result of experiments on Geo Life data set shows that the location prediction model proposed in this thesis is more accurate than traditional location prediction models,in terms of predicting location.