AI for prediction of usage pattern for a car sharing service

Important results from the project

The aim of the project was to identify how AI-based data analysis can be used on existing data from our car sharing service. This goal has been met by the project.

Expected long term effects

The regression problem using LSTM gave reasonable results (similar to what a human had predicted), but they differed from the real trips. This is most likely a problem with randomness in the data. When the prediction of trips was changed to a simpler problem where the end positions were clustered together w.r.t geographical position, the correlations could be predicted. The larger classes were predicted correctly to a high extent while the network had more problems with the smaller classes. The binary prediction on simulated data succeeded in predicting the relationships we simulated.

Approach and implementation

The algorithms were implemented and evaluated using Kera´s library. Throughout the project, a literature study has been done. The regression problem using LSTM was evaluated using different metrics for regression problems, as well as by plotting the positions on a map. The prediction of the end positions as a classification problem was evaluated by studying the confusion matrix. The prediction of the simulated data was studied by visualizing the predicted outcome depending on the parameter.