Personalized Navigation: Context-Based Preference Mining Using TensorFlow

Klokkervoll, Joachim Højbak ; Pedersen, Mike ; Pedersen, Samuel Nygaard

4. term

Computer Science, Master

2017

2017-06-05

14 pages

Existing navigation systems provide route suggestions with very limited room for customization or personalization, and if present, must be configured manually by the user. We propose a context-based personalized navigation framework capable of performing personal route suggestion based on a preference towards certain road features, with the preference inferred from previously driven trips. The framework infers a small set of unique driving preferences, and use contextual information to learn personal preference choice in the given context. This allows for generalization of preferences, fast learning, and good scalability while allowing for context-based personalization within the globally optimal preference set. Training data stems from the ITS project and after filtering consists of 958 thousand driven routes mainly in North Jutland, Denmark from 458 drivers. We also use a speedmap of North Jutland provided by Aalborg University to infer speed information about roads. Lastly, OpenStreetMap is used to extract appropriate road features and construct a graph structure used for pathfinding. After training with a large real-world dataset, our framework can propose personal route suggestions for future navigation trips with an average Jaccard distance of 0.48 when compared to the actual driven route preferred by the driver. The framework is trained using a combination of clustering and classification implemented in TensorFlow. The preferences are learned by comparing driven routes to alternative routes.

Download
View record in AAU Student Projects