Semantic Annotation of GPS Traces: Activity Type Inference

Due to the rapid development of technology, larger data sets concerning activity travel behavior become available. These data sets often lack semantic interpretation. This implies that annotation in terms of activity type and transportation mode is necessary. This paper aims to infer activity types from GPS traces by developing a decision tree-based model. The model only considers activity start times and activity durations. Based on the decision tree classification, a probability distribution and a point prediction model were constructed. The probability matrix describes the probability of each activity type for each class (i.e. combination of activity start time and activity duration). In each class, the point prediction model selects the activity type that has the highest probability. Two types of data were collected in 2006 and 2007 in Flanders, Belgium, i.e. activity travel data and GPS data. The optimal classification tree constructed comprises 18 leaves. Consequently, 18 if-then rules were derived. An accuracy of 74% was achieved when training the tree. The accuracy of the model for the validation set, i.e. 72.5%, shows that overfitting is minimal. When applying the model to the test set, the accuracy was almost 76%. The models indicate the importance of time information in the semantic enrichment process. This study contributes to future data collection in that it enables researchers to directly infer activity types from activity start time and duration information obtained from GPS data. Because no location information is needed, this research can be easily and readily implemented to millions of individual agents.