About me
Experiments of Pattern Recognition and Anomaly Detection in Observed Daily Activity Travel Sequences using Multichannel Autoencoders. - Developing taxonomies of patterns of daily activities is something that has gone on for a while, and it is usually at the center of most activity-based models for travel demand forecasting. However, with the increase in the size of data available and a plethora of heterogeneous behavior, alternate ways are required to develop pattern taxonomies that are also designed to distinguish between “normal” behavior from “anomalous” behavior. Activities reported by survey participants in a diary are represented as sequences of occurrence in a day. Hence, this study adopted a novel Multichannel Autoencoder method, where each of a finite set of sequences of activities were analyzed separately distinct channels of information. The Autoencoder is then used for dimension reduction with each resulting latent space used as a representation to cluster each of the activities individually. Using the silhouette scores, any activity that had very low scores when independently clustered pointed to an activity that does not play an important role in distinguishing activity sequences from each other. In essence, this method reveals to us the most informative activities, meaning that these informative activities can be used as the minimum amount of information to find distinct patterns instead of the whole sequence of activities. In this way, we will still get very similar results of pattern identification and classification as when the complete sequence of activities is used. This brings in the benefits of faster analysis and then it brings about the possibility of real time pattern recognition. For the anomaly detection, using the Multichannel Autoencoder reconstruction errors from the encoded and decoded activities are used. In addition, using indicators such as entropy and complexity are also useful for the characterization of the sequences analyzed here.
