Life patterns : structure from wearable sensors

Abstract

In this thesis I develop and evaluate computational methods for extracting life's patterns from wearable sensor data. Life patterns are the reoccurring events in daily behavior, such as those induced by the regular cycle of night and day, weekdays and weekends, work and play, eating and sleeping. My hypothesis is that since a "raw, low-level" wearable sensor stream is intimately connected to the individual's life, it provides the means to directly match similar events, statistically model habitual behavior and highlight hidden structures in a corpus of recorded memories. I approach the problem of computationally modeling daily human experience as a task of statistical data mining similar to the earlier efforts of speech researchers searching for the building block that were believed to make up speech. First we find the atomic immutable events that mark the succession of our daily activities. These are like the "phonemes" of our lives, but don't necessarily take on their finite and discrete nature. Since our activities and behaviors operate at multiple time-scales from seconds to weeks, we look at how these events combine into sequences, and then sequences of sequences, and so on. These are the words, sentences and grammars of an individual's daily experience. I have collected 100 days of wearable sensor data from an individual's life. I show through quantitative experiments that clustering, classification, and prediction is feasible on a data set of this nature. I give methods and results for determining the similarity between memories recorded at different moments in time, which allow me to associate almost every moment of an individual's life to another similar moment. I present models that accurately and automatically classify the sensor data into location and activity.

(cont.) Finally, I show how to use the redundancies in an individual's life to predict his actions from his past behavior.