First page Back Continue Last page Overview Text


I see Reality Mining as bridging these two approaches – collecting real, objective data and with such quantity and detail to be satisfactory to the machine learning community as training data.

A main driver of the idea is to leverage the tools that are quickly becoming ubiquitous within the workplace – cell phones and PDAs. For our initial experiments we are using a system of linux PDAs and either wired, or wireless cell phone headset microphones.

Depending on the system, the individuals simply keep the PDA in their pocket, or in a briefcase or purse if they areusing the bluetooth mics, and audio is continuously streamed over the wireless network to a central server which is also collecting AP and wireless traffic info for each user to give a rough idea of location and others in the users proximity

The audio analysis leads to an interference of the user’s situation and activity patterns, given some training data.
We are running commercial speech recognition engines to mine out keywords and infer topics.
And given multiple synchronized streams of audio we can establish who is involved in the interaction, how they are interacting.

This analysis leads us to a large set of features which then can be input into probabilistic models for parameter estimation and prediction.