|
I need to address a problem of predicting the number of steps, taken by a user, using mobile sensor data generated by accelerometer, gyroscope and so on. But I'm not sure of the way I formulate this problem (or is it even a machine learning problem?). Sensor data will be sampled at some rate (e.g., every 1 second). At each sampling point, the sensor data are used for predicting whether a step was taken. So it seems obvious that this is a supervised machine learning problem. However, the groundtruth collected from a user is not acquired at the same level, but the number of steps taken during some time. For example, I ask a user to walk for 20 seconds and get his/her feedback of the actual number of steps taken. (I can't ask the user to label the groundtruth at each sampling point because it's impossible for the user to know that) Is this still a supervised learning problem or something else (like an optimization problem)? How should I formulate this problem appropriately? |
|
I am guessing it could be a Hidden Markov model or something similar. ie you have a hidden discrete time/state process which generates a continuous set of measurements... your job is to recover hidden state transitions from the measurements |
A problem can be either supervised or unsupervised, optimization is what you do to solve any of those two.