Machine Learning and Classification Problems

At this point, I want to evaluate the classes and class tests I have designed.  I want to evaluate their error rate, do they work at all?  Do they tell me what action a user is performing with any reliability? If they do not work reliably, are they fundamentally flawed or is the way they are being used causing the error?(do they need to be weighted)  If they do work reliably what can be gained from having this knowledge in realtime.  How can I leverage the data to reveal something, and on what time scales does the data work best/have the most impact?

I have studied a few very basic machine learning algorithms and will also present this research in more detail at a later time when it is more complete.  For now I will use the quick survey I have taken to discuss my initial plan for my dataset and classifier testing and possible more complex trajectories.

1) Implement a simple counting algorithm: iterate through all data, perform needed analysis (moving window integration, smoothing, peak detection and fill variables and arrays)  — run all tests from previous post and add “1” to variables for different classes, I.E.

IF TEST 1 = TRUE{

line horizontal  =++

}

Do nothing if the test is FALSE, just continue iterating through all tests and counting positives.

2) Weighting the results of the tests so that classes that have greater chances to be true are weighted less.  This would be to fix the lopsidedness of the number of tests for certain classes.

3) Integrating a false response to a test to subtract from a class counter…possibly also to weight the subtractions, so that multiple tests returning false more greatly impact a variable than only one test returning false.

4) Implementing decision trees to a) work through each test in a certain order making distinctions in the order of testing and combining certain tests to get to a classification or b) using decision trees to analyze the counter variables and make final classifications about the current and past actions of the user.

5) Rigorous research needed to implement Markov State Chains with transition states and weighting of those transitions.  These state chains are very powerful statistical models and are commonly used for all sorts of state based problems.  They are used for generating text or audio as they can be used to predict common next words and sounds based on previous states in the chain.  This seems a promising avenues for research one action in sketching is preceded by a transitional state, the probability of which would be a combination of the specific users past and physical reality.

Advertisements
This entry was posted in Thesis. Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s