normalizing attributes

normalizing attribute values

Hi all,
I am a researcher in the field of Human Computer interaction.
Now, I am doing research on automatic detection of users skill level
using user interface events.
Machine learning algorithms were used to build statistical predictive
models of skill (decision trees).
we asked some subjects to perform a specific task for 15 trials, so that they acquire skill by repetition.
The specified task was a seven step image manipulation job with a paint software.
All their interactions were logged automatically. We have built our classifiers based
on the attribute values extracted from these log files. Attribute values such as as task completion time, mouse distance traveled,mouse velocity and acceleration, pause counts and etc. were extracted from logged high frequency user interface events(mouse and keyboard events).
instances (attribute values) from early trials of users were labeled as novice and high trials were labeled as skilled.
First we built task-dependent classifiers,
i.e. the classifier is trained and built using attribute values extracted from
interactions to perform a specific step or task in a specific application.
Thus, this sort of classifiers can only be used to classify new instances of attribute
values from the same task in the same application.
For example, consider mouseDistanseTravelled (distance traveled by the mouse to perform a task in an application in pixels) attribute, the values for this attribute learn by a classifier for a specific step or task are dependent on that task and cannot be used for classification in other steps.
In order to make the classifiers reusable in other applications and tasks, We normalized the attributes. Normalizing means, making the range of attribute values identical or making attribute values from different tasks comparable to each other. we considered that some of the attributes were intrinsically normalized for a diverse range of applications and tasks (perhaps with a little approximation) and no further normalizations were necessary, such as mouse velocity.
But, some of the attributes such as mouseDistanseTravelled required normalization. e.g. To normalize distanceTravelled, it was divided by the minimum distance required to be traveled by the mouse to perform that task. Or in order to normalize pauseCount (number of pauses during performance of a task), it was divided by the number of actions (e.g. a mouse click is an action ) performed in that task. This new quantity gives us the average number of
pauses taken place before an action. then we trained our classifiers using this normalized values and built task-independent classifiers that can be used to classify instances from every desired UI task.

My question is:
is our approach in making our classifiers task-independent correct?
are there theories in machine learning for this sort of problems?
does my problem relates to multi-task learning?
Multi-task learning is an approach to machine learning that learns a problem together with other
related problems at the same time, using a shared representation. This often leads to a better
model for the main task, because it allows the learner to use the commonality among the tasks.
but as i interpret, my problem does not relate to multi-task learning, because
i have completely identical tasks and not similar (I want to classify the skill level in different ui tasks).
but multi-task learning is used in situations when multiple related (but not identical) tasks in a domain are
going to be learn such as,
learning the phonemes and stresses to give a speech synthesizer to pronounce the words given it as inputs, or
given a newswire story, predicting its subject categories as well
as the regional categories of reported events based on the same text.

I will be thankful for any guides, comments or any ideas related!
Arin

normalize attributes

Hi,

looks like u r doing pretty interesting project! kudos!!!

here is waht i think abt it.
1. abt the task independence of features. the idea of using normalization is a good one. but i have a few things to question.
say u want to identify the skill of a subject then u would be ask the subject to perform all the 15 trails to determine its skill. if this is true then the attributes of all the trials can be combined to form the attributes of a subject and hence can be used to classify the skill of each subject. if this is not as it works then pardon me for my ignorance.

2. multi tasking is not wat u r doing. multi tasking is learning more than one task at the same time with same set of attributes...or say using the same info.

Hope this helps n waiting to hear frm u

Rakesh

attributes of member for normalization

hi,
1)I try to detect two kinds of skill: general skill (user's general computer usage expertise) and task skill (skill in doing this specific paint task).
task skill is improved by the trials, i.e. data from the first trial is labeled as novice in Task_skill and data from 10-15th trials is considered as skilled in task skill.
we already know users general skill by an interview, for example the data from first trial of a user with skilled general skill is labeled skilled/novice (general skill/ task skill).

2) as far as i know, there are two typs of multi task learning :
1- same data but different tasks, such as detecting news region and category from its text.
2- different data but identical task: predicting some users will like a special painting. in this case a seperate classifier is created for each user. tasks are identical: whether each user will like a specfifc painting or not?. but the data for each user is different (previous log of the paintngs the user had liked). in this case the data from other users is used to improve the individual user's classifier performance

hi arin_g
good job!
Your approach is on the correct path
I'd suggest that you may need to use one more level of optimization using pattern based data gathering to enhance the intelligence,
hope this gives you a near human behavior

thanks for your reply,
what do you mean by "more level of optimization using pattern based data gathering", if u can please explain more.

Hi arin_g
What I meant was quite similar to the way we learn,
we can make a computer learn.
Have you studied the experiment named "Skinner's Rats/Mice"
We learn the same way. My suggestion was to improvise the learning process by adding one more optimizer based on different responses by the same person to the same test at different time instances.
that depends on the fact that two people trained in exactly the same manner can react to a situation in completely opposite way.
thats human nature...
guess that makes it a bit clearer.

hi arin_g,
I recon that you've not replied.
does it mean you got what I said or you are still confused.
If its the latter, My apologies...
Please let me know how can I be of help...
regards