A User-specific machine learning approach for improving touch ...

zoomzurichAI and Robotics

Oct 16, 2013 (3 years and 10 months ago)

67 views

TOUCH INTERFACE ISSUES


Electrostatic interference (noise)


Jittery touch registration


Unintentional selection


Screen Misalignments


Users constantly miss targets


“Fat Finger Problem”


Imprecise selection

TOUCH INTERFACE ISSUES (2)


Results in:


Trivial actions requiring mental involvement


“Why won’t this button activate when I press
it?”


Users losing trust in the system


Users cannot be confident in their selections


Increasing error proneness


Users must spend more time accommodating
for mistakes


EXISTING SOLUTIONS


Electrostatic sensor sensitivity
hardware adjustments


3
-
point or 5
-
point calibration
methodology


UI Adjustments


Error state recovery improvements


Interface design alterations


We’re missing one major concept
here…

USER PROFILING


Every user performs differently


User Profiling


An association of specific data to specific
users



Why does this apply here?



How do we obtain and apply the data?

INTELLIGENT UI


Machine Learning approach


Consists of training/test phases


TRAINING PHASE


Obtain data from some source (sensors)


Process the data and generate a pattern (offset)


TESTING PHASE


Utilize the pattern to adjust the data collection
process (recalibrate)


Analyze how the adjustment affected the data
(improvements)


Lather, rinse, repeat until satisfied

HYPOTHESIS


Users possess distinct touch offsets
which hinder performance


Machine learning can be implemented
on raw data to calculate offsets


Offsets can be used to calibrate the
touch screen to provide a more
consistent interface for the user


Finally, the correction procedure will
greatly improve user touch accuracy

CAPTURING THE
DATA


Nokia N9
MeeGo


Capturing sensor data


Uses
Guassian

Process
Regression


THE EXPERIMENT


Environment


Uses the tester designated model phone


Nokia


Program on the phone prompts the user to
touch crosshairs


Records intended location (crosshair location)
and physical touch location


Phone held in landscape position


THE EXPERIMENT (2)


Test Population


8 Different Participants


Age between 23 and 34


Most experiment subjects owned and
regularly used
smartphones

though this
wasn’t a requirement

THE EXPERIMENT (3)


Procedure


Prompt the user to touch 1,000 crosshairs


Record intent/actual touch data for each
attempt


Split data into training/testing phase


Repeat sets of 1,000 attempts for alternative
experimental cases


THE RESULTS


1
st

Experiment


Used raw values from sensors


Data analysis with 3 button radii and 2
training set sizes


Resulted in statistically significant results

THE RESULTS (2)


2
nd

Experiment


Used interpreted touch location information
specified by the phone


Decreased button radius for each set


Resulted in statistically significant improvements
with even smaller training sets


FURTHER EXPERIMENTATION


Testing in alternative cases to ensure
genuine data


Two users experimented in portrait mode as
opposed to landscape


Alternative phone models were used for
some users


Data still showed statistically significant
improvements in touch accuracy when
the offsets were applied to the test data

CONCLUSIONS


Utilizing a ML approach proves a viable
solution to user
specifity


This solution is versatile


Research was thorough but for a small
sample size


Future work is necessary to further the
study