Predictive (UX)Analysis: (LSTM) Model for Cursor Kinematics & Intent Modeling
Hello Community,
I am C. Rajesha, a final-year (AIML) engineering student, and I wanted to share a research-focused project I’ve developed that applies (Deep Learning) to the field of Human-Computer Interaction (HCI) and (UX) analytics.
The goal was to transition standard (UX) logging (which is reactive) into a predictive (AI) system that can infer user intent and measure cognitive load before a click occurs.
The Technical Approach
-
Data Source: Used Windows Problem Steps Recorder (MHT)files to extract raw {X/Y} coordinate time-series data.
-
Feature Engineering: Transformed raw data into kinematic features essential for behavioral modeling: Total Velocity and Total Acceleration.
-
Model: Designed and trained an (LSTM) Time-Series Classification Network to map sequences of movement (e.g., 1 second preceding an action) to specific user intents ({System Utility} vs.{Web Browser}).
Key Advantages and Insights
-
Prediction vs. Logging: The model can classify user intent with \\text{X}\\% accuracy based purely on the movement pattern, making it a valuable tool for proactive \\text{AI} assistants.
-
Cognitive Load Measurement: The analysis of acceleration variance and velocity profiles allows the system to quantify user hesitation, providing objective data to pinpoint (UI) friction points.
-
Ethical {AI}: The project incorporates a Privacy by Design policy that halts data collection when the active window title contains sensitive keywords (e.g., ‘password’).
Seeking Feedback
I would be interested in feedback, particularly on:
-
Scaling the {LSTM} architecture for real-time inference.
-
Methods for incorporating visual {UI}context (e.g., image embeddings of the button clicked) to improve model accuracy.
-
Ideas for deploying this type of {HCI} model within a large-scale application environment.
Thank you for your time!
C.Rajesha
