Behavioral cloning
Behavioral cloning is the supervised form of imitation learning: train a model to predict the expert's action from the state the expert saw, over a dataset of state-action pairs. A DOM-level session recording is that dataset's raw material, since it stores exactly the paired sequence of interface states and user actions.
How it works
The pipeline is direct: each recorded moment where a user acted becomes a training example, with the interface state (the reconstructed DOM at that timestamp, or a derived representation of it) as input and the action taken as label. The method's known weakness is compounding error, since the cloned policy drifts into states the experts never visited and has learned nothing about them; mitigations range from corrections data to reinforcement fine-tuning, all of which consume more interaction recordings. The practical dependency runs through data quality: cloning reproduces whatever the demonstrations contain, including user confusion, so corpus curation (task-completed sessions, low frustration scores, deduplicated flows) does much of the work. Consent and licensing govern all of it: sessions recorded for analytics are not automatically usable as training data, and the permission question belongs in the data design, not after it.
Where you encounter it
Agent training pipelines, robotics and web-agent papers, and dataset licensing discussions.