E4UL: efficient unlearning for recommender systems

B.S. thesis. Removes specific users' data from a trained recommender without full retraining: matches retraining accuracy (~69.3%) with ~5× less compute.

Status: Research
Period: Nov 2024 – Jul 2025
Role: B.S. thesis, VNUHCM – University of Science
Code: Code not public
Stack: Python
PyTorch
MovieLens
Goodreads

Accuracy, matching full retraining: ~69.3%
Less computation time than retraining: ~5×

Trained recommender
Deletion request
Unlearning update
Recommender without that data

Fig. 2 Data flow, left to right (top to bottom on small screens).

Problem

Privacy requests such as the right to be forgotten mean a trained recommender must forget specific user data. Retraining from scratch every time is expensive.

Approach

I developed an efficient machine unlearning method that removes the influence of specific user data from a trained recommender without retraining from scratch.

Architecture

A deletion request names the data to forget. The unlearning step updates the trained model to remove that data's influence, and the result is evaluated against a model fully retrained without the data.

Results

Matched full retraining accuracy (~69.3%).
Cut computation time by ~5×.
Evaluated on the MovieLens and Goodreads datasets.