Decoupled Architecture for Privacy-Compliant Recommendation

Overview

This repository implements a decoupled recommender system to address the "Right to be Forgotten" in production environments. Standard deep learning models (like monolithic Transformers) mix user and item data, which makes deleting a single user impossible without retraining the entire model.

This project uses a Decoupled Architecture that physically separates item learning from user history processing. This separation allows for the effective removal of user data by discarding local parameters while keeping the global item model intact.

System Architecture

The solution combines Neural Collaborative Filtering (NCF) and Self-Attentive Sequential Recommendation (SASRec):

1. Item Tower (Cloud / Public): A Two-Tower network trains on global interactions to learn item-to-item compatibility. These embeddings are extracted and frozen, preventing user-specific gradients from leaking into the central model.
2. User Tower (Edge / Private): A local Adapter Layer (Linear-ReLU-Linear) projects the frozen item embeddings into a latent space for a SASRec Transformer, which learns sequential history locally.

Data & Methodology

• Dataset: KuaiRec 2.0 (Filtered for positive interactions with watch_ratio ≥ 1.0).
• Debiasing: Applied Inverse Propensity Weighting (IPW) during training to fix popularity bias and ensure niche content representation.
• Validation Method: Exact Retraining. To verify the unlearning, we trained two distinct models from scratch: a "Baseline" (containing all users) and an "Unlearned Model" (excluding the target user). This provides a mathematical ground truth for evaluation.

Evaluation Results

1. General Utility (Hit Rate @ 10)

Impact on the general user base was measured using Hit Rate @ 10 on a held-out test set.

Figure 1: Comparison of recommendation quality. The unlearned model demonstrates stable utility (0.0260) compared to the baseline (0.0240), confirming that unlearning a user does not degrade system performance for the remaining population.

2. Privacy Verification (Cross Entropy Loss)

Erasure was verified by calculating Cross Entropy Loss on the target user's historical training sequences. A spike in loss indicates the model has lost its predictive capability for that specific user.

Figure 2: Privacy verification. The sharp increase in loss (3.61 → 8.32) for the unlearned model confirms the effective removal of the user's data patterns, approaching random guessing levels.

Repository Structure

• 01_Data_Prep_Debiasing.ipynb: Data preprocessing, sequence generation, and calculation of Inverse Propensity Weights (IPW).
• 02_Cloud_Item_Tower.ipynb: Training of the Two-Tower candidate generation model and extraction of the frozen item embedding matrix.
• 03_Edge_SASRec_Training.ipynb: Implementation of the Adapter Layer and SASRec Transformer. Execution of the "Exact Retraining" simulation (Baseline vs. Unlearned models).
• 04_Evaluation.ipynb: Calculation of Utility (HR@10) and Privacy (Cross Entropy) metrics, and generation of visualization charts.

References

1. KuaiRec Dataset: Gao, C., et al. (2022). "KuaiRec: A Fully-Observed Dataset...". CIKM.
2. IPW: Schnabel, T., et al. (2016). "Recommendations as Treatments: Debiasing Learning and Evaluation". ICML.
3. SASRec: Kang, W. C., & McAuley, J. (2018). "Self-Attentive Sequential Recommendation". ICDM.