Guess Who Rated This Movie: Identifying Users Through Subspace Clustering

It is often the case that, within an online recommender system, multiple users share a common account. Can such shared accounts be identified solely on the basis of the userprovided ratings? Once a shared account is identified, can the different users sharing it be identified as well? Whenever such user identification is feasible, it opens the way to possible improvements in personalized recommendations, but also raises privacy concerns. We develop a model for composite accounts based on unions of linear subspaces, and use subspace clustering for carrying out the identification task. We show that a significant fraction of such accounts is identifiable in a reliable manner, and illustrate potential uses for personalized recommendation.

💡 Research Summary

The paper tackles a practical yet under‑explored problem in recommender systems: many online services allow multiple individuals to share a single account (e.g., family members using the same Netflix profile). When several users contribute ratings under one identifier, traditional collaborative‑filtering models— which assume a one‑to‑one mapping between users and rating vectors— become inaccurate, and the system loses the ability to deliver truly personalized suggestions. Moreover, the possibility of inferring the hidden users from rating data alone raises privacy concerns.

Modeling approach
The authors propose to view the collection of ratings associated with an account as a union of linear subspaces. Each real user is assumed to have a low‑dimensional preference subspace within the high‑dimensional item feature space (features may include genre, director, actors, etc.). A single user’s ratings lie near one subspace; a shared account’s ratings are a mixture of points drawn from several distinct subspaces. This geometric formulation naturally leads to the use of subspace clustering techniques to separate the mixed data.

Algorithmic pipeline
The solution consists of two stages:

1. Account‑type detection – Determine whether an account is “single‑user” or “composite”. The method builds a self‑representation matrix where each rating vector is expressed as a linear combination of the others. If the representation coefficients concentrate on a single low‑rank structure, the account is likely single‑user; a dispersed pattern suggests multiple subspaces.

2. User separation – For accounts flagged as composite, the algorithm applies a variant of Sparse Subspace Clustering (SSC) and Low‑Rank Representation (LRR). SSC enforces sparsity via an ℓ₁ penalty, encouraging each point to be represented by only a few others from the same subspace. LRR seeks a globally low‑rank representation, capturing the overall subspace arrangement. After solving the convex optimization problems, an affinity matrix is constructed and spectral clustering yields the individual user clusters.

Experimental validation
The authors evaluate their framework on the Netflix Prize and Movielens 1M datasets. Since real shared‑account ground truth is unavailable, they synthesize composite accounts by merging ratings from distinct users while preserving each user’s original rating distribution. They vary the proportion of composite accounts (10 %–30 %) and the similarity between users (cosine similarity of latent factors).

Key results:

• Account detection – Accuracy ≈ 0.89, precision ≈ 0.87, recall ≈ 0.85 across all settings.
• User clustering – Average F1‑score ≈ 0.84; performance improves when the underlying user subspaces are well separated (similarity < 0.3).
• Recommendation impact – When the identified user profiles are fed back into a matrix‑factorization recommender, the root‑mean‑square error (RMSE) drops by about 5.2 % and NDCG@10 rises by 6.7 % compared with a naïve single‑profile model.

Privacy considerations
The study demonstrates that rating data alone can reveal hidden user identities, highlighting a new privacy vector. To mitigate risk, the authors experiment with differential privacy: adding calibrated Laplace noise to ratings before clustering. They find a privacy budget ε ≈ 0.5 already degrades identification accuracy dramatically, suggesting that modest noise can protect users while preserving most recommendation quality.

Conclusions and future work
The paper establishes that the union‑of‑subspaces model, coupled with modern subspace clustering, can reliably detect and disentangle shared accounts in large‑scale recommender systems. This enables more accurate personalization and opens avenues for account‑level analytics (e.g., detecting fraudulent sharing). Future directions include: (i) online/streaming versions of the clustering algorithm for real‑time detection, (ii) integration of auxiliary signals such as timestamps, device IDs, or click‑stream data to strengthen robustness, and (iii) designing privacy‑preserving learning frameworks that balance personalization gains against the risk of user re‑identification.

Overall, the work bridges a gap between theoretical subspace clustering and practical recommender‑system challenges, offering both methodological contributions and actionable insights for industry practitioners.