[Paper Review] Extendable Neural Matrix Completion

Duc Minh Nguyen, Evaggelia Tsiligianni+1 more|arXiv (Cornell University)|May 13, 2018

Image and Signal Denoising Methods17 references4 citations

TL;DR

This paper proposes a two-branch deep neural network model, Neural Matrix Completion (NMC), for matrix completion that enables extendability to unseen users or items without retraining. By learning joint latent representations via a factorization-based architecture with convolutional summarization layers, the model achieves state-of-the-art accuracy and robust extendability on benchmark datasets like ML-1M and Netflix.

ABSTRACT

Matrix completion is one of the key problems in signal processing and machine learning, with applications ranging from image pro- cessing and data gathering to classification and recommender sys- tems. Recently, deep neural networks have been proposed as la- tent factor models for matrix completion and have achieved state- of-the-art performance. Nevertheless, a major problem with existing neural-network-based models is their limited capabilities to extend to samples unavailable at the training stage. In this paper, we propose a deep two-branch neural network model for matrix completion. The proposed model not only inherits the predictive power of neural net- works, but is also capable of extending to partially observed samples outside the training set, without the need of retraining or fine-tuning. Experimental studies on popular movie rating datasets prove the ef- fectiveness of our model compared to the state of the art, in terms of both accuracy and extendability.

Motivation & Objective

• Address the limitation of existing neural matrix completion models that cannot generalize to new users or items without retraining.
• Develop a method that maintains high predictive accuracy while enabling efficient inference on previously unseen samples.
• Enable application of matrix completion to high-dimensional, real-world datasets such as the Netflix dataset with extreme sparsity and large scale.
• Design a model that relies solely on explicit feedback (observed ratings), avoiding reliance on implicit feedback or external data.
• Ensure the model is scalable and effective across varying matrix dimensions and sparsity levels.

Proposed method

• Propose a two-branch deep neural network that maps row and column samples (e.g., users and items) into a shared latent space.
• Use matrix factorization principles, modeling user-item interactions as inner products in the latent space.
• Incorporate convolutional summarization layers with learnable filters and strides to compress input representations and enable scalability to high-dimensional matrices.
• Train the model end-to-end using observed entries in the training set via Frobenius norm minimization of the reconstruction error.
• Enable extendability by freezing the encoder branches and using the learned latent representations to infer missing values for new samples via inner product computation.
• Apply dropout regularization (ratio 0.6) on the ML-1M dataset to prevent overfitting due to limited training samples.

Experimental results

Research questions

• RQ1 Can a deep neural network-based matrix completion model generalize to users or items not seen during training without retraining or fine-tuning?
• RQ2 How does the proposed model perform in terms of accuracy and extendability compared to state-of-the-art methods on real-world, high-dimensional datasets?
• RQ3 To what extent do convolutional summarization layers improve scalability and performance on large-scale matrices like Netflix?
• RQ4 Does the model maintain strong performance when tested on data not used during training, particularly in areas with no overlap in training samples?
• RQ5 Can the model achieve competitive performance using only explicit feedback, without relying on implicit user interactions?

Key findings

• The NMC-S model achieves state-of-the-art performance on the ML-1M dataset, outperforming I-Autorec and U-Autorec in areas (II) and (III), and is the only model capable of extending to area (IV).
• On the Netflix dataset, NMC-S delivers the best performance in areas (II) and (IV), despite not being the top performer in (I) or (III), demonstrating strong generalization and extendability.
• The model maintains competitive accuracy across all evaluation areas, with RMSE and MAE values consistently below or comparable to SOTA baselines, particularly excelling in unseen data scenarios.
• The use of summarization layers enables effective handling of high-dimensional matrices, such as the Netflix dataset, where input dimensions are extremely large.
• NMC-S achieves strong performance with only explicit feedback, avoiding the need for implicit feedback or external information, which enhances applicability in real-world settings.
• The model’s extendability is validated empirically: it can predict missing entries for new users or items by leveraging learned latent representations without retraining.