[Paper Review] Learning Deep Latent Spaces for Multi-Label Classification

Chih‐Kuan Yeh, Wei-Chieh Wu+2 more|arXiv (Cornell University)|Jul 3, 2017

Text and Document Classification Technologies18 references52 citations

TL;DR

C2AE integrates Deep Canonical Correlation Analysis and autoencoders to learn a feature-aware latent space for joint feature/label embedding and label-correlation aware prediction, with robustness to missing labels.

ABSTRACT

Multi-label classification is a practical yet challenging task in machine learning related fields, since it requires the prediction of more than one label category for each input instance. We propose a novel deep neural networks (DNN) based model, Canonical Correlated AutoEncoder (C2AE), for solving this task. Aiming at better relating feature and label domain data for improved classification, we uniquely perform joint feature and label embedding by deriving a deep latent space, followed by the introduction of label-correlation sensitive loss function for recovering the predicted label outputs. Our C2AE is achieved by integrating the DNN architectures of canonical correlation analysis and autoencoder, which allows end-to-end learning and prediction with the ability to exploit label dependency. Moreover, our C2AE can be easily extended to address the learning problem with missing labels. Our experiments on multiple datasets with different scales confirm the effectiveness and robustness of our proposed method, which is shown to perform favorably against state-of-the-art methods for multi-label classification.

Motivation & Objective

• Motivate and address the challenge of predicting multiple labels per instance in real-world data.
• Propose a deep learning framework that jointly learns feature and label embeddings in a common latent space.
• Incorporate a label-correlation aware loss to exploit cross-label dependencies during decoding.
• Extend the model to handle training data with missing labels without significant performance loss.

Proposed method

• Propose Canonical-Correlated Autoencoder (C2AE) that combines Deep Canonical Correlation Analysis (DCCA) with an autoencoder.
• Learn a shared latent space L by optimizing a joint loss: Phi(Fx,Fe) for feature/label embedding and Gamma(Fe,Fd) for label output recovery.
• Phi enforces alignment of Fx(X) and Fe(Y) in L with orthogonality constraints to mimic CCA (F_x(X)F_x(X)^T = F_e(Y)F_e(Y)^T = I).
• Gamma implements a label-correlation aware decoding loss that compares positive vs. negative label pairs to preserve label co-occurrence.
• Extend to missing labels by using known positive/negative pairs to compute Gamma and preprocessing to mask missing labels.
• Optimize end-to-end via gradient descent updating Fx, Fe, and Fd.

Experimental results

Research questions

• RQ1 Does C2AE outperform state-of-the-art label embedding and DNN-based methods on standard multi-label datasets?
• RQ2 Does incorporating a deep latent space that ties features and labels improve prediction accuracy by exploiting label dependencies?
• RQ3 Can C2AE handle missing labels without substantial degradation in performance?
• RQ4 How does the latent space dimensionality relative to the label space affect performance?

Key findings

• C2AE achieves favorable performance against label-embedding baselines on multiple datasets, exploiting nonlinear feature/label embeddings and label dependencies.
• On NUS-WIDE, C2AE attains the best macro-F1 and micro-F1 among DNN-based methods, with Macro-F1 48.6 and Micro-F1 67.6.
• C2AE substantially outperforms CNN-WARP, CNN-RNN, DNN-BCE, and BP-MLL in both per-class and overall metrics.
• C2AE trains efficiently (10–15 minutes on NUS-WIDE with a Titan X GPU) compared to other DNN approaches.
• C2AE remains robust under missing-label scenarios, outperforming LEML, MLML, and ML-MG across varying missing-rate levels.