[Paper Review] From Common to Special: When Multi-Attribute Learning Meets Personalized Opinions

Zhiyong Yang, Qianqian Xu+2 more|arXiv (Cornell University)|Nov 18, 2017

Text and Document Classification Technologies4 citations

TL;DR

This paper proposes a hierarchical multi-task learning model that unifies user-specific attribute prediction with multi-attribute correlation through a three-component decomposition: common cognition, attribute-specific bias, and user-specific bias. By integrating Lasso and group Lasso penalties for feature selection, the model achieves state-of-the-art performance on both binary and relative attribute learning tasks, outperforming global, user-specific, and multi-task baselines in empirical evaluations.

ABSTRACT

Visual attributes, which refer to human-labeled semantic annotations, have gained increasing popularity in a wide range of real world applications. Generally, the existing attribute learning methods fall into two categories: one focuses on learning user-specific labels separately for different attributes, while the other one focuses on learning crowd-sourced global labels jointly for multiple attributes. However, both categories ignore the joint effect of the two mentioned factors: the personal diversity with respect to the global consensus; and the intrinsic correlation among multiple attributes. To overcome this challenge, we propose a novel model to learn user-specific predictors across multiple attributes. In our proposed model, the diversity of personalized opinions and the intrinsic relationship among multiple attributes are unified in a common-to-special manner. To this end, we adopt a three-component decomposition. Specifically, our model integrates a common cognition factor, an attribute-specific bias factor and a user-specific bias factor. Meanwhile Lasso and group Lasso penalties are adopted to leverage efficient feature selection. Furthermore, theoretical analysis is conducted to show that our proposed method could reach reasonable performance. Eventually, the empirical study carried out in this paper demonstrates the effectiveness of our proposed method.

Motivation & Objective

• To address the limitation of existing methods that treat user-specific attributes and multi-attribute correlation separately.
• To model both personal diversity in attribute perception and intrinsic relationships among multiple attributes simultaneously.
• To develop a unified framework that captures shared knowledge across attributes while adapting to individual annotator preferences.
• To improve generalization and feature selection in personalized attribute learning through structured regularization.
• To validate the effectiveness of the proposed model on both binary and relative attribute learning benchmarks.

Proposed method

• The model decomposes predictor weights into three components: a shared common cognition factor, an attribute-specific bias factor, and a user-specific bias factor.
• A three-component decomposition is used to jointly model global consensus and individual annotator preferences across multiple attributes.
• Lasso and group Lasso penalties are applied to enable efficient, structured feature selection across attributes and users.
• The optimization is performed using an accelerated proximal gradient method to solve the regularized learning problem.
• The model is trained end-to-end on crowd-sourced labels with adaptation to user-specific data via fine-tuning.
• For relative attribute learning, image pair features are processed to fit the ranking-based formulation of the model.

Experimental results

Research questions

• RQ1 How can we jointly model user-specific attribute preferences and intrinsic correlations among multiple attributes in a unified framework?
• RQ2 Can a common-to-special decomposition of model weights effectively capture both shared knowledge and individual differences in attribute perception?
• RQ3 To what extent does incorporating structured regularization (Lasso and group Lasso) improve feature selection and model generalization in personalized attribute learning?
• RQ4 How does the proposed model compare to global, user-specific, and multi-task baselines in terms of accuracy and robustness across different data splits?
• RQ5 Does the theoretical analysis support the convergence and consistency of the proposed method under realistic assumptions?

Key findings

• The proposed model achieves the highest average classification accuracy on both binary and relative attribute learning tasks across all data splits.
• On the binary attribute dataset, the model outperforms the best baseline (rMTFL-U) by up to 4.2% in accuracy when 40% of the data is used for training.
• On the relative attribute dataset, the model achieves a significant improvement in average ranking accuracy over all baselines, particularly under low-data regimes.
• Theoretical analysis shows that the method can estimate true parameters with high probability and achieve good generalization performance.
• Empirical results confirm that the integration of common cognition, attribute-specific, and user-specific factors leads to better performance than isolated modeling of either factor.
• The use of Lasso and group Lasso penalties enables effective feature selection, reducing overfitting and improving model interpretability.