Context-Aware REpresentation: Jointly Learning Item Features and Selection From Triplets.

Area of Science:

• Machine Learning
• Cognitive Modeling
• Recommendation Systems

Background:

• User feedback in machine learning is often context-dependent, influencing decisions in recommendation and cognitive modeling.
• The "odd-one-out" learning setting highlights how item context affects choices, with participants selecting the most dissimilar item from a group.
• Existing methods struggle to provide interpretable feature representations for both individual items and their context.

Purpose of the Study:

• To develop a model that accurately predicts item selection from triplets.
• To generate interpretable feature representations for individual items and their context.
• To address the challenge of learning from context-dependent user feedback using only triplet data.

Main Methods:

• Introduction of CARE (Context-Aware REpresentations), a specialized neural network architecture.
• Training CARE using only triplet responses (three items) to predict which item will be selected.
• Proving parameter counting generalization bounds in the i.i.d. setting to demonstrate learning efficiency.

Main Results:

• CARE achieves state-of-the-art performance in the item selection task.
• The model successfully generates meaningful, interpretable representations for both individual items and the context (triplet).
• CARE learns effectively despite the combinatorial complexity of triplets and lack of supervised item-level information.

Conclusions:

• CARE provides accurate predictions and interpretable representations in context-dependent machine learning tasks.
• The architecture demonstrates efficient learning from triplet data, overcoming sample sparsity challenges.
• CARE offers a novel approach for understanding item and context relationships in areas like recommendation and cognitive modeling.