Proactive Transfer Learning for Heterogeneous Feature and Label ...

Proactive Transfer Learning for Heterogeneous Feature and Label Spaces Seungwhan Shane Moon Jaime Carbonell Email: [email protected]

Human Learning Is From Heterogeneous Knowledge Sources

Knowledge sources: text, visuals, communications, thought process, …, etc.

Example Experimental Setup Hetero-lingual Transfer Learning Dataset

Source

Target

R8

Multi-lingual Reuters

Heterogeneous Feature Space Language

English

Italian

Heterogeneous Label Space Class Names

’money’, ‘interest’, ‘trade’, ‘ship’, ‘crude’, ‘grain’, …

'corporate', 'government', 'finance’, 'economics', 'equity’, …

A Novel Task (Text Classification) Label Assumption

Fully Labeled (100%)

Partially Labeled (