CHEER: Rich Model Helps Poor Model via Knowledge Infusion

Abstract: There is a growing interest in applying deep learning (DL) to healthcare, driven by the availability of multi-modal data in rich-data environments (e.g., intensive care units). However, in many practical situations we can only access a few noisy modalities extracted from poor-data environments (e.g., at home), which prevents DL models trained on rich data working accurately on those poor data. How can we boost the performance of models learned using poor-data by leveraging models trained using rich-data? To address this question, we develop a knowledge infusion framework named CHEER that can succinctly summarize the rich model into transferable representations, which can help the poor model improve its performance. The infused model is analyzed theoretically and evaluated empirically on several real-world datasets. Our empirical results showed that CHEER outperformed baselines by 5.60% to 46.80% in terms of the macro-F1 score on multiple physiological datasets.

Paper Bibtex