# Efficient Shapley Values Estimation by Amortization for Text Classification

**Authors:** Chenghao Yang, Fan Yin, He He, Kai-Wei Chang, Xiaofei Ma, Bing Xiang  
**arXiv:** [2305.19998](https://arxiv.org/abs/2305.19998)  
**Date:** 2023-05-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.19998v1/

## Key points

- Authors address the challenge of computing Shapley Values for large pretrained models in neural text classification
- Computation of Shapley Values is time-consuming due to the large number of model evaluations required
- Authors propose an amortized model that directly predicts each input feature's Shapley Value without additional model evaluations
- Estimated Shapley Values are sensitive to random seed choices, especially for examples with longer input texts
- Proposed amortized model provides stable estimations as the inference is deterministic
- Experimental results show that the amortized model accurately estimates Shapley Values with a substantial speedup compared to traditional methods (60 times faster process)
- Functionality of the model evaluated by examining quality of explanations in downstream tasks such as feature selection and domain calibration
- Model compared with computationally expensive KernelSHAP (KS) method and demonstrates superior performance
- Overall, paper presents an efficient and effective approach for estimating Shapley Values in neural text classification models

## Summary

In this paper, the authors address the challenge of computing Shapley Values for large pretrained models in neural text classification. They highlight that while Shapley Values are popular for explaining these models, their computation is time-consuming due to the large number of model evaluations required. To overcome this limitation, the authors propose an amortized model that directly predicts each input feature's Shapley Value without additional model evaluations. The authors demonstrate that estimated Shapley Values are sensitive to random seed choices, especially for examples with longer input texts. This lack of stability can only be mitigated by aggregating thousands of model evaluations, which introduces significant computational overheads. In contrast, their proposed amortized model provides stable estimations as the inference is deterministic. Experimental results on two text classification datasets show that the amortized model accurately estimates Shapley Values with a substantial speedup compared to traditional methods. The computation time per instance is reduced from about 3.47 seconds to less than 50 milliseconds, resulting in a 60 times faster process. Furthermore, the authors evaluate the functionality of their model by examining the quality of explanations in downstream tasks such as feature selection and domain calibration. They compare their method with the computationally expensive KernelSHAP (KS) method and demonstrate superior performance. Overall, this paper presents an efficient and effective approach for estimating Shapley Values in neural text classification models. The proposed amortized model significantly reduces computation time while providing stable estimations, making it a valuable tool for interpreting and understanding these models more effectively and efficiently than before.