INFFEED: Leveraging Influence Functions for Improved Model Performance in Subjective Tasks

Core Concepts

Incorporating influence functions as feedback improves model performance in subjective tasks.

Abstract

Abstract: Influence functions quantify perturbations of individual train instances impacting test predictions. INFFEED uses influence functions to improve model performance and identify data points for annotation. Introduction: Real-world data distribution imbalance affects model performance. Transparency enhances model explainability and trustworthiness. Data Extraction: "INFFEED outperforms state-of-the-art baselines by 4% for hate speech classification, 3.5% for stance classification, and 3% for irony and sarcasm detection." Related Work: Interpretability issues in deep learning models. Influence functions provide insights into training data impact on model predictions. Methodology: INFFEED framework adjusts labels based on influencers to enhance model performance. Evaluation on subjective tasks shows significant F1 score improvements over baselines. Experimental Setup: Three setups with varying dataset sizes demonstrate improved performance with increased data. Results: INFFEED outperforms baselines across multiple datasets, showing statistically significant results.

Stats

INFFEEDはヘイトスピーチ分類で最大マクロF1スコアの改善を示す。

Quotes

Key Insights Distilled From

InfFeed

by Somnath Bane... at arxiv.org 03-12-2024

https://arxiv.org/pdf/2402.14702.pdf

Deeper Inquiries

データセットの拡張における手動アノテーションの削減方法は他の分野でも有効ですか？

この研究で提案された手法は、影響関数を使用してデータポイントのラベルを更新することで、データセットの拡張時に手動アノテーションを削減する効果的な方法であることが示されています。このアプローチは主観的なタスクに特に適しており、他の機械学習タスクでも同様に有効である可能性があります。例えば、感情分析や意図理解などの自然言語処理タスクや画像認識など幅広い領域で利用される可能性があります。