In recent years, Explainable Artificial Intelligence (XAI) has attracted significant attention due to the growing complexity and opacity of ML models. While traditional XAI tools have focused on feature interaction analysis, there is a gap in understanding data point interactions and their impact on model performance. This research addresses this gap by studying data point interactions in ML models, specifically KNN, and proposing novel algorithms to enhance prediction performance.
Existing literature in XAI has extensively explored feature interactions, providing insights into how
diferent features contribute to model predictions. Previous works, such as Ribeiro et al.’s LIME [
Tynes et al. introduced pairwise diference regressor [
We question the maturity and usability of explanation methods for the data science community. The abundance of xAI algorithms can be overwhelming, making it hard for practitioners to select the right one for their needs. The difering requirements and implementations of xAI algorithms pose challenges for data scientists in accurately evaluating them and staying current with their development. This
CEUR
ceur-ws.org
issue manifests as the illusion of explanatory depth [
Existing literature primarily focuses on feature interactions, lacking methodologies for evaluating data point interactions. How do data points interact in forming patterns? How can we measure this interaction? And how can we leverage this explanation to improve the ML pipeline?
Exploring whether understanding data point interactions can lead to improved model accuracy and robustness. How can we design and implement algorithms that utilize data point interactions for prediction tasks? Developing and evaluating algorithms that incorporate data point interactions in their predictive mechanisms.
3.0.1. RQ.1 Given the unsolved burden of evaluating and correctly choosing xAI algorithms, we propose ComparexAI that mitigates two issues: non-unified benchmark for xAI algorithms and the illusion of explanatory depth during the interpretation of results. Compare-xAI emerges as a unique and valuable benchmark. Its distinct contributions lie in its simplicity, scalability, ability to integrate any dataset and ML model, and, most importantly, its focus on the user’s expected explanation. By addressing the pitfalls highlighted in surveys of xAI algorithms through concrete functional tests, Compare-xAI provides a robust evaluation framework. 3.0.2. RQ.2 3.0.3. RQ.3 We propose, STI-KNN, the first algorithm that calculates the exact pair-interaction Shapley values in ( 2) rather than (2 ). STI-KNN is the first algorithm that allows studying the exact interaction on large real-world datasets. This research is the first to consider two disjoint fields: Data valuation and Interaction in Explainable AI. Finally, we study various cases of positive and negative data interactions using STI-KNN.
Leveraging the concept of data point interactions, we introduce the Pairwise Diference Learning (PDL) Classifier. This classifier employs a dual representation of the ML task, achieving better prediction performance by integrating pair interaction data, see Figure 1. The empirical evaluation contains 99 diverse datasets, times 25 CV repetitions. We use the macro F1 metric.
4.0.1. RQ.1 With 15 post-hoc xAI algorithms, 25 tests, and 50 research papers indexed, Compare-xAI ofers a unified benchmark that accurately reproduces experiments. Through a rigorous selection protocol, it highlights the contrast between theoretical foundations and practical implementations, making the limitations of each method transparent. Compare-xAI uses an intuitive scoring method to absorb the vast quantity of xAI-related papers and reduce human errors in interpreting xAI outputs. Its goal is to unify post-hoc xAI evaluation methods into a multi-dimensional benchmark, providing insights into the strengths and weaknesses of diferent approaches. Link: https://karim-53.github.io/cxai/ 4.0.2. RQ.2 Thanks to the STI-KNN algorithm, the data interaction can quickly be visualized using a heatmap of the Shapley interaction values. The matrix shows an example of interaction. We observe, first, a contrast between in-class and out-of-class interactions, second, a reduction in interaction due to data redundancy, and third, an unusual pattern when data contains outliers. 4.0.3. RQ.3 Our benchmark demonstrates that PDL consistently outperforms state-of-the-art ML models, resulting in improved F1 scores in a majority of cases. This highlights PDL’s efectiveness in enhancing performance over baseline methods, facilitated through its straightforward integration via our Python package. Link: https://github.com/Karim-53/pdll
Our research indicates that data point interactions play a crucial role in the performance of ML models.
By shifting the focus from feature interactions to data interactions, we have opened up new avenues
for enhancing model interpretability and accuracy. For more detailed results, refer to the following
papers[
Confirming the eficiency of the PDL algorithm by studying its calibration and uncertainty estimation. By continuing to explore the interactions between data points, we hope to contribute significantly to the field of xAI and ML, ultimately leading to more transparent, accurate, and robust models. of explanatory depth in explainable ai, in: 26th International Conference on Intelligent User Interfaces, 2021, pp. 307–317. [16] H. Kaur, H. Nori, S. Jenkins, R. Caruana, H. Wallach, J. Wortman Vaughan, Interpreting interpretability: understanding data scientists’ use of interpretability tools for machine learning, in: Proceedings of the 2020 CHI conference on human factors in computing systems, 2020, pp. 1–14.