Local Explanations via Necessity and Sufficiency: Unifying Theory and Practice

Watson, David S.; Gultchin, Limor; Taly, Ankur; Floridi, Luciano

doi:10.1007/s11023-022-09598-7

Cited by 21 publications

(13 citation statements)

References 64 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…As observed by Watson, Gultchin et al [57], necessity and sufficiency are the building blocks of any successful explanation. Necessary and sufficient explanations are complementary to each other: given a source entity, a relation and the predicted target entity:…”

Section: Explaining Link Predictionsmentioning

confidence: 87%

“…Recently, LIME seems to have been outclassed by frameworks that formulate the relevance of input features in terms of Shapley values [46]. As observed by Watson, Gultchin et al [57] these methods, such as SHAP [30], have gained popularity due to their solid theoretical backing derived from Game Theory. We provide a detailed comparison between SHAP and our approach in Section 4.3.…”

Section: General Purpose Frameworkmentioning

confidence: 99%

See 1 more Smart Citation

Explaining Link Prediction Systems based on Knowledge Graph Embeddings

Rossi

Firmani

Merialdo

et al. 2022

Proceedings of the 2022 International Conference on Management of Data

View full text Add to dashboard Cite

Link Prediction (LP) aims at tackling Knowledge Graph incompleteness by inferring new, missing facts from the already known ones. The rise of novel Machine Learning techniques has led researchers to develop LP models that represent Knowledge Graph elements as vectors in an embedding space. These models can outperform traditional approaches and they can be employed in multiple downstream tasks; nonetheless, they tend to be opaque, and are mostly regarded as black boxes. Their lack of interpretability limits our understanding of their inner mechanisms, and undermines the trust that users can place in them. In this paper, we propose the novel Kelpie explainability framework. Kelpie can be applied to any embedding-based LP models independently from their architecture, and it explains predictions by identifying the combinations of training facts that have enabled them. Kelpie can extract two complementary types of explanations, that we dub necessary and sufficient. We describe in detail both the structure and the implementation details of Kelpie, and thoroughly analyze its performance through extensive experiments. Our results show that Kelpie significantly outperforms baselines across almost all scenarios.

show abstract

Section: Explaining Link Predictionsmentioning

confidence: 87%

Section: General Purpose Frameworkmentioning

confidence: 99%

Explaining Link Prediction Systems based on Knowledge Graph Embeddings

Rossi

Firmani

Merialdo

et al. 2022

Proceedings of the 2022 International Conference on Management of Data

View full text Add to dashboard Cite

show abstract

“…SHAP is preferred over LIME because: (1) SHAP has been well formulated for tree-based interpretations, (2) improvements in mathematical guarantees over LIME have been well documented, (3) SHAP provides global feature importances, and (4) reports salient interactions. However, this does not necessarily indicate SHAP as the “go to” feature importance technique; discussion of display outputs from other feature importance approaches (e.g., counterfactual explanations; DiCE, Anchor) in collaboration with the clinical domain expert is important when making a final selection ( 58 – 62 ). While the ultimate utility in using SHAP lies in the ability to fit explanatory models for each individual in the case that machine learning approaches dominate, SHAP, in any model application, can generate instance-wise importance values for useful, patient-specific readouts for the clinician.…”

Section: Discussionmentioning

confidence: 99%

Machine Learning Approaches for Hospital Acquired Pressure Injuries: A Retrospective Study of Electronic Medical Records

Levy

Lima

Miller

et al. 2022

Front. Med. Technol.

View full text Add to dashboard Cite

BackgroundMany machine learning heuristics integrate well with Electronic Medical Record (EMR) systems yet often fail to surpass traditional statistical models for biomedical applications.ObjectiveWe sought to compare predictive performances of 12 machine learning and traditional statistical techniques to predict the occurrence of Hospital Acquired Pressure Injuries (HAPI).MethodsEMR information was collected from 57,227 hospitalizations acquired from Dartmouth Hitchcock Medical Center (April 2011 to December 2016). Twelve classification algorithms, chosen based upon classic regression and recent machine learning techniques, were trained to predict HAPI incidence and performance was assessed using the Area Under the Receiver Operating Characteristic Curve (AUC).ResultsLogistic regression achieved a performance (AUC = 0.91 ± 0.034) comparable to the other machine learning approaches. We report discordance between machine learning derived predictors compared to the traditional statistical model. We visually assessed important patient-specific factors through Shapley Additive Explanations.ConclusionsMachine learning models will continue to inform clinical decision-making processes but should be compared to traditional modeling approaches to ensure proper utilization. Disagreements between important predictors found by traditional and machine learning modeling approaches can potentially confuse clinicians and need to be reconciled. These developments represent important steps forward in developing real-time predictive models that can be integrated into EMR systems to reduce unnecessary harm.

show abstract

“…Galhotra et al (2021) suggested an approach to capture the notions of necessity and sufficiency from probabilistic causal models (Pearl, 2009). Watson et al (2021) presented a different method for quantifying necessity and sufficiency over subsets of features. We follow the framework of probabilistic causal models, and adopt the definitions from Galhotra et al (2021).…”

Section: Background and Related Workmentioning

confidence: 99%

Necessity and Sufficiency for Explaining Text Classifiers: A Case Study in Hate Speech Detection

Balkir¹,

Nejadgholi²,

Fraser³

et al. 2022

Preprint

View full text Add to dashboard Cite

We present a novel feature attribution method for explaining text classifiers, and analyze it in the context of hate speech detection. Although feature attribution models usually provide a single importance score for each token, we instead provide two complementary and theoretically-grounded scores -necessity and sufficiency -resulting in more informative explanations. We propose a transparent method that calculates these values by generating explicit perturbations of the input text, allowing the importance scores themselves to be explainable. We employ our method to explain the predictions of different hate speech detection models on the same set of curated examples from a test suite, and show that different values of necessity and sufficiency for identity terms correspond to different kinds of false positive errors, exposing sources of classifier bias against marginalized groups.

show abstract

Local Explanations via Necessity and Sufficiency: Unifying Theory and Practice

Cited by 21 publications

References 64 publications

Explaining Link Prediction Systems based on Knowledge Graph Embeddings

Explaining Link Prediction Systems based on Knowledge Graph Embeddings

Machine Learning Approaches for Hospital Acquired Pressure Injuries: A Retrospective Study of Electronic Medical Records

Necessity and Sufficiency for Explaining Text Classifiers: A Case Study in Hate Speech Detection

Contact Info

Product

Resources

About