Interpretable Local Concept-based Explanation with Human Feedback to Predict All-cause Mortality
Authors: Radwa EL Shawi, Mouaz H. Al-Mallah
JAIR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we introduce the dataset used in this work and the concepts definition in Sections 4.1, and 4.2, respectively. We define baselines in Section 4.3 to be compared to the proposed approach in Section 4.4. The faithfulness of the proposed approach is evaluated in Section 4.5. We evaluate the trust in the explanations of CLEF in Section 4.6. We show in Section 4.7 the effectiveness of the explanations of CLEF in detecting bias in data. |
| Researcher Affiliation | Collaboration | Radwa Elshawi EMAIL Institute of Computer Science Tartu University, Estonia Mouaz H Al-Mallah EMAIL Houston Methodist De Bakey Heart & Vascular Center Houston, TX, USA |
| Pseudocode | Yes | Algorithm 1: Algorithm for interactively proposing intuitive and interpretable concepts with human feedback |
| Open Source Code | No | The paper mentions using "scikit-learn implementations (Pedregosa, Varoquaux, Gramfort, Michel, Thirion, Grisel, Blondel, Prettenhofer, Weiss, Dubourg, et al., 2011)" for training. However, it does not explicitly state that the authors' own implementation of CLEF is open-source, nor does it provide any links to a code repository. |
| Open Datasets | Yes | 4.1 Henry Ford FIT Dataset The dataset of this study was collected from patients who underwent treadmill stress testing by physician referrals at Henry Ford Affiliated Hospitals in metropolitan Detroit, MI in the United States, FIT Project (Al-Mallah et al., 2014). |
| Dataset Splits | Yes | The dataset used in this work is split 60% for training, 20% for validation and 20% for testing. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for running the experiments. It only mentions training models like random forest and support vector machine. |
| Software Dependencies | No | The paper states: "We train all approaches using the scikit-learn implementations (Pedregosa, Varoquaux, Gramfort, Michel, Thirion, Grisel, Blondel, Prettenhofer, Weiss, Dubourg, et al., 2011)." While scikit-learn is mentioned, a specific version number is not provided, only a citation to its foundational paper. |
| Experiment Setup | Yes | In this work, we use a fixed depth of 4, leaving the exploration of dynamic depth to future work. More specifically, each concept cj is associated with two list of features; the explored list lj consists of features that have been proposed to a clinician to be associated with concept cj and the other list uj consists of the set of features that have not been proposed yet for concept cj. If the clinician accepts the proposed feature-concept association, then the proposed feature is added to the concept definition and thus feature-concept matrix Ai,j = 1; otherwise, the feature-concept matrix remains unchanged. List lj is first initialized with a single feature i, such that Ai,j = 1 for each concept j and uj is initialized with the rest of features that are not included in lj. Algorithm 1 models the human feedback while proposing feature-concept associations by incorporating the clinician s prior acceptance of feature-concept associations to improve future proposals made by the algorithm and refit model f each time g is updated. To do so, we store a set of labels of the proposals that the user has previously accepted or rejected in matrix intuit. This matrix is first initialized so that intuiti,j = 1 and intuiti,j =j = 0 if Ai,j = 1 in the concept definitions initialized by the user. The matrix is then updated such that intuiti ,j = 1 if the user accepts the proposed feature-concept association; otherwise, it remains unchanged. We assume that a single feature can be associated with different concepts. The key challenge is to propose feature-concept associations that are intuitive for the clinician and equally highly faithful to the model being explained. If the proposal is highly For each concept, we make a fixed number of proposals before moving to the next concept. In this work, we use a fixed number of proposals per concept numproposals = 7. |