Explainable ai Research Papers PDF
Explainable AI (XAI) Project Ideas 2026 — Healthcare, Finance, Autonomous Systems & Legal AIExplainable Artificial Intelligence (XAI) is the discipline of designing or augmenting AI systems so that their internal logic, feature contributions and decision boundaries can be understood by domain experts, affected individuals, regulators and the public. As AI models permeate high-stakes domains — clinical diagnosis, algorithmic trading, autonomous vehicle control and criminal sentencing — the demand for transparent, accountable and legally compliant AI has surged dramatically. The EU AI Act (2024), GDPR Article 22 and the US FDA SaMD guidance all mandate meaningful human-readable explanations for high-risk AI decisions, making XAI no longer optional but regulatory necessity.
At ProjectsatBangalore, we offer 40+ IEEE 2026 Explainable AI project ideas spanning four critical domains with complete Python implementations, IEEE 2026 base papers, quantitative explanation evaluation (faithfulness, stability, comprehensibility) and fairness auditing. Projects are available for BE, MTech and PhD students and include SHAP, LIME, GradCAM, Integrated Gradients, Counterfactual Explanations (DiCE, Alibi), Concept-Based XAI (TCAV) and Causal XAI (DoWhy).
XAI Research Areas We Cover
- Post-hoc local explanations — SHAP, LIME, Anchors
- Gradient-based saliency — GradCAM, ScoreCAM, Integrated Gradients
- Counterfactual explanations — DiCE, Alibi, NICE
- Concept-based XAI — TCAV, Concept Bottleneck Models
- Causal XAI — DoWhy, CausalNex, structural causal models
- Attention map visualisation — Transformer, BERT, ViT
- Explainable NLP — token attribution, rationale extraction
- Fairness auditing — SHAP disparity, demographic parity, Equalized Odds
- Medical imaging XAI — GradCAM on X-ray, CT, MRI, pathology
- Autonomous driving XAI — LiDAR, camera decision explanation
- Intrinsically interpretable models — EBMs, decision rules, monotone networks
- XAI evaluation — faithfulness, stability, AOPC, insertion/deletion curves
Explainable ai Based Projects
Core XAI Explanation Methods — Quick ReferenceSeven major families of XAI techniques used across all 40+ project ideas. Each method has distinct strengths — choose based on model type, data modality, explanation audience and regulatory context.
SHAP
SHapley Additive exPlanations
Game-theory-based feature attribution. Assigns each feature a Shapley value — its marginal contribution averaged over all feature subsets. Consistent, locally accurate, model-agnostic.
✦ Best for: tabular data, tree models, any classifier/regressor requiring feature importance
LIME
Local Interpretable Model-Agnostic Explanations
Perturbs the input around a specific instance, queries the black-box model and fits a sparse linear surrogate to explain that individual prediction. Works on tabular, text and image data.
✦ Best for: explaining any single prediction to a non-technical stakeholder
GradCAM / ScoreCAM
Gradient-weighted Class Activation Maps
Uses gradients of the class score with respect to feature maps in a CNN's last convolutional layer to produce a coarse localisation heatmap highlighting the discriminative image regions used for prediction.
✦ Best for: CNN classification explanations in medical imaging and visual AI
Counterfactual (DiCE / Alibi)
Counterfactual Explanations
Generates the minimum-change input that would alter the model's decision — "your loan would be approved if your income were ₹8,000 higher." Actionable, human-interpretable and legally defensible under GDPR.
✦ Best for: finance, HR and legal AI where actionable recourse is needed
Causal XAI (DoWhy)
Causal Inference & Structural Causal Models
Goes beyond correlation-based attribution to identify causal relationships between variables. Intervention and counterfactual queries via do-calculus. Enables debiasing, fairness and what-if analysis grounded in causality rather than statistics.
✦ Best for: policy AI, medical treatment effect estimation, legal fairness
Attention Maps
Transformer Attention Visualisation
Extracts and visualises self-attention weights from BERT, GPT, ViT and other transformer models to show which input tokens or image patches the model focused on when making a prediction. Used for NLP and vision XAI.
✦ Best for: explainable NLP (text classification, QA) and Vision Transformer image analysis
TCAV
Concept-Based XAI (Testing with CAVs)
Tests how sensitive a model's predictions are to user-defined concepts (e.g., "striped texture", "optic disc", "high-risk neighbourhood") using linear classifiers trained on concept examples. Produces human-interpretable concept importance scores.
✦ Best for: validating medical AI (do doctors' concepts align with model features?)
Integrated Gradients (Captum)
Integrated Gradients & DeepLIFT
Attributes the model output to input features by integrating gradients along a straight-line path from a baseline to the input. Satisfies completeness, sensitivity and implementation invariance axioms. Available in PyTorch via Captum.
✦ Best for: deep neural network explanation where SHAP KernelExplainer is too slow
Explainable ai Based Intrusion Detection System
Tools & Libraries Used in XAI ProjectsComplete Python XAI toolkit used across all 40+ IEEE 2026 Explainable AI project ideas — from model training to explanation generation, evaluation and fairness auditing.
SHAP — TreeExplainer / DeepExplainer
LIME — Tabular / Image / Text
Captum (PyTorch XAI)
pytorch-grad-cam (GradCAM / ScoreCAM)
DiCE — Diverse Counterfactuals
Alibi — Anchors / CF / KernelSHAP
DoWhy — Causal Inference
IBM AI Explainability 360
InterpretML — EBM / SHAP
PyTorch / Lightning
Scikit-learn / XGBoost / LightGBM
Fairlearn / AIF360 Fairness Audit
Plotly / Matplotlib — XAI Visualisation
Domain 1 — Healthcare & Medicine
XAI for clinical decision support, medical imaging, drug discovery, ICU mortality prediction and patient outcome modelling — with GradCAM, SHAP, Concept-Based and Counterfactual methods. Regulatory context: FDA SaMD guidance, EU AI Act Article 13 and clinical trial transparency requirements.
H-01
SHAP-Guided Explainable Sepsis Prediction from ICU EHR Data — 72-Hour Early Warning
XGBoost + SHAP TreeExplainer trained on MIMIC-IV ICU data to predict sepsis onset 72 hours before clinical diagnosis. SHAP waterfall plots explain individual patient predictions; beeswarm plots reveal global feature importance. Includes AUROC, sensitivity/specificity and SHAP-based feature interaction analysis.
H-02
GradCAM + ScoreCAM Explainable Chest X-Ray Diagnosis for Pneumonia, COVID-19 and TB
ResNet-50 / DenseNet-121 trained on CheXpert + NIH ChestX-ray14 with GradCAM, GradCAM++, ScoreCAM and EigenGradCAM heatmap comparison. Radiologist alignment study using Intersection over Union (IoU) of heatmap vs. clinical ground-truth bounding boxes. Faithfulness evaluated via insertion/deletion curves.
H-03
Counterfactual Explanations for Diabetic Readmission Prediction — Actionable Clinical Recourse
LightGBM readmission classifier on UCI Diabetes 130-US dataset with DiCE diverse counterfactual generation. Clinicians receive actionable recourse: "Readmission risk drops below 20% if HbA1c is managed below 7.5% and discharge instructions are improved." Includes proximity, sparsity and plausibility metrics for counterfactual quality.
H-04
Concept-Based XAI (TCAV) for Breast Cancer Histopathology Classification — Validating Pathologist Concepts
EfficientNet-B4 trained on BreakHis + TCGA-BRCA histopathology images. TCAV tests whether pathologist-defined concepts ("nuclear pleomorphism", "mitotic figures", "gland formation") align with model-internal feature representations. CAV sensitivity scores compared across benign, malignant and grade classes.
H-05
Causal XAI with DoWhy for Estimating Drug Treatment Effect in Sepsis — Observational Study Debiasing
Structural Causal Model (SCM) on MIMIC-III ICU data to estimate Average Treatment Effect (ATE) of vasopressors on 28-day mortality, adjusting for confounders (SOFA score, age, comorbidities) using inverse probability weighting and doubly-robust estimation. DoWhy refutation tests validate causal assumptions.
H-06
Attention-Map Explainable Vision Transformer (ViT) for MRI Brain Tumour Grading — Glioblastoma vs. Meningioma
ViT-Base/16 fine-tuned on BraTS 2024 MRI dataset with Attention Rollout and DINO self-supervised attention visualisation. Explainability panels show which MRI slices and spatial regions drove grade-IV vs. grade-II classification. Grad-Rollout outperforms raw attention in clinician-blinded localisation study.
H-07
Integrated Gradients (Captum) for Explainable Drug–Drug Interaction Prediction using Graph Neural Networks
GraphSAGE trained on DrugBank + TWOSIDES DDI dataset. Captum Integrated Gradients attributed to node and edge features to explain why a drug pair is flagged as interacting — highlighting substructures (functional groups, pharmacophores) responsible. SHAP GNN extension compared for node-level attribution consistency.
H-08
SHAP + LIME Dual Explainer Framework for Explainable Mental Health Crisis Detection from Clinical Notes (NLP)
BioBERT fine-tuned on MIMIC-III discharge notes for depression/suicide risk classification. SHAP token-level attribution vs. LIME token perturbation explanations compared on faithfulness (AOPC), stability (std. deviation over 100 runs) and agreement score. Clinician user study measures explanation utility and trust.
Domain 2 — Finance & Banking
XAI for credit scoring, fraud detection, algorithmic trading, risk assessment and regulatory compliance — ensuring GDPR Article 22 "right to explanation" and RBI/SEBI model risk management mandates are met through SHAP, LIME, EBM and Counterfactual explanations.
F-01
SHAP-Based Explainable Credit Scoring with GDPR Right-to-Explanation Compliance — Indian Banking Dataset
XGBoost / LightGBM credit default model on RBI-published Loan Default dataset + CIBIL bureau data. SHAP force plots for individual loan rejection explanation letters (GDPR Art. 22 compliant). SHAP dependence plots reveal non-linear income–debt interactions. Fairness audit on gender and age attributes using Fairlearn.
F-02
LIME + Anchors Explainable Fraud Detection for UPI and Credit Card Transactions — Real-Time Explanation
Imbalanced-class fraud detection using SMOTE + Random Forest / Isolation Forest on PaySim + IEEE-CIS Fraud dataset. LIME explains why a transaction was flagged (e.g., "unusual merchant category + 3 AM transaction + new device"). Anchors generate rule-based explanations. Real-time explanation latency benchmarked (<50 ms target).
F-03
Explainable Boosting Machine (EBM) for Mortgage Default Prediction — Intrinsically Interpretable Model with SHAP
Microsoft InterpretML's Explainable Boosting Machine trained on Fannie Mae mortgage dataset — achieves near-XGBoost AUROC while being glass-box interpretable. Each feature's shape function and interaction terms visualised. SHAP and EBM explanations compared for regulator-facing model documentation. Monotonicity constraints ensure income → risk relationship is economically rational.
F-04
Causal XAI for Loan Default Prediction — Debiasing Protected Attributes with DoWhy + Structural Causal Models
DoWhy SCM on HMDA (Home Mortgage Disclosure Act) dataset to distinguish direct causal effect of income on default from indirect paths through neighbourhood race composition (redlining proxy). Interventional fairness enforced by blocking discriminatory causal paths. Compares with post-processing fairness correction baselines.
F-05
Counterfactual Recourse for Explainable Insurance Premium Pricing — NICE Algorithm for Minimum-Effort Recourse
NICE (Nearest Instance Counterfactual Explanation) on health insurance premium prediction model — generates the minimum lifestyle changes (BMI, smoking, activity level) needed to reduce a customer's premium category. Proximity, plausibility and diversity of counterfactuals evaluated. Regulatory compliance with IRDAI model transparency guidelines.
F-06
SHAP Temporal Analysis for Explainable Algorithmic Trading — Which Market Signals Drive Buy/Sell Decisions?
LSTM + XGBoost ensemble stock price prediction on NSE/BSE data with time-step-level SHAP DeepExplainer values showing which technical indicators (RSI, MACD, volume spike, sentiment score) drove each trading signal. Rolling SHAP window analysis reveals regime changes. Backtested on Nifty50 data 2020–2025.
F-07
Explainable Anti-Money Laundering (AML) Graph Neural Network — SHAP-Based Transaction Network Explanation
GraphSAGE trained on Elliptic Bitcoin dataset for money laundering detection. SHAP GNN explainer highlights which transaction-graph neighbours and edge features (transaction volume, time delta, address reuse) contributed most to the suspicious classification. Attention-based GNN explanation compared as baseline.
F-08
Fairness-Aware XAI for SME Loan Approval — SHAP Disparity Decomposition across Firm Size, Region and Gender
Random Forest + SHAP on RBI MUDRA / SIDBI SME loan dataset. SHAP disparity decomposition separates explained disparity (legitimate financial features) from unexplained disparity (proxy discrimination via firm location, owner gender). AIF360 and Fairlearn post-processing compared. Output: regulator-ready fairness report with per-group SHAP profiles.
Domain 3 — Autonomous Systems & Defense
XAI for autonomous vehicles, UAVs/drones, robot navigation, sensor fusion and defense AI — enabling safety certification (ISO 26262, SOTIF, MIL-STD-882), edge-case failure analysis and human-on-the-loop decision transparency for safety-critical deployments.
A-01
GradCAM Explainability for Autonomous Vehicle Lane Detection CNN — Failure Case Analysis at Adversarial Weather
DeepLab-v3+ lane segmentation model trained on BDD100K + nuScenes. GradCAM heatmaps visualise which road regions drive lane prediction at night, rain and fog. Failure analysis: GradCAM reveals the model attends to guardrails instead of lane markings in heavy rain. Eigen-GradCAM reduces noise; post-hoc segmentation XAI verified against human annotations.
A-02
SHAP-Based Sensor Fusion Explainability for LiDAR-Camera 3D Object Detection — Why Did the AV Miss a Pedestrian?
PointPillars + camera fusion trained on KITTI / Waymo Open Dataset. SHAP KernelExplainer quantifies each sensor stream's (LiDAR intensity, RGB camera, radar) contribution to each 3D bounding box confidence score. Edge case analysis: SHAP reveals camera dominates at low LiDAR point density — motivating adaptive fusion weighting.
A-03
Counterfactual XAI for Autonomous Drone Trajectory Planning — Minimum-Change Obstacle Avoidance Explanations
Deep Reinforcement Learning (DRL) policy for UAV path planning in cluttered 3D environments. DiCE counterfactual generation explains: "The drone chose path A over path B because path B had an obstacle within 1.2 m at 3.4-second lookahead." State-space counterfactuals validated in Gazebo simulation. Human pilot study measures explanation utility for safety auditing.
A-04
Attention-Map XAI for Transformer-Based Pedestrian Intent Prediction — Explaining "Will This Person Cross the Road?"
Pedestrian action transformer (PAT) on JAAD + PIE pedestrian datasets. Multi-head attention rollout visualises which past body poses, gaze direction frames and crossing-context features (signal state, kerb proximity) the model attends to when predicting crossing intent. Critical for AV ethics: the model must not discriminate by pedestrian appearance.
A-05
SHAP + LIME Explainable Anomaly Detection for Industrial Robot Predictive Maintenance — Bearing Fault XAI
Autoencoder + XGBoost anomaly detection on CWRU Bearing + PHM Challenge datasets. SHAP waterfall plots explain which frequency-domain vibration features (RMS, kurtosis, envelope spectrum peak at BPFO) triggered a bearing fault alert. LIME time-series explainer applied to raw waveform for maintenance engineer dashboards.
A-06
Causal XAI for Explainable Target Classification in SAR (Synthetic Aperture Radar) Defense Imagery
CNN trained on MSTAR + OpenSARShip SAR datasets for military vehicle classification. Causal intervention (do-calculus via DoWhy) identifies whether the model classifies based on genuine radar backscatter signatures or on spurious correlations (image capture angle, operating frequency artefacts). GradCAM confirms causal attribution; adversarial patch robustness verified.
A-07
Integrated Gradients Explainability for End-to-End Autonomous Driving Neural Network — DAVE-2 / PilotNet Attribution
Captum Integrated Gradients on PilotNet (NVIDIA DAVE-2 style end-to-end AV network) trained on CARLA simulation + comma2k19 real-world data. Attribution maps show which road scene pixels cause steering corrections — verifying the network attends to lane markings and not sky texture. Sensitivity to adversarial road patches analysed for safety certification.
A-08
Explainable Multi-Agent Reinforcement Learning (MARL) for Swarm Drone Coordination — Role-Based SHAP Attribution
Multi-agent PPO for swarm UAV target search on AirSim. SHAP values computed per agent at each timestep — explaining which neighbour positions, relative velocities and target-detection signals drove each agent's velocity command. Role attribution reveals emergent specialisation (scouts vs. pursuers). Explanation fidelity measured by removing high-SHAP features and observing performance degradation.
Domain 4 — Legal & Criminal Justice
XAI for recidivism prediction, judicial decision support, legal document analysis and criminal sentencing AI — addressing fairness, racial bias, due-process transparency and the constitutional requirement for explainability of government AI decisions under the EU AI Act and US algorithmic accountability frameworks.
L-01
Explainable Recidivism Prediction Beyond COMPAS — SHAP Fairness Audit and Counterfactual Recourse for Defendants
XGBoost recidivism model on ProPublica COMPAS + NIJ Recidivism Challenge datasets. SHAP racial bias audit decomposing prediction disparity into feature-based (criminal history) vs. residual (unexplained racial gap) components. DiCE counterfactuals provide defendants: "Your recidivism risk score drops below medium if employment history spans 18+ months." SHAP vs. COMPAS risk scores compared with lawyers' assessments.
L-02
Concept-Based XAI (TCAV) for Legal Judgement Prediction — Validating Judges' Legal Concepts in AI Court Models
BERT / Legal-BERT fine-tuned on ECHR (European Court of Human Rights) + Indian Supreme Court datasets for judgement outcome prediction. TCAV tests whether legal concepts ("proportionality", "state interference", "procedural fairness") are captured in model activations — or whether the model relies on spurious surface patterns (defendant nationality, case length).
L-03
Causal Fairness Auditing of Criminal Sentencing AI — Separating Legitimate Predictors from Racial Proxy Variables
DoWhy structural causal model on US Sentencing Commission dataset (n=500,000 federal sentences). Intervention on race variable via do-calculus quantifies direct racial discrimination in sentencing AI vs. indirect paths through neighbourhood poverty index (redlining proxy). Counterfactual fairness implemented to guarantee predictions invariant to race. Compared with demographic parity and Equalized Odds post-processing.
L-04
LIME + Attention Explainable Contract Clause Risk Classification — Legal AI for Indian Contract Act Compliance
Legal-BERT fine-tuned on CUAD (Contract Understanding Atticus Dataset) for risky clause detection (indemnification, limitation of liability, governing law). LIME token perturbation explanations identify which specific words/phrases triggered the risk flag. Attention head visualisation reveals cross-clause dependencies. Lawyer user study validates explanation quality and reduces review time by 40%.
L-05
SHAP-Based Explainable Bail Decision Support System — Transparency and Consistency Audit for Pre-Trial Detention AI
XGBoost bail risk model on Arnold Foundation Public Safety Assessment data. SHAP dependence plots reveal race-correlated proxy features (zip code arrest history, prior failure-to-appear — both correlated with poverty). AIF360 Equalized Odds post-processing applied. Output: per-judge consistency dashboard showing SHAP profiles across demographically matched defendants for judicial quality assurance.
L-06
Counterfactual Explanations for Explainable Parole Decision AI — Minimum Requirements for Parole Eligibility
Random Forest parole recommendation model on US Bureau of Prisons data. DiCE generates diverse counterfactuals for each denied prisoner: "Parole recommended if: (1) no disciplinary incidents for 18 months AND (2) vocational programme completion." Actionability and legal validity of counterfactuals reviewed by criminal justice experts. Fairness: counterfactual difficulty analysed by race and age groups.
L-07
Attention + SHAP Explainable Hate Speech and Extremist Content Detection — Transparency for Content Moderation AI
RoBERTa-large fine-tuned on HateXplain + Measuring Hate Speech datasets. SHAP token-level attribution identifies which words drove the hate speech classification — enabling appeals process: "This post was flagged due to terms X and Y in context Z." Attention heads visualised for cross-token dependency analysis. Fairness audit: detection rates equalised across racial and religious groups using Fairlearn.
L-08
Integrated Gradients + Anchors Explainable Fake Evidence Detection in Court Documents — Forensic Document AI
BERT + ResNet multimodal model classifying forged vs. authentic court documents (text tampering + image watermark manipulation). Captum Integrated Gradients highlights altered text regions; Alibi Anchors generate IF-THEN rules: "IF date field font differs from body text font AND metadata timestamp inconsistent THEN document flagged as forged." Evaluated on curated legal document forgery dataset.
All 32 XAI Project Ideas — Quick Reference Table
All topics with XAI method, domain, tools and level at a glance. Call 9591912372 to confirm topic availability, get the full research specification and start within 48 hours.
| # | Project Title (Short) | Domain | XAI Method | Tools | Level |
|---|---|---|---|---|---|
| H-01 | Explainable Sepsis Prediction — 72-Hour ICU Early Warning | Healthcare | SHAP | XGBoost, SHAP, MIMIC-IV, Pandas | MTech |
| H-02 | GradCAM Chest X-Ray Diagnosis Explanation | Healthcare | GradCAM | PyTorch, pytorch-grad-cam, CheXpert, ResNet-50 | MTech |
| H-03 | Counterfactual Explanations for Diabetic Readmission | Healthcare | DiCE / CF | DiCE, LightGBM, UCI Diabetes, Python | BE/BTech |
| H-04 | TCAV Concept XAI for Breast Cancer Histopathology | Healthcare | TCAV | PyTorch, TCAV, BreakHis, EfficientNet-B4 | PhD |
| H-05 | Causal XAI — Drug Treatment Effect in Sepsis (DoWhy) | Healthcare | Causal | DoWhy, MIMIC-III, Python, CausalNex | PhD |
| H-06 | Attention XAI for ViT Brain Tumour MRI Grading | Healthcare | Attention | ViT, Captum, BraTS 2024, PyTorch | PhD |
| H-07 | Integrated Gradients for Drug–Drug Interaction GNN | Healthcare | Integrated Grad. | Captum, PyG, DrugBank, GraphSAGE | MTech |
| H-08 | SHAP + LIME — Mental Health Crisis Detection NLP | Healthcare | SHAP+LIME | SHAP, LIME, BioBERT, MIMIC-III NLP | MTech |
| F-01 | SHAP Explainable Credit Scoring — GDPR Compliant | Finance | SHAP | XGBoost, SHAP, Fairlearn, Python | MTech |
| F-02 | LIME + Anchors Explainable UPI Fraud Detection | Finance | LIME / Anchors | LIME, Alibi, Random Forest, PaySim | MTech |
| F-03 | Explainable Boosting Machine for Mortgage Default | Finance | EBM / SHAP | InterpretML, EBM, SHAP, Fannie Mae data | MTech |
| F-04 | Causal XAI for Loan Default — Debiasing via DoWhy | Finance | Causal | DoWhy, HMDA Dataset, Python, AIF360 | PhD |
| F-05 | Counterfactual Recourse for Insurance Premium (NICE) | Finance | NICE / CF | Alibi NICE, Scikit-learn, Python | BE/BTech |
| F-06 | SHAP Temporal Analysis for Algorithmic Trading (NSE) | Finance | SHAP DeepExplainer | SHAP, LSTM, XGBoost, NSE/BSE data | MTech |
| F-07 | Explainable AML GNN — Transaction Network SHAP | Finance | SHAP + GNN | PyG, GraphSAGE, SHAP, Elliptic dataset | PhD |
| F-08 | Fairness-Aware XAI for SME Loan Approval — India | Finance | SHAP Disparity | SHAP, Fairlearn, AIF360, Random Forest | PhD |
| A-01 | GradCAM AV Lane Detection — Adverse Weather XAI | Autonomous | GradCAM | pytorch-grad-cam, DeepLab-v3+, BDD100K | MTech |
| A-02 | SHAP LiDAR-Camera Sensor Fusion Explainability | Autonomous | SHAP | SHAP, PointPillars, KITTI, PyTorch | PhD |
| A-03 | DiCE Counterfactual XAI for Drone Trajectory DRL | Autonomous | DiCE / CF | DiCE, Stable-Baselines3, Gazebo, Python | PhD |
| A-04 | Attention XAI for Pedestrian Intent Prediction | Autonomous | Attention Rollout | Attention Rollout, PAT, JAAD, PyTorch | MTech |
| A-05 | SHAP + LIME — Predictive Maintenance Bearing Fault | Autonomous | SHAP + LIME | SHAP, LIME, XGBoost, CWRU dataset | MTech |
| A-06 | Causal XAI for SAR Defense Target Classification | Autonomous | Causal + GradCAM | DoWhy, GradCAM, PyTorch, MSTAR | PhD |
| A-07 | Integrated Gradients — DAVE-2 End-to-End AV XAI | Autonomous | Integrated Grad. | Captum, PilotNet, CARLA, PyTorch | MTech |
| A-08 | SHAP MARL for Swarm UAV Coordination XAI | Autonomous | SHAP MARL | SHAP, PPO, AirSim, Stable-Baselines3 | PhD |
| L-01 | Explainable Recidivism — Beyond COMPAS with SHAP | Legal / CJ | SHAP + DiCE | SHAP, DiCE, XGBoost, COMPAS / NIJ data | PhD |
| L-02 | TCAV Concept XAI for Legal Judgement Prediction | Legal / CJ | TCAV | TCAV, Legal-BERT, ECHR, PyTorch | PhD |
| L-03 | Causal Fairness Audit for Criminal Sentencing AI | Legal / CJ | Causal Fairness | DoWhy, AIF360, Python, US Sentencing data | PhD |
| L-04 | LIME + Attention — Contract Clause Risk Classification | Legal / CJ | LIME + Attention | LIME, Legal-BERT, CUAD, HuggingFace | MTech |
| L-05 | SHAP Explainable Bail Decision Audit System | Legal / CJ | SHAP | SHAP, AIF360, Fairlearn, XGBoost | PhD |
| L-06 | DiCE Counterfactuals for Parole Decision AI | Legal / CJ | DiCE / CF | DiCE, Random Forest, Python, Fairlearn | PhD |
| L-07 | Attention + SHAP — Explainable Hate Speech Detection | Legal / CJ | SHAP + Attention | SHAP, RoBERTa, HateXplain, Fairlearn | MTech |
| L-08 | Integrated Gradients + Anchors — Forensic Document AI | Legal / CJ | Integrated Grad. | Captum, Alibi, BERT, ResNet, Python | MTech |
ℹ️ 10+ additional XAI topics available — including XAI for NLP summarisation, recommendation systems, energy grid forecasting and cybersecurity intrusion detection. WhatsApp +91 9591912372 with your domain preference, model type and submission deadline.
Need a complete IEEE 2026 Explainable AI project — implemented and ready to submit?
Share your preferred domain (Healthcare / Finance / Autonomous / Legal), XAI method (SHAP / LIME / GradCAM / Counterfactual / Causal), model type (tabular ML, CNN, Transformer, GNN, RL) and submission deadline — we'll confirm the project specification, IEEE 2026 base paper and start within 48 hours. Full package: Python code, trained model, XAI explanation visualisations, fairness audit, IEEE paper, project report, PPT and viva Q&A support.
Explainable AI Project Lab — Gallery
XAI project implementations — SHAP beeswarm plots, GradCAM heatmaps, DiCE counterfactual dashboards, causal graphs, fairness audit reports and IEEE paper drafting for Healthcare, Finance, Autonomous and Legal AI domains.
SHAP Feature Importance — Healthcare
GradCAM Heatmap — Medical Imaging
DiCE Counterfactuals — Finance
Causal Graph — Legal Fairness Audit
LIME Token Attribution — NLP XAI
Attention Rollout — ViT XAI
Fairness Audit — AIF360 / Fairlearn
IEEE 2026 XAI Paper DraftingFAQ — Explainable AI Project Ideas
What is the difference between post-hoc and ante-hoc (intrinsic) XAI — and which should I choose for my project?
Post-hoc XAI applies an explanation method to an already-trained black-box model — SHAP, LIME, GradCAM and Counterfactual Explanations are all post-hoc. The model itself is unchanged; the explainer reverse-engineers its behaviour. Post-hoc methods work with any existing model (XGBoost, CNN, BERT) and are therefore practical for industry settings where the model is fixed. They are currently the dominant approach in IEEE 2026 research — SHAP TreeExplainer is the most commonly used in tabular ML papers, GradCAM in medical imaging. Ante-hoc (intrinsic) XAI builds interpretability directly into the model architecture — Explainable Boosting Machines (EBMs), Decision Trees, Concept Bottleneck Models, Monotone Neural Networks and Linear Models are all intrinsically interpretable. They sacrifice some predictive accuracy for transparency. Ante-hoc models are preferred by regulators (RBI, FDA, EBA) because they are glass-box — no explanation fidelity gap. For MTech projects, post-hoc XAI on a state-of-the-art model (XGBoost + SHAP, CNN + GradCAM) typically produces the best AUROC and the richest IEEE paper; for PhD research, comparing post-hoc vs. ante-hoc explanation quality with faithfulness metrics (AOPC, insertion/deletion) provides a strong novel contribution.
How is explanation faithfulness evaluated in an XAI project?
Explanation faithfulness measures how accurately an explanation reflects the model's actual reasoning — not just what sounds plausible to a human. Key evaluation metrics: AOPC (Area Over the Perturbation Curve) — iteratively removes the most important features (per SHAP/LIME) and measures the resulting model performance drop; a faithful explanation causes larger drops. Insertion and Deletion Curves (Samek et al.) — start with a fully masked input and insert features in order of importance; AUC of the resulting accuracy curve measures faithfulness. Stability / Consistency — run the explainer 100 times with different random seeds and measure the standard deviation of feature attribution vectors; LIME is notably unstable (high std), SHAP TreeExplainer is deterministic. Sanity Checks (Adebayo et al.) — randomise model weights and verify that explanations change (if they don't, the explainer is not actually explaining the model). Human-Grounded Evaluation — domain expert (clinician, financial analyst, lawyer) judges whether the top-5 SHAP features match their professional understanding of the decision. All these evaluation methods are implemented in our XAI project packages to ensure IEEE-publishable explanation quality assessment.
Which XAI method is most legally defensible under GDPR Article 22 and the EU AI Act for Finance and Legal AI projects?
Under GDPR Article 22 ("right to explanation" for automated decisions), the European Data Protection Board (EDPB) guidelines favour Counterfactual Explanations as most legally defensible — they directly answer "what would need to change for the decision to be different?" (e.g., "your loan would be approved if income exceeded ₹8 lakh") and enable actionable recourse, which is a key GDPR requirement. SHAP explanations are accepted by regulators as model risk management evidence (under EBA/RBI model risk guidelines) but require careful communication — raw SHAP waterfall plots are not intelligible to a layperson, so they must be translated into plain-English explanation letters. The EU AI Act (2024) Article 13 (Transparency) and Annex IV (technical documentation) require high-risk AI providers to document feature importance — SHAP global importance plots satisfy this. Intrinsically interpretable models (EBMs, monotone networks) are the gold standard for regulatory compliance as they require no post-hoc explanation gap. For PhD research in legal AI or finance XAI, we recommend implementing SHAP + DiCE together — SHAP for global model audit and DiCE counterfactuals for individual decision explanation, covering both regulatory documentation and consumer-facing recourse requirements simultaneously.
Can SHAP be applied to deep learning / neural networks — or is it only for tree models?
SHAP has multiple backend explainers for different model architectures: TreeExplainer — exact, fast SHAP values for tree ensembles (XGBoost, LightGBM, Random Forest, CatBoost); runs in O(TLD²) — suitable for models with millions of parameters. DeepExplainer — approximate SHAP for PyTorch and TensorFlow deep neural networks using a variant of DeepLIFT; much faster than KernelExplainer for large networks. GradientExplainer — uses expected gradients (a SHAP-consistent extension of Integrated Gradients) for neural networks; supports convolutional feature maps. KernelExplainer — model-agnostic, works on any black-box including LLMs, RL policies and ensemble stacks; computationally expensive (O(2^n) exact or Monte Carlo approximation) but universally applicable. LinearExplainer — exact SHAP for linear models; analytically computed from model coefficients. For medical imaging CNNs, GradientExplainer or GradCAM-SHAP (pixel-level SHAP using gradient-weighting on conv feature maps) is recommended. For BERT/transformer NLP, SHAP KernelExplainer on sentence embeddings or token-level Integrated Gradients via Captum are preferred — raw attention weights are not SHAP-consistent and can be misleading as standalone explanations.
What datasets are used in Healthcare, Finance, Autonomous and Legal XAI projects?
Healthcare: MIMIC-III / MIMIC-IV (ICU EHR, 40,000+ patients, PhysioNet), CheXpert (224,000 chest X-rays, Stanford), NIH ChestX-ray14, BreakHis (breast histopathology), BraTS 2024 (brain tumour MRI), UCI Diabetes 130-US (readmission), DrugBank + TWOSIDES (drug interaction). Finance: UCI Credit Card Default, Kaggle IEEE-CIS Fraud Detection, Fannie Mae Mortgage Performance, HMDA (Home Mortgage Disclosure Act), Elliptic Bitcoin Transactions (AML), PaySim synthetic mobile money, NSE / BSE historical OHLCV data (via yfinance). Autonomous Systems: KITTI Vision Benchmark (3D detection), BDD100K (lane detection), nuScenes (sensor fusion), JAAD + PIE (pedestrian intent), CWRU Bearing Dataset (predictive maintenance), MSTAR (SAR defense), comma2k19 (end-to-end driving), Waymo Open Dataset. Legal / Criminal Justice: ProPublica COMPAS (recidivism, n=7,200), NIJ Recidivism Challenge (n=26,000), ECHR (European Court of Human Rights judgements), CUAD (500 contracts, 41 clause types), US Sentencing Commission data, Arnold Foundation PSA (bail), HateXplain (hate speech with human rationales). All datasets are publicly available, freely downloadable and used within academic fair-use / research-only terms.