Replaces a degree-optimized C++ B+Tree with a two-stage Recursive Model Index backed by an LSM-style delta buffer, shifting lookups from memory-bound pointer-chasing to cache-friendly arithmetic.
- 10M-row lognormal benchmark: 334.71 MB B+Tree footprint reduced to 117.97 KB — a 99.96% reduction.
- Read throughput scales from 4.74 MQPS to 14.91 MQPS, a 3.14× improvement, by fitting the model within L2 cache.
- Real-time inserts sustained at 0.0230 μs/write via the delta buffer, with read latency degrading predictably as it fills.
- GPU batch sweep (MPS) identifies a throughput sweet spot at batch = 512 (2.72 MQPS).
Systems intersect
Investigates how shifting workloads from memory-bound pointer chasing to compute-bound inference changes CPU cache hit rates and instruction throughput.
CH.02
Drift-Lab: Proactive Drift Detection & Conformal Mitigation
drift-lab
A statistical hypothesis cascade (KS, MMD, Sinkhorn optimal transport) classifies distribution shift into a formal taxonomy, then routes each type to a cost-minimal, decision-theoretic remedy instead of blanket retraining.
- Classifies drift as covariate, concept, prior-probability, or full dataset shift, each mapped to a distinct remedy — re-weighting, active learning, threshold recalibration, or retraining.
- Non-exchangeable weighted conformal prediction holds coverage near the 90% nominal target (α = 0.10) as standard conformal bounds fall below 70% under shift.
- Synthetic drift injected on the UCI Credit Card dataset across baselines: Deep MLP, TabNet, and XGBoost.
- Targeted remediation lowers maintenance compute by 40–60% versus global retraining.
Research focus
Aims to identify the exact typology and statistical threshold of a drift event, so pipelines trigger proactive, targeted intervention rather than reactive full retraining.
Benchmarks GNNExplainer, Subgraph MCTS, and counterfactual explanations on a GCN under adversarial edge perturbations, and introduces an explanation-guided defense and a homophily-based attribution metric.
- On Cora, Subgraph MCTS holds 0.68 Jaccard similarity at 10% edge perturbation versus 0.31 for GNNExplainer — connected-subgraph search is structurally more stable than gradient-based masking.
- Proposes an XAI Sparsifier: pruning low-attribution edges to filter adversarial noise and restore classification accuracy.
- Introduces the Homophily Attribution Index to quantify how far an explanation drifts toward cross-class edges under attack.
- Fidelity-plus/minus evaluated across all three explanation paradigms on a two-layer GCN (81.30% test accuracy).
Research focus
Investigates the vulnerability of XAI attribution logic under subtle, adversarial changes to graph topology.
