Data Mining Research Paper Topics

We see data mining powering recommendations, catching fraud, and speeding up scientific discovery, and we know it offers students plenty of doable, high‑impact paper ideas. We work as the TopicSuggestions team and, as practicing academic researchers, we treat data mining as the craft of extracting reliable patterns from large, messy data using statistics, machine learning, and domain knowledge while respecting constraints like privacy and fairness.

Today we will share a concise set of original research paper topics that you can scope for a semester and justify with current literature and accessible datasets.

Research Paper Topic Ideas on Data Mining

We will group the list into foundations and evaluation, methods and optimization, applications in health/finance/social media/education/sustainability, and cross‑cutting concerns like data quality, bias, privacy, interpretability, and scalability. We will also note a suggested angle, candidate methods, and sample data sources for each topic so you can move from idea to outline fast.

1. Case Study: Orchestrating a pop-up 5G/private Wi‑Fi/LoRaWAN network for a nomadic desert city

How can we choreograph cross-RAT self-healing under dust, heat, and power scarcity? Can we enforce airtime fairness across art installations with heterogeneous burstiness? How do we allocate ephemeral spectrum slices to volunteer microcells without stable backhaul? To what extent can we trade energy for latency via coordinated sleep/wake without breaking real-time safety beacons?

2. Case Study: Carbon-aware campus routing co-optimized with building microgrid states

Can we route flows through switches and APs powered by surplus solar to minimize grams CO2 per bit? How do we co-schedule traffic shifting with HVAC and battery dispatch so we don’t worsen peak demand? Can we prove stability when we embed carbon signals into BGP/OSPF metrics? What do we gain if we prioritize “green” edges for non-latency-sensitive research data movement?

3. Case Study: Delay-tolerant bus-to-bus mesh for homework sync in snowbound rural districts

How can we exploit predictable bus encounters to guarantee homework delivery by day’s end? Can we design social-aware forwarding that uses drivers’ routines as network anchors? What reliability do we achieve if we cache on heater-powered routers during engine-off dwell? How do we bound privacy risks when we carry student content across peer vehicles?

4. Case Study: Privacy-preserving Wi‑Fi sensing coexisting with WPA3 traffic in eldercare residences

Can we schedule CSI collection so we detect falls while we respect resident data privacy and throughput? How do we obfuscate motion signatures at the edge without degrading detection sensitivity? What interference patterns do we induce if we multiplex sensing with video calls at peak times? Can we formalize consent when we infer presence from neighboring apartments’ reflections?

5. Case Study: Reef-safe underwater acoustic sensing with tide-powered surface gateways

How can we adapt MAC timing to tidal energy availability without desynchronizing nodes? Can we shape packet trains to avoid startling marine life while maintaining detection accuracy? What do we learn if we migrate scarce compute between gateways using wave-driven duty cycles? How do we validate reliability when we can only retrieve logs during calm sea windows?

6. Case Study: Drone-swarm mmWave relays that trail marathon packs to offload uplink video

Can we maintain beam alignment while we orbit runner clusters at variable paces? How do we hand over flows between drones without oscillations as packs split and merge? What energy–capacity frontier do we reach if we co-locate compute for in-flight transcoding? Can we guarantee safety margins when we share spectrum with public safety channels on race day?

7. Case Study: Backscatter tag cooperatives as “parasite caches” in big-box retail aisles

Can we coordinate passive tags to cache hot content and reduce AP airtime under shopper density spikes? How do we incentivize batteryless tags to participate when we only pay them with harvested RF? What consistency can we achieve if we let tags gossip via ambient TV carriers? Can we bound inventory privacy leakage when tags overhear each other’s requests?

8. Case Study: Federated learning over metro transit infotainment networks without leaving the bus

How can we schedule on-bus model updates over dead zones using inter-bus exchanges at depots? Can we prevent gradient leakage when we piggyback on captive portals for aggregation? What latency–accuracy trade-offs do we observe if we adapt rounds to route timetables? How do we throttle training to respect vehicle power budgets and rider QoE?

9. Case Study: Community mesh as a neutral measurement fabric for citywide QoS and throttling audits

Can we crowd-orchestrate synchronized probes to detect time-of-day throttling without ISP cooperation? How do we calibrate heterogeneous CPE hardware so we compare apples to apples? What legal and ethical safeguards do we need if we publish per-street performance heatmaps? Can we prove attribution when we separate access bottlenecks from peering congestion?

10. Case Study: LEO satellite fallback with predictive edge prefetching for artisanal fishing fleets

How can we pre-stage weather charts and market bulletins during coastal Wi‑Fi windows to minimize LEO spend offshore? Can we exploit catch patterns to predict which content we cache on each vessel? What resilience do we gain if we let boats form opportunistic raft meshes at night? How do we secure updates when crews share tablets across shifts with intermittent identity checks?

11. Temporal Causality-Aware Clustering for Sensor Drift Correction

— We ask: how can we cluster time series segments by underlying causal regimes to correct for sensor drift without supervision?; We ask: how can causal change-points be disentangled from noise and seasonal effects in streaming sensors? — We will build a pipeline that (1) extracts short-window causal signatures using Granger/transfer entropy variants, (2) clusters signatures with a causality-distance metric, (3) learns per-cluster drift-correction transforms, and (4) validates on synthetic drifted sensors and two real-world multi-sensor deployments.

12. Privacy-Preserving Outlier Explanation Using Synthetic Counterfactuals

— We ask: how can we generate human-interpretable explanations for outliers while ensuring differential privacy for sensitive features?; We ask: can synthetic counterfactuals provide faithful explanations under privacy constraints? — We will design a mechanism that (1) trains a private generative model (e.g., DP-VAE) to synthesize local counterfactuals, (2) derives sparse explanation rules from synthetic neighborhoods, (3) quantifies fidelity vs. privacy trade-offs, and (4) evaluate on healthcare and finance datasets with real privacy budgets.

13. Cross-Modality Concept Drift Detection in Multimodal Streams

— We ask: how can we detect concept drift when different modalities drift asynchronously (e.g., audio drifts before video)?; We ask: how can cross-modal alignment signals improve early detection and attribution of drift? — We will create a detector that (1) learns modality-specific feature monitors, (2) models cross-modal lead-lag relationships via transfer entropy/time-lagged CCA, (3) raises drift alerts when predictive consistency degrades, and (4) tests on multimodal social-media and autonomous vehicle streams.

14. Fairness-Aware Pattern Mining for Underreported Populations

— We ask: how can we mine association patterns that reliably represent underreported groups without amplifying sampling bias?; We ask: what constraints or reweighting schemes preserve pattern utility while improving representation? — We will develop a constrained frequent-pattern mining method that (1) integrates adaptive sampling weights based on estimated underreporting, (2) enforces fairness constraints on support/lift thresholds, and (3) evaluates explanatory power and fairness on crime, health, and census-derived datasets.

15. Energy-Efficient On-Device Frequent Pattern Mining

— We ask: how can devices mine frequent itemsets under strict CPU/RAM/energy budgets with provable approximation guarantees?; We ask: can we trade computation for communication by opportunistic cooperative mining across devices? — We will propose lightweight sketch-based and substream-summarization algorithms with (1) energy-aware sampling, (2) opportunistic peer exchange protocols to merge summaries, and (3) bounds on approximation vs. energy cost; we will implement on mobile/edge hardware and measure battery impact and accuracy.

16. Graph Neural Network Explainability via Subgraph Frequency Mining

— We ask: how can frequent discriminative subgraphs be mined to explain GNN predictions across instances?; We ask: can we integrate mined subgraphs into GNN training to improve both interpretability and robustness? — We will mine discriminative subgraphs using pattern growth constrained by GNN attention maps, (1) score subgraphs by class-discriminative frequency, (2) validate explanations by ablation and counterfactual insertion, and (3) optionally regularize GNNs to favor interpretable substructures, evaluated on molecular and social graphs.

17. Adaptive Sampling for Mining Rare Event Precursors in IoT

— We ask: how can we adaptively sample high-rate IoT streams to maximize discovery of precursors to rare critical events under limited storage?; We ask: how can learned budget-allocation policies generalize across device types? — We will frame sampling as a reinforcement-learning bandit that (1) uses lightweight precursor detectors to allocate sampling budget, (2) incorporates long-term reward for capturing rare-event windows, and (3) evaluates on industrial sensor logs and simulated rare-failure scenarios.

18. Mining Socio-Environmental Event Cascades from Sparse Mobile Data

— We ask: how can we infer cascading socio-environmental events (e.g., protests following weather shocks) when mobile-device signals are spatially sparse and biased?; We ask: how can causal cascade models be robust to sampling heterogeneity? — We will combine spatial point-process models with user-representative reweighting, (1) infer cascade trigger and propagation kernels, (2) correct for sampling bias via auxiliary population datasets, and (3) validate against ground truth event logs and remote-sensing measurements.

19. Automated Hypothesis Generation from Incomplete Scientific Databases

— We ask: how can we mine plausible, testable hypotheses by linking partial facts across heterogeneous, incomplete scientific databases?; We ask: how can we score generated hypotheses for novelty and experimental feasibility? — We will build a hypothesis-assembler that (1) extracts relation triples from structured and unstructured sources, (2) completes missing links via probabilistic knowledge graph embedding with uncertainty, (3) generates candidate hypotheses with provenance, and (4) ranks them by novelty, confidence, and resource-estimated feasibility with expert-in-the-loop evaluation.

20. Benchmarking Robustness of Data Mining Pipelines to Adversarial Data Poisoning at Scale

— We ask: how resilient are end-to-end data mining pipelines (preprocessing, feature extraction, mining algorithms) to realistic, resource-constrained poisoning attacks?; We ask: which pipeline stages are the most vulnerable and which defenses are most cost-effective at scale? — We will construct a benchmark suite that (1) exposes modular pipeline components, (2) defines attacker cost models and poisoning strategies for each stage, (3) measures downstream impact on mined outputs (patterns, clusters, explanations), and (4) provides defense baselines and an open leaderboard for reproducible evaluation.

21. Federated Causal Pattern Mining in Edge-IoT Streams Under Differential Privacy

We (TopicSuggestions) pose research questions: 1) How can we discover causal relationships across heterogeneous edge-IoT streams while guaranteeing differential privacy at each device? 2) What are the trade-offs between causal discovery accuracy and privacy budget in a federated setting? 3) How do asynchronous updates and intermittent connectivity affect federated causal inference?
We (TopicSuggestions) outline how to work: synthesize or collect diverse IoT stream benchmarks, implement federated causal discovery algorithms (e.g., adaptation of PC/FCI or Granger causality) with DP noise mechanisms, evaluate trade-offs via simulation and real deployments, and analyze robustness to connectivity patterns.

22. Adversarial Robustness of Synthetic Tabular Data Generators Across Domain Shift

We (TopicSuggestions) pose research questions: 1) How do common tabular data synthesizers fail under realistic domain shifts and adversarial perturbations? 2) Can we define robustness metrics that predict downstream model performance under label or covariate shift? 3) How can generators be trained to produce synthetic data that preserves robustness properties?
We (TopicSuggestions) outline how to work: build domain-shift benchmarks across multiple domains, measure generator-induced shifts with proposed robustness metrics, apply adversarial training or distributionally robust objectives to generators, and validate by training downstream models on synthetic data and testing on shifted real data.

23. Mining Temporal Signatures of Supply-Chain “Black Swan” Precursors Using Sparse Event Embeddings

We (TopicSuggestions) pose research questions: 1) Which sparse temporal event patterns act as early precursors to major supply-chain disruptions? 2) Can sparse event embedding models amplify weak precursor signals while controlling false alarms? 3) How early can we reliably predict systemic risk under noisy reporting?
We (TopicSuggestions) outline how to work: curate multi-source supply-chain event logs, design sparse/time-aware embedding methods (e.g., temporal LASSO + event embeddings), develop precursor scoring and early-warning thresholds, and validate lead-time vs. precision trade-offs on historical disruption cases.

24. Explainability-Aware Active Learning for Label-Scarce Medical Time-Series

We (TopicSuggestions) pose research questions: 1) How can explainability constraints be integrated into acquisition functions so clinicians prefer queried samples? 2) Does prioritizing explainable samples accelerate clinically useful model calibration? 3) How to quantify and optimize the trade-off between explainability, informativeness, and annotation cost?
We (TopicSuggestions) outline how to work: implement acquisition strategies that score candidates by informativeness and explainability (saliency stability, counterfactual clarity), run simulation studies with clinician-in-the-loop or proxy labelers, measure label efficiency and clinician trust, and iterate with user studies.

25. Mining Simulation Outputs to Optimize Wet-Lab Experiment Design Under Resource Constraints

We (TopicSuggestions) pose research questions: 1) How can we mine high-fidelity simulator outputs to recommend a minimal set of wet-lab experiments maximizing discovery probability under budget limits? 2) Which active-learning / Bayesian optimization hybrids best transfer from simulation to wet-lab reality? 3) How do uncertainty calibration and simulator bias affect experiment selection?
We (TopicSuggestions) outline how to work: pair computational simulators with meta-learning or Bayesian optimization to propose experiments, model cost-aware acquisition functions, validate with retrospective case studies and limited prospective lab collaborations, and incorporate simulator bias correction techniques.

26. Cross-Modal Concept Drift Detection in Augmented Reality Interaction Logs

We (TopicSuggestions) pose research questions: 1) How do concept drifts manifest across visual, gestural, and voice modalities in AR systems? 2) Can we detect modality-consistent vs. modality-specific drift and adapt models accordingly? 3) What are lightweight online adaptation strategies suitable for latency-sensitive AR pipelines?
We (TopicSuggestions) outline how to work: collect synchronized multi-modal AR interaction datasets, design drift detectors that fuse modality-specific statistics and cross-modal alignment signals, test online adaptation (e.g., modular re-weighting, continual learning) and measure latency, accuracy, and user experience.

27. Mining Economic Incentive Structures in Decentralized AI Marketplaces

We (TopicSuggestions) pose research questions: 1) How do payment rules and reputation systems influence strategic behavior, collusion, and model quality in decentralized AI marketplaces? 2) Can mining marketplace logs reveal early signs of incentive misalignment or Sybil attacks? 3) What incentive-compatible mechanisms minimize manipulative behaviors while preserving participation?
We (TopicSuggestions) outline how to work: simulate decentralized marketplaces with agent-based models, mine real or synthetic transaction logs to extract behavioral features, apply causal inference and anomaly detection to detect collusion patterns, and evaluate alternative mechanism designs via economic simulations.

28. Privacy-Preserving Discovery of Rare Phenotypes from Heterogeneous EHRs using Synthetic Cohorts

We (TopicSuggestions) pose research questions: 1) Can differentially private synthetic cohorts reliably preserve signals for ultra-rare phenotypes across heterogeneous EHR systems? 2) What synthesis and validation pipelines best separate privacy leakage from phenotype utility? 3) How to calibrate privacy budgets to protect individuals while enabling rare-event discovery?
We (TopicSuggestions) outline how to work: implement multi-site DP synthetic generation methods, create evaluation metrics for rare-phenotype fidelity, perform cross-site validation with holdout real cohorts, and produce guidelines for privacy-utility budgeting.

29. Graph Mining of Dynamic Research Collaborations to Predict Emergent Interdisciplinary Fields

We (TopicSuggestions) pose research questions: 1) Which dynamic graph features and temporal motifs predict the birth of enduring interdisciplinary subfields? 2) Can early identification of bridging teams forecast lasting citation and funding growth? 3) How do institutional and geographic factors mediate field emergence signals?
We (TopicSuggestions) outline how to work: construct time-resolved author-paper-institution graphs from bibliographic databases, extract temporal motifs and embedding trajectories, train predictive models for field emergence events, and validate predictions against retrospective field formation cases.

30. Mining Human-AI Interaction Logs to Identify and Mitigate Habit-Forming Recommender Behaviors

We (TopicSuggestions) pose research questions: 1) What temporal interaction patterns constitute habit-forming behavior induced by recommender systems? 2) How can mining logs reveal causal pathways from recommendation policies to decreased user autonomy? 3) What interventions (diversity injections, friction, timing) effectively mitigate habit formation without harming engagement?
We (TopicSuggestions) outline how to work: define operational metrics for habit formation (repeat-rate, narrowing of choices, reduced fallback behavior), mine large-scale interaction logs with causal inference tools, run A/B tests of mitigation strategies in simulation or partner platforms, and measure downstream user well-being proxies.

31. Federated, Multi‑Modal Concept‑Drift Detection with Differential Privacy

We ask: How can we detect and localize concept drift across heterogeneous, privacy‑sensitive clients that stream different modalities (text, sensor, images) without centralizing raw data? What tradeoffs arise between drift detection accuracy, localization granularity, and formal privacy budgets?
We overview how to work on it: We collect or simulate federated clients with multimodal streams, implement client‑side summary sketches (e.g., compact feature histograms, model gradient signatures, and representation drift statistics) and apply differentially private aggregation. We design drift tests that operate on aggregated summaries (e.g., change‑point tests on representation distributions, graph‑based divergence measures) and evaluate localization by attributing drift to clients/modalities. We measure detection delay, false alarm rate, attribution accuracy, and privacy loss (ε). We compare against centralized baselines and ablations varying sketch fidelity and privacy budgets.

32. Causal Pattern Mining in Outputs of Agent‑Based Simulations

We ask: How can we mine recurring causal substructures (not mere correlations) from large ensembles of agent‑based simulation traces to explain emergent phenomena and guide policy interventions? What mining algorithms can extract interpretable causal motifs under model stochasticity?
We overview how to work on it: We generate varied simulation runs (e.g., epidemiology, urban mobility) with controlled interventions. We encode traces as temporal graphs or event sequences, apply causal discovery (Granger, PCMCI, or constraint‑based causal pattern mining) at subgraph scale, and mine frequent causal motifs using pattern growth with causal constraints. We validate motifs by counterfactual re‑simulation and measure stability across noise and parameter changes. We produce visual, human‑interpretable motifs linked to actionable interventions.

33. Mining Ethical Bias Emergence in Synthetic Data Generation Pipelines

We ask: How can we detect, quantify, and attribute emergent ethical biases introduced at different stages of synthetic data pipelines (sampling, augmentation, generative modeling, filtering)? Can we design automated audits that localize which pipeline stage most contributes to specific harms?
We overview how to work on it: We assemble pipelines combining real datasets and multiple synthetic generators (GANs, diffusion, rule‑based). We define bias probes (task success disparity, stereotype amplification, subgroup calibration) and perform ablation by swapping pipeline components. We apply contribution analysis (Shapley or influence functions) over pipeline outputs to attribute bias, and propose corrective mechanisms (stagewise reweighting, adversarial debiasing). We evaluate on fairness benchmarks and human‑annotated harm assessments.

34. Mining Privacy‑Leakage Signatures in Feature Embeddings Across Transfer Learning

We ask: What measurable signatures in intermediate feature embeddings indicate private attribute leakage when models are fine‑tuned or transferred across tasks? How can we detect and mitigate leakage before deployment?
We overview how to work on it: We pretrain large models and fine‑tune on target tasks, then extract embeddings from layers. We design leakage detectors that predict sensitive attributes from embeddings (using probing classifiers, mutual information estimators, and clustering leakage metrics). We study how fine‑tuning, dataset shifts, and adapter layers change leakage signatures and develop layerwise mitigation (privacy heads, adversarially trained scramblers). We report leakage AUC, downstream task utility, and mitigation cost.

35. Mining Semantic Erosion in Long‑Lived Ontologies and Taxonomies

We ask: How can we detect and quantify semantic drift and erosion in ontologies/taxonomies that evolve over years due to new concepts, usage shifts, and annotation practices? Can we recommend principled merges/splits to restore consistency?
We overview how to work on it: We collect version histories of ontologies (biomedical, product catalogs) and associated corpora. We mine changes via embedding‑based concept representations, measure semantic drift (cosine drift, clustering changes), and identify erosion patterns (concept dilution, blurred boundaries). We propose corrective operations (relabeling, hierarchical rebalancing) through optimization that preserves backward compatibility. We validate by simulated user tasks (search accuracy, annotation agreement) before/after repairs.

36. Mining Adversarial Collaboration Patterns in Human–AI Co‑Creation Logs

We ask: How can we mine logs of human–AI co‑creation (code pairs, design iterates, writing drafts) to detect adversarial or manipulative behaviors (gaming, overreliance, adversarial prompting) and to recommend equitable interaction protocols?
We overview how to work on it: We instrument co‑creation platforms to capture turn‑level actions, prompts, and outcomes. We define adversarial collaboration behaviors (prompt injection, surprise edits, credit hijacking) and mine them using sequence mining, interaction motifs, and anomaly detection. We correlate patterns with outcomes (quality, fairness, authorship attribution) and design intervention models—interface nudges or constraint monitors—to reduce harmful patterns. We evaluate via A/B studies and quality/fairness metrics.

37. Mining Dataset Lifecycle Provenance to Predict Future Dataset Drift and Maintenance Needs

We ask: Can we use provenance metadata (collection methods, annotator profiles, preprocessing logs, update schedules) to predict where and when datasets will require maintenance due to drift, label decay, or hidden biases?
We overview how to work on it: We assemble provenance graphs across multiple datasets and timepoints, featurize lifecycle events, and train survival/time‑to‑event models to predict maintenance needs. We use explainable models to surface contributing factors (annotation protocol changes, sample sourcing shifts). We validate predictions against observed drift incidents and propose priority schedules for dataset audits and automated tests guided by the model.

38. Mining Privacy‑Preserving Graph Relational Patterns in Sparse, Heterogeneous Networks

We ask: How can we discover robust relational patterns (motifs, roles) in extremely sparse and heterogeneous graphs (multi‑relation knowledge graphs, IoT networks) while guaranteeing node/edge privacy under sublinear queries?
We overview how to work on it: We design private graph mining algorithms that operate on compressed summaries (random walk sketches, motif-counting via differentially private mechanisms) and adapt role discovery to heterogeneous edge types with regularization for sparsity. We analyze utility vs. privacy tradeoffs both theoretically and empirically on synthetic and real sparse networks, and benchmark motif recovery, role stability, and downstream link‑prediction performance.

39. Mining Real‑World Performance Degradation Signals for Continual Learning Systems

We ask: Which low‑level model telemetry signals (loss landscapes, gradient norms, activation distributions) most reliably predict future catastrophic forgetting or negative transfer in continual learning setups? Can we mine these signals online to trigger selective replay or architecture adaptation?
We overview how to work on it: We run continual learning experiments across task sequences and instrument extensive telemetry. We apply time‑series mining and feature importance analyses to map telemetry patterns to future performance drops. We design triggers and adaptive strategies (memory allocation, dynamic expansion) based on mined predictors and evaluate reductions in forgetting and resource use compared to fixed schedules.

40. Mining Policy Compliance from Multimodal Surveillance Streams with Fairness Guarantees

We ask: How can we extract interpretable policy‑compliance patterns (e.g., safety protocol adherence) from multimodal surveillance (video, audio, IoT sensors) while ensuring group fairness and minimizing surveillance bias?
We overview how to work on it: We create multimodal datasets with labeled compliance events and protected attributes, develop multimodal pattern miners that produce symbolic, human‑readable rules (hybrid neuro‑symbolic pipelines), and enforce fairness constraints during mining (reweighing, constrained optimization). We evaluate compliance detection accuracy, subgroup parity metrics, and interpretability via expert review, and iterate with human‑in‑the‑loop refinement.

41. Privacy-Utility Drift Quantification for Continually Updated Synthetic Datasets

We propose a framework to measure how privacy guarantees and downstream utility drift over time for synthetic datasets that are continuously retrained.
We ask: How does incremental retraining change the empirical privacy leakage and task utility? What metrics capture privacy-utility drift succinctly? How can we design alerts or retraining policies to bound drift?
We outline: We will instrument continuous-generation pipelines, compute membership and attribute inference risks periodically, track downstream model performance on held-out tasks, and develop composite drift metrics and thresholding policies. We will validate on healthcare and finance longitudinal datasets.

42. Multimodal Micro-Stressor Mining from Passive Wearable Streams

We explore mining short, low-intensity stress events (“micro-stressors”) using asynchronous multimodal wearable signals (HRV, skin conductance, audio snippets, motion).
We ask: What feature representations best capture micro-stressor onset and recovery? Can we mine individualized micro-stressor signatures without labeled events? How do micro-stressor patterns predict cumulative health outcomes?
We outline: We will apply unsupervised segmentation, contrastive multimodal representation learning, and weak supervision from calendar/context signals; then link discovered micro-stressor clusters to longitudinal health metrics.

43. Federated Attack-Surface Mining for Differential-Privacy Parameters in Real-World Deployments

We develop methods to mine and visualize vulnerable combinations of federated learning configuration and DP parameter settings that lead to practical privacy breaches.
We ask: Which combinations of client sampling, communication frequency, and noise schedules produce exploitable privacy gaps? How can we algorithmically enumerate worst-case configurations for a given deployment?
We outline: We will simulate federated settings with diverse heterogeneity, perform red-team membership inference and gradient inversion across parameter grids, and build an automated attack-surface mapper to recommend safe parameter regimes.

44. Graph Change-Point Mining in Decentralized Social Platforms under Local-Only Observability

We target detection and characterization of structural change points in social graphs when only local ego-network snapshots are accessible (e.g., decentralized platforms).
We ask: How can we infer global change events (coordination campaigns, sudden polarization) from sampled local views? What statistical tests and graph summarizations are robust under sampling bias?
We outline: We will formulate likelihood-based and embedding-drift detectors adapted to ego-sampled streams, validate with synthetic cascade injections, and apply to datasets from decentralized protocols or partial crawls.

45. Energy-Aware Data Mining: Mining Patterns to Drive Adaptive Compute Scheduling for Large-Scale Pipelines

We study mining temporal and workload patterns in data pipelines to drive adaptive scheduling that minimizes energy while preserving SLA.
We ask: Which mined features best predict energy spikes and idle windows? How can mined motifs be converted to scheduling heuristics that reduce carbon intensity?
We outline: We will collect telemetry from diverse ETL/ML pipelines, mine recurring workload motifs with time-series motif discovery, and integrate results into a scheduler simulator to quantify energy and latency trade-offs.

46. Causal Discovery from Reinforcement Learning Trajectories with Confounded Observations

We investigate methods to recover causal structure from RL trajectories where state observations are confounded or partially observed.
We ask: Under what conditions can we identify state-action causal graphs from policy-driven data? How do exploration strategies impact identifiability?
We outline: We will combine instrumental-variable style approaches with constraint-based causal discovery adapted to sequential data, run controlled simulation experiments, and test on logged interaction datasets from robotics and recommender systems.

47. Mining Bias Amplification in LLM-Feedback Loops across Automated Content Moderation Chains

We examine how biases are amplified when large language models are used in iterative moderation pipelines that feed outputs back into training data.
We ask: What quantitative signatures indicate bias amplification across moderation-feedback cycles? How can we detect and mitigate runaway stereotype reinforcement in automated pipelines?
We outline: We will simulate multi-round moderation pipelines, mine trajectories of bias metrics (e.g., sentiment/stereotype scores per subgroup), and propose intervention strategies such as resampling, adversarial debiasing, or calibrated uncertainty thresholds.

48. Predicting API Evolution Impact by Mining Low-Resource Open-Source Repositories

We propose mining small, low-star repositories to predict the impact of API changes on dependent code that is not well-indexed by large dataset sources.
We ask: How do API usage patterns in low-resource repos differ, and how do they influence downstream breakage risk? Can we predict high-impact API changes before widespread adoption?
We outline: We will aggregate and mine usage patterns from shallow forks and niche packages, build models to estimate sensitivity of code to API modifications, and validate predictions against historical API deprecations.

49. Provenance Trace Mining to Predict Reproducibility Failure Modes in Data Science Workflows

We mine provenance metadata (tool versions, parameter changes, data lineage) to predict specific reproducibility failure modes before full reruns.
We ask: Which provenance features are most predictive of failures (e.g., nondeterministic randomness, floating-point drift, dependency mismatch)? How can we prioritize checks to preempt costly reruns?
We outline: We will collect provenance from reproducibility studies and CI logs, apply supervised failure-mode classification and SHAP-style attribution, and produce a checklist/prioritization model for pipeline validation.

50. Latent-Activation Trajectory Mining for Early Anomaly Detection in Medical Imaging Pipelines

We explore mining activation-space trajectories of deep imaging models during training/inference to detect anomalies (scanner drift, preprocessing faults, poisoning) early.
We ask: What trajectory patterns in latent activations correlate with specific anomaly types? Can we deploy lightweight continuous monitors on activation summaries to trigger alerts?
We outline: We will instrument imaging models to record compact activation summaries, cluster and model trajectory dynamics with sequential autoencoders, and evaluate detection lead-time on simulated scanner faults and adversarial interventions.