Data Repair Practices: Algorithms & Workflows

• Data Repair Practices are systematic approaches that combine rule-driven, data-driven, and hybrid techniques to restore integrity in databases, cloud systems, and more.
• They use methodologies such as probabilistic inference, graph-based repair, and active learning to minimize errors and improve downstream model performance.
• Practical workflows integrate automated algorithms with human-in-the-loop adjustments to address technical challenges and socio-economic constraints in data restoration.

Data repair practices are the suite of methodologies, algorithms, social protocols, and computational frameworks developed to restore, correct, or render usable data that is compromised by errors, inconsistencies, corruption, incompleteness, or inaccessibility. This domain encompasses a spectrum of technologies—from rule- and data-driven algorithmic cleaning to labor-intensive, expert-guided recovery—applied in settings ranging from relational databases and cloud storage systems to informal "data recovery" shops in resource-constrained environments. Contemporary research emphasizes both the technical rigor of algorithmic strategies and the socio-technical realities governing data repair in practice.

1. Taxonomies and Classes of Data Repair Approaches

Recent surveys structure the field of data repair along two principal axes: information exploitation and repair-goal orientation. The major classes are as follows (Ni et al., 2023):

• Rule-Driven Repair: Methods use a fixed set of integrity constraints such as FDs (functional dependencies), DCs (denial constraints), CFDs (conditional FDs), or user-defined logic. Examples: Holistic [Chu et al.], BigDansing, Horizon, MLNClean, Daisy, NADEEF.
• Data-Driven Repair: Algorithms infer statistical regularities or train predictive models on dirty data (possibly with seed labels) to correct errors, often optimally (Scare, Baran).
• Combined Rule+Data-Driven Repair: These blend the above, leveraging both hand-crafted or mined rules and distributional/statistical signals for holistic or tolerant repairs (HoloClean, Unified, Relative).
• Model-Driven Repair: The repair process is guided by performance metrics over downstream ML models, selecting fixes to maximize predictive accuracy or minimize error (BoostClean).

The principal goals addressed include:

• Consistency Repair—restoring rule compliance with minimal changes.
• Holistic Repair—maximally correcting all truly erroneous cells, even those not explicitly violating rules.
• Tolerant Repair—jointly editing data and rules, informed by trust bounds or MDL principles.
• Model-Performance Repair—fixing only what demonstrably improves predictive tasks.

2. Algorithmic Frameworks and Representative Techniques

A broad array of algorithmic strategies structure the research landscape:

Methodology	Paradigm	Information	Repair Principle
Holistic	Rule-Driven	DCs	Conflict hypergraph, enumerate repair contexts
BigDansing	Rule-Driven	Rule DSL	Logical plan pruning, staged fix generation
Horizon	Rule-Driven	FDs	Pattern-graph, support-based path selection
HoloClean	Rule+Data	Constraints+Stats	Probabilistic inference (factor graph), MAP repair
Scare	Data-Driven	Classifiers	Block-wise prediction, optimize global consistency
Baran	Data-Driven	Labeled/Unlabeled	Rich candidate sets, few-shot classifier training
Unified	Rule+Data	DCs+Distributions	MDL-based joint optimization (data+rules)
BoostClean	Model-Driven	Validation metric	Greedy boosting over candidate repairs

(Fully detailed method descriptions and their relative trade-offs are systematically compared in the experimental survey (Ni et al., 2023).)

Specializations of data repair also occur within storage and knowledge-base systems:

• Distributed Storage Repair: Techniques such as opportunistic regenerating codes adapt the repair protocol dynamically based on network or failure conditions, achieving orders-of-magnitude boosts in reliability while minimizing bandwidth overhead (Aggarwal et al., 2013, Hu et al., 2017, Calis et al., 2017).
• Graph Database Repair: Repair models for knowledge graphs utilize Reg-GXPath-based integrity constraints, and introduce weight/multiset-based preference criteria for selecting among subset/superset repairs, with complexity that sharply depends on the expressiveness of the constraints (Abriola et al., 2023).

3. Metrics and Evaluation Methodologies

Traditional metrics—precision, recall, and F1—report the correctness of individual repaired cells but may mask the true net effect of repair operations. As high ER-F1 can coincide with negative net error reduction, an alternative metric, the Error Drop Rate (EDR), has been proposed (Ni et al., 2023):

$$\text{EDR} = \frac{\mathrm{DEC} - \mathrm{IEC}}{\mathrm{OEC}}$$

where DEC = true errors fixed, IEC = initially correct cells corrupted by repair, OEC = count of original errors. EDR>0 indicates net error reduction. This metric enables fairer comparisons across data sets and provides a more meaningful measure in practical deployments.

Evaluation protocols increasingly also consider downstream effects, such as changes in classifier or regressor accuracy and robustness to various classes of simulated error (e.g., domain-swaps, typographic corruption).

4. Practical Repair Workflows and Social Dimensions

Fieldwork in varied contexts demonstrates that technical approaches to data repair exist in a complex ecosystem of social, economic, and infrastructural constraints (Rahman et al., 2 Feb 2026). For example, in the context of urban Bangladesh:

• Data repair encompasses not only digital restoration (e.g., recovering files, repairing corrupted images) but also analog-to-digital conversion and account restoration.
• Resource limitations drive reliance on pirated software, hardware improvisation, and informal knowledge-sharing via invitation-only social channels (Telegram groups).
• Specialist expertise is protected by tight knowledge curation, with economic reasoning mediating decisions on training, documentation, and pricing.
• LLMs (e.g., ChatGPT) are a last resort; repairers frequently encounter hallucinated instructions and refusal to discuss unauthorized practices.
• Pricing structures incorporate technical effort, data emotionality, customer profile, and market sustainability, with tactical variations for repeat/organizational clients.

These factors are critical for understanding the durability and scalability of repair practices outside formal organizations and must be considered in the development of equitable data-infrastructure policy and technology.

5. Recent Advances: Fairness, Human-in-the-Loop, and Subset Repairs

Fairness-Aware Repair

Recent regulatory demands require precision repairs to mitigate algorithmic unfairness. Algorithms employing optimal-transport (OT) plans transform feature distributions to enforce conditional independence between protected and non-protected attributes (u-conditional fairness), applying sequential repair plans scalable to archival data (Langbridge et al., 2024). Performance is assessed via proxy metrics such as the KL-symmetrized divergence across fairness slices and practical guidelines are outlined for grid-size, KDE bandwidth, and regularization.

Human-Centric and Guided Repair

Human-in-the-loop frameworks target the minimization of user burden via principled intervention selection. Guided Data Repair (GDR) employs value-of-information-based group selection followed by active-learning-driven user consultation, achieving high data-quality improvements at fractional annotation effort (Yakout et al., 2011). This workflow integrates classifier training, groupwise repair proposal, and user-in-the-loop confirmation, with utility framed in terms of the expected reduction in a data-quality loss function.

Subset Repair with Topological Adaptivity

Modern relational repair focuses on subset repair under functional or conditional dependencies, often by deleting tuples to resolve constraint violations. Topology-aware approaches combine local-density estimates, conflict-degree penalties, and graph decomposition to generate robust repairs, optimizing both data preservation and violation removal (Zhao et al., 27 Jan 2026). The integration of entropy-weighted densities (EntroCFDensity) combats density bias, while mixed-integer programming and efficient heuristics (PPIS, MICO) are empirically shown to yield superior F1 and clean-data retention rates across a range of schemas.

6. Operational Challenges and Recommendations

Comprehensive experimental studies document that most repair algorithms, when used off-the-shelf, risk inflating error rates by inappropriately overwriting correct values. A lightweight and effective optimization is to restrict modification to cells flagged as erroneous by a state-of-the-art error detector (Ni et al., 2023). This “flag-then-repair” filter sharply reduces overfitting and net error creation, with notable improvements for rule-based engines. Furthermore, empirical results invalidate the assumption that maximal cleaning always yields optimal downstream model performance, as measured models sometimes outperform on judiciously repaired (even non-perfectly-clean) data.

Practical guidelines are case-dependent:

• For maximum error reduction, few-shot data-driven methods (Baran) are preferable.
• For rapid, good-enough repair with tight time constraints, probabilistic (MLNClean) or hybrid (Unified) techniques are recommended.
• When downstream ML task performance is critical, repair algorithms should be benchmarked on end-task metrics, and model-aware strategies (BoostClean, Scare) are suitable.
• For mostly numeric regression tables, mean/mode imputations can match full-scale repair.

Challenges include algorithmic scalability to very large tables, integration of rule discovery and repair, fairness-aware cleaning, and effective cell-level influence estimation for targeted repair selection. There is ongoing investigation into leveraging LLMs for candidate suggestion and schema-adaptive correction, though domain expertise bottlenecks and hallucination risk remain substantial barriers.

---

The data repair landscape is thus highly multidisciplinary, integrating formal algorithmics, statistical inference, ML, human judgment, and ethnographic insight. Technical progress in repair algorithms must be matched with sensitivity to the lived realities and infrastructures that mediate practical access to data, especially in resource-poor or non-Western settings (Rahman et al., 2 Feb 2026, Ni et al., 2023).