Data wrangling (DW) is a fundamental step to prepare data for downstream mining tasks. Recent studies explore large language models (LLMs) to form a lightweight DW paradigm. Such studies typically require prompting an LLM with a DW task together with a few examples as task demonstrations (i.e., in-context learning). A problem yet to be explored is how to select the examples, to maximize task effectiveness given constraints on the size of the examples. To fill this gap, we introduce the constrained Shapley value (CSV), a tailored variant of the Shapley value with a constraint on the LLM prompt size, to guide example selection. We show that CSV has desirable properties in example importance estimation. Using CSV directly for LLM-based DW is still computationally intractable. We further propose activated contribution (ACSV) as an unbiased estimation for CSV and sample allocation algorithms with approximation guarantees. Empirical results show that, compared with DW examples manually selected by experts, CSV improves the effectiveness of LLMs for DW tasks including schema mapping, entity matching, error detection, and missing value imputation by 5.90% averagly in F1 score, demonstrating the general applicability of CSV for in-context learning example selection towards DW tasks.

Entity Matching (EM) is a crucial task in data integration. Compared to deep learning-based EM solutions, tree-based machine learning models are more computationally effective and explainable, making them more applicable in real-world EM scenarios. However, Random Forest-based EM methods select features with a static feature engineering rule set for all attributes. Consequently, they suffer model retraining cost to select features, and can hardly customize to different EM tasks. To tackle this problem, we propose Description-Similarity Rules (DSR) for EM feature engineering. DSR introduces diverse attribute value distribution metrics and data-driven thresholds to traditional EM feature engineering rules. Unfortunately, both the DSR search space and its online model retraining costs are exponential. By pushing the model retraining to the offline stage, our DSR set mining algorithm is one order of magnitude faster than the baseline algorithms, taking only seconds for online selection. Empirically, DSR outperforms the feature engineering methods by 3.75 % on average F1 score, while reaching the state-of-the-art EM performance on several datasets.

Data cleaning is an essential technique to enhance data quality. Despite the proposal of various algorithms with different cleaning strategies, current automated cleaning technologies still fall short of practical requirements when dealing with large-scale data containing mixed errors. This paper presents UniClean to efficiently solve the mixed error cleaning problem with three key technical contributions. (1) A unified construction and extension method for cleaners, enabling cleaning methods to easily utilize various cleaners to perform cleaning tasks. (2) Three optimization strategies to achieve efficiency-oriented cleaning preparation. (3) A cleaning algorithm based on an optimized cleaning process to effectively clean mixed errors. UniClean achieves a time complexity of O (| D error | 4 · | Op | + |D| · | D error |), significantly enhancing scalability. Experiments on public and large-scale enterprise datasets demonstrate that UniClean achieves over 40% improvement across five metrics, compared to five state-of-the-art cleaning methods, and delivers more than 30% gains in F1 and REDR on complex datasets, while completing the cleaning process within hours even for millions of records.