A Unified Framework for Task-Driven Data Quality Management

High-quality data is a critical enabler for high-quality Machine Learning (ML) applications. However, due to inevitable errors, bias, and adversarial attacks occurring during the data generation and collection processes, real-world datasets often suffer various defects that can adversely impact the learned ML models. Hence, Data Quality Management (DQM) has become an essential prerequisite step for building ML applications.

DQM has been extensively studied by the database community in the past. Early works [1, 2, 3] consider DQM as a standalone exercise without considering its connection with downstream ML applications. Studies have shown that such ML-oblivious DQM may not necessarily improve model performance [1]; worse yet, it may even degrade model performance [4]. More recent work started to tailor the DQM strategies to specific ML applications [5, 6, 7]. Still, they apply only to simple models such as convex models, nearest neighbors, and specific data quality issues such as outlier detection. In parallel with these research efforts, the ML community has intensively investigated techniques focused on addressing a broad variety of data quality issues, such as adversarial [8, 9] and mislabeled [10, 11] data detection, anomaly detection [12], dataset debiasing [13, 14, 15]. However, a DQM scheme that can comprehensively remedy various types of data defects is still lacking.

Our paper aims to address the limitations of prior DQM schemes by developing a unified DQM framework with the following properties: (1) multi-purpose – to handle various data quality issues; (2) task-driven – to effectively utilize the information from downstream ML tasks; and (3) model-agnostic – to incorporate different ML models. The line of existing work closest to achieving these goals is what we will later refer to as data valuation-based approaches. These approaches first adopt some importance quantification metric, e.g., influence functions [10], Shapley values [16, 17] and least cores [18], to quantify each training point according to the contributions toward the training processes, then decide which data to retain or remove based on the valuation rankings. While some existing data valuation-based approaches satisfy the three desiderata, empirical studies have shown that their performance varies considerably based on the underlying data and the learning process. Moreover, there is no clear understanding of such performance variation or formal characterization of the worst-case performance.

In this paper, we start by formulating various DQM tasks into optimal data selection problems. The goal is to find a subset of data points that achieve the highest performance for a given ML task. We propose DataSifter, a multi-purpose, task-driven, model-agnostic DQM framework that first learns a data utility model from a small validation set, then selects the subset of data points by optimizing the acquired utility model. With the acquired data utility model, DataSifter can go beyond the functionalities offered by existing DQM schemes and further estimate the utility of selected data points. Such information could help data analysts to decide how many data points to choose or whether there is a need to acquire new data. Furthermore, we present a novel theoretical framework based on domination analysis which allows one to rigorously analyze the worst-case performance of data valuation-based DQM approaches and compare them with our approach. Specifically, we show that data valuation-based DQM approaches have unsatisfying worst-case performance guarantees. In particular, the popular Shapley value-based approach will select the worst data in some commonly occurring scenarios. We conduct a thorough empirical study on a range of ML tasks, including adversarially perturbed data detection, noisy label/feature detection, data summarization, and data debiasing. Our experiments demonstrate that DataSifter achieves and most often significantly improves the state-of-the-art performance of data valuation-based approaches on various tasks.

The major differences between this paper and the related works are summarized in Table 1.

Data Cleaning. Classical data cleaning methods are based on simple attributes of a dataset such as completeness [3], consistency [21], and timeliness [2]; however, these attributes may not necessarily correlate with the actual utility of data for training machine learning models. Recent works leverage the information about downstream ML tasks to guide the cleaning process. ActiveClean [5] explored task-driven data cleaning for convex models by selecting data points for human screening. BoostClean [22] sought to automate the manual cleaning by determining a predefined cleaning strategy from a library using boosting. AlphaClean [7] also aimed to automate the cleaning process but relied on parallelized tree search. However, those framework’s efficacy and generalizability are limited by the cleaning library. Furthermore, the recursive nature of the automatic selection process constrained the use case of those methods to only small models and datasets. CPClean [6] proposed a different strategy for nearest neighbor models based on the concept of Certain Prediction. Still, CPClean is designed explicitly for SQL datasets with greedy repairing, making it difficult to generalize to larger-scaled cases like image datasets. In summary, the state-of-the-art data cleaning methods are only applicable to certain classes of ML models and datasets. Besides, adapting those cleaning works to other domains requires manual recollection of the cleaning library or human intervention, which can be impractical in many cases.

Data Importance Quantification. One simple idea to quantify data importance is to use the leave-one-out error. [10] provides an efficient algorithm to approximate leave-one-out error for each training point. Recent works leverage credit allocation schemes originated from cooperative game theory to quantify data importance. Particularly, Shapley value has been widely used [16, 17, 20, 23, 24], as it uniquely satisfies a set of desirable axiomatic properties. More recently, [18] suggests that the Least core is also a viable alternative to Shapley value for measuring data importance. However, computing the exact Shapley and Least core values are generally NP-hard. Several approximation heuristics, such as TMC-Shapley [16], G-Shapley [16], KNN-Shapley [20], have been proposed for the Shapley value. Despite their computational advantage, they are biased in nature. On the other hand, unbiased estimators such as Permutation Sampling [19] and Group Testing [17] still require retraining models many times for any descent approximation accuracy. TracIn [11] estimates the importance by tracing the test loss change caused by a training example during the training process. The representer point method [25] captures the importance of that training point by decomposing the pre-activation prediction of a neural network into a linear combination of activations of training points. Many of the aforementioned works can only be applied to differentiable models.

In general, DQM aims to find a subset of data points with the highest utility. We use the data utility function to characterize the mapping from a set of data points to its utility. Formally, given a dataset $\mathcal{D}$ of size $n$, a data utility function $U:2^{\mathcal{D}}\rightarrow\mathbb{R}$ maps a set of data points $S\subseteq\mathcal{D}$ to a real number indicating the performance of the ML model trained on the set, such as test accuracy and fairness.

With the notion of the data utility function, one can abstract DQM tasks as a data selection problem: $\max_{|S|=k}U(S)$, where $k$ indicates the selection budget with $0<k<n$, which can be predetermined (e.g., based on the prior knowledge about potential data defects or computational requirements). Moreover, the DQM tasks without a specific selection budget can be reduced to a sequence of data selection problems with different values of $k$.

With the abstraction above, one straightforward way to optimally select data is to exhaustively evaluate $U(S)$ for all possible size-$k$ subsets $S\subseteq\mathcal{D}$ and choose the one that achieves the highest utility. Of course, this naive algorithm requires prohibitively large computational resources because the number of utility evaluations is exponential in $k$, and worse yet, each evaluation of data utility function requires retraining the model. Fortunately, recent work shows that many common data utility functions can be effectively learned with a relatively small amount of samples [26] because they are “approximately” submodular [27]. The “approximate submodularity” property allows efficient maximization of data utility functions through simple greedy algorithms [28, 29, 30, 31, 32]. Hence, combining data utility function learning and greedy search enables an efficient algorithm for data selection problems.

Specifically, we extend and generalize the data utility learning and optimization technique originally proposed in [26] for active learning to DQM. The proposed DQM framework, termed DataSifter, proceeds in two phases: learning and selection phase.

Learning Phase. Figure 1 depicts the learning phase of the DataSifter, which consists of a utility sampling step and a utility model training step. In particular, we assume that we have access to a small validation set representative for potential test samples. Thus, the utility of any given subset can be estimated by feeding the model with this subset then evaluating its performance over the validation set. In the utility sampling step, we randomly sample subsets of the training set, estimate the utility of each sampled subset, and label each using its utility score. We will refer to the scored subset as utility samples hereinafter. To accelerate this step for large models such as deep nets, a small proxy model (such as logistic regression) can be used for approximating the utility since data utilities evaluated on deep nets and logistic regression are positively correlated, as shown in [15]. In the utility model training step, we learn a parametric model for the data utility function using the utility samples; particularly, our experiments adopt DeepSets [33] as the utility model. For a large dataset, the utility sampling step could be conducted on a small portion of the dataset. Our empirical studies show that the learned utility model can still extrapolate the utility for the unseen part of the dataset.

Selection Phase. We select high-quality data through optimizing the learned model for data utility functions obtained from the previous phase; specifically, we adopt a linear-time stochastic greedy algorithm [34] to perform optimization.

Clearly, DataSifter is an optimal solution to the data selection problem if the validation data matches the test data exactly and there are no computational constraints. In practice, despite limited validation dataset and limited computational resources, DataSifter is still very effective in selecting high-quality data or filtering bad data as we will show in the evaluation section. In addition, with the learned data utility model, DataSifter can provide an estimate of the utility for the selected dataset (see example in Figure 2), which will be useful for data analysts to decide the number of data points to select.

This section presents a theoretical framework for comparing the worst-case performance between DataSifter and data valuation-based DQM schemes, such as leave-one-out (LOO), Shapley value, and Least core¹¹1Least core may not be unique. In this paper, when we talk about the least core, we always refer to the least core vector that has the smallest $\ell_{2}$ norm, following the tie-breaking rule in the original literature [18]., and we assume no computational constraints.

We start by abstracting a general notion from data valuation-based DQM schemes in the literature. We call an algorithm that returns $S\subseteq\mathcal{D}$ of size $k$ a heuristic to a (size-$k$) data selection problem on $\mathcal{D}$. The typical pattern of data valuation-based heuristics is that they first rank the data points according to their corresponding data importance metric and then prioritize the points with the highest importance scores. We will define the heuristics matching this selection pattern as linear heuristics.

Our theoretical framework for studying the worst-case performance of data selection heuristics extends the domination analysis initially proposed in [35]. Our worst-case performance metric is domination number, which measures how many subsets achieve lower utility than the selected set in the worst-case scenario.

The domination number is well defined for every data selection heuristic. A heuristic with a higher domination number may be a better choice than a heuristic with a smaller domination number due to the better worst-case guarantee. The best heuristic for data selection has domination number $d(n,k)={n\choose k}$ for every $k\leq n$, which means that it will select the size-$k$ data subset with the highest utility for every possible data utility function.

Clearly, assuming no computational constraints, DataSifter is among the best heuristics which achieve the largest possible domination number. In contrast, the following result shows that no linear heuristic is the best whenever $n\geq 3$. We will defer all proofs to Appendix.

Furthermore, we can tighten the upper bound of the domination number for data valuation-based heuristics by noticing another common property: two data points will receive the same importance score if they contribute equally to all possible subsets of the training data. This property is often referred to as symmetry axiom in the literature.

The symmetry axiom may be desired for application scenarios requiring fairness, e.g., data importance scores are used to assign monetary rewards for data sharing or responsibility for ML decisions. However, for data selection, symmetry axiom may be undesirable because simply gathering high-value data points may lead to a set of redundant points. Based on this intuition, the following theorem gives an upper bound of domination number for non-trivial linear heuristics that with symmetry property.

To better illustrate the issue raised by symmetry axiom, we evaluate the LOO, Shapley, and least core heuristic on a synthetic dataset with 15 training data points (so that we can compute the exact Shapley and least core values, as well as obtain the optimal solution for data selection problem). The utility metric is the test accuracy of a Support Vector Machine (SVM) classifier trained on the dataset. We simulate the natural redundancy in a dataset by replicating 5 data points three times and adding slight Gaussian noise to differentiate. Figure 3 shows that with small selection budgets, the subsets selected by all the heuristics have low utility as the heuristics fail to promote diversity during selection.

Notably, we show that the Shapley value heuristic would select the data subset with the lowest utility for certain data utility functions, including submodular ones. The Shapley value of a training point is calculated by taking a weighted average of the contribution of the point to all possible subsets of the training set, and the weights are independent of the selection budget $k$. Moreover, the Shapley value of training data weights higher for its marginal contributions on small datasets. Thus, data points that make a larger contribution on tiny datasets may be assigned with higher Shapley value, even if they make little or negative contributions in every dataset of desired selection size $k$.

We evaluate DataSifter on six DQM tasks, as listed in Table 2. We consider various benchmark models and datasets used in past literature for each DQM task. Since we can observe similar results on different datasets, this section will only describe the result on one representative dataset for each task and leave the other dataset in the Appendix. Finally, we discuss the scalability of the DataSifter on larger datasets. The implementation details and the additional results are presented in the Appendix.

We now discuss two limitations of the DataSifter.

This paper presents DataSifter as a unified framework for realizing task-driven, multi-purpose, model-agnostic data quality management. We theoretically analyzed the worst-case performance of existing data valuation-based DQM schemes and show that these approaches suffer unsatisfying performance guarantees. This sheds light on the empirical observations that existing data valuation-based DQM schemes exhibit significant performance variation over different datasets and tasks. Based on an extensive evaluation of the DataSifter over six types of DQM tasks and eight different datasets, we showed that DataSifter is more comprehensive and robust than the state-of-the-art DQM approaches with similar design goals. For future work, we would like to further improve the scalability of the DataSifter as well as design utility learning models that are better aligned with the properties of data utility functions.