Data Valuation in Machine Learning: "Ingredients", Strategies, and Open Challenges

1 Introduction

As machine learning (ML) and data sharing become more ubiquitous, there is greater interest by data owners in how much their data is worth and by model owners in how to explain model predictions and what data to add or remove to improve model performance. The former is evidenced by the emergence of data marketplaces (e.g., Ocean Protocol) for users to buy and sell data and initiatives such as the Data Dividend Project where consumers are banding together to demand property rights to their data under the California Consumer Privacy Act and receive payments from big tech companies [Rozemberczki and Sarkar, 2021]. The latter is evidenced by the popularity of interpretable ML methods in various fields, including healthcare.

The above concerns can be addressed via data valuation. In data valuation, the value of data contributed by a data owner is influenced by the data from other owners. Data valuation can be used in collaborative ML (CML) to determine a fair compensation (or share of a fixed monetary sum) for every data owner to incentivize them to share data [Sim et al., 2020; Tay et al., 2022]. The fair compensation is usually some affine or monotonic transformation of the contributed data value [Jia et al., 2019a; Sim et al., 2020]. Data valuation can also be used in interpretable ML to attribute the model predictions and accuracy to the most responsible training examples from the data. Often, only the ranking of data values (instead of their absolute values) is needed. Model owners can improve their models by actively collecting data similar to the highvalue training examples or removing noisy and misleading training examples with low values [Ghorbani and Zou, 2019; Yoon et al., 2020]. Lastly, data valuation is useful for domain adaptation by assigning higher value to data with an underlying distribution similar to the target one [Yoon et al., 2020].

∗Equal contribution.

It is thus timely to introduce a technical survey that (a) provides a new formal study of data valuation in ML through its "ingredients" and the corresponding properties (e.g., performance metric and its design principles) (Sec. 2), (b) grounds the discussion of common desiderata satisfied by existing data valuation strategies (e.g., Shapley value) on our proposed ingredients (Sec. 3), and (c) identifies open research challenges for designing new ingredients, data valuation strategies, and cost reduction techniques with specific examples (Sec. 4). All these distinguish the novel contributions of our survey from that of Cong et al. [2021] and Pei [2020]. Other surveys have either focused on the economics perspective (e.g., nature of data as a merchandise, costs for data acquisition, government regulation) [Fricker andMaksimov, 2017; Liang et al., 2018; Muschalle et al., 2013; Raskar et al., 2019; Zhang and Beltr´an, 2020] or the cooperative game theory (CGT) based valuation strategies [Bax, 2019].

2 "Ingredients" of Data Valuation

Suppose that there is a set of n data owners denoted as N ≜ {1, . . . , n}. The objective of data valuation is to determine the value (or contribution/payoff) {ϕk}k∈N of dataset Dk for data owner k. The valuation function ϕ is a function of several "ingredients": the family of datasets D ≜ {Dk}k∈N from all data owners, the learning algorithm A, and the performance metric U. Formally, ϕk ≜ ϕ(k, D, A, U), as shown in Fig. 1. The learning algorithm A is a function that takes in an arbitrary training dataset D and outputs a model. The performance metric U takes in any model A(D) or dataset D to output a real-valued performance score. For any subset C ⊆ N of data owners, DC ≜ ⋃k∈C Dk, so, DN denotes the aggregated dataset of all owners. Before giving examples of the data valuation strategy ϕ, we will discuss how the choices of the ingredients can affect the choice and computation of ϕ.

2.1 Data, D

Both the number n of data owners and the maximum dataset size maxk∈N |Dk| affect the cost of computing or approximating ϕ. The presence of noisy data affects the choice of the performance metric U [Ghorbani and Zou, 2019]. Lastly, whether Dk is a fixed dataset or a random/biased sample from an underlying distribution may also affect the choice of ϕ [Abay et al., 2019; Koenecke and Varian, 2020].

2.2 Learning Algorithm, A

The value of data may vary across learning algorithm A, and the choice of A affects the choice of performance metric U. When A is undecided, U would solely be designed based on data-driven principles. As another example, when the model has no parameters or the parameters are not directly comparable across models (e.g., neural networks), the performance metric cannot be directly based on the parameters. Furthermore, the choice of A (and U) affects the cost of computing ϕ and its approximations. We will outline several learning algorithms that are important to our discussion of data valuation.

The k-nearest neighbours (k-NN) is a commonly used learning algorithm that predicts the label for a datum based on the known values of several similar data/nearest neighbours [Jia et al., 2019a]. Another class of learning algorithms involves empirical risk minimization (ERM) using gradientbased techniques such as logistic regression and deep neural networks (DNNs) [Ghorbani and Zou, 2019]. Bayesian models such as Bayesian linear regression offer the distinct advantage of providing an uncertainty measure in the predictions; intuitively, a lower uncertainty is more desirable [Sim et al., 2020]. Generative models [Goodfellow et al., 2014; Kingma and Welling, 2014] are typically used to learn an effective representation of the data and can be used to generate synthetic data that can be operationally easier to share than real data due to privacy regulation [Tay et al., 2022].

2.3 Performance Metric, U

The performance score UA(D) ≜ U(A(D)) or U(D) for more desirable models or datasets should be higher. We have identified a few design principles which are model- or datadriven. Table 1 summarizes how different performance metrics in existing work satisfy multiple design principles.

Model-driven Principles

• P1 Validation performance (VP). Intuitively, a model with better performance on a validation set is more valuable. Possible choices of VP include accuracy [Ghorbani and Zou, 2019; Jia et al., 2019b] and negated loss functions. VP also depends on the task, such as whether we are predicting the likelihood vs. severity of illness (in a medical use case). VP can distinguish between data that improve model performance from data that worsen model performance (e.g., noisy/adversarial data or from a different domain). Thus, VP is suitable for use cases that involve model interpretability, domain adaptation, and noisy/adversarial data. However, it may be practically difficult to obtain a validation set that all data owners agree on [Sim et al., 2020].
• P2 Uncertainty. A model with a lower uncertainty in its model parameters and predictions is more valuable. This principle is theoretically ensured by using information gain (IG) on the model parameters/function from observing the data outputs as the performance metric [Sim et al., 2020]. IG measures the expected reduction in entropy capturing uncertainty across all possible data outputs. Moreover, IG can be used as a surrogate measure of the predictive accuracy of a trained model [Krause and Guestrin, 2007; Kruschke, 2008] when the test queries are not known a priori. Hence, it is especially useful when a validation set is not available.
• P3 Order of data arrival. For certain A and optimization techniques (e.g., deep neural networks (DNNs) and gradient-based optimization), sharing the same data earlier is more valuable. The rationale is that model performance is less likely to be improved or influenced by new data if the model (e.g., DNN) is closer to convergence at the later stages in a gradient-based optimization. Thus, data owners should be more highly valued for sharing data in the early stage of such optimization, like in federated learning [Song et al., 2019].

Data-driven Principles

These principles are defined based on D instead of the model (i.e., U(D) directly instead of U(A(D))).

• P4 Monotonicity. Having more data is more (or at least equally) valuable [Sim et al., 2020; Tay et al., 2022; Xu et al., 2021b].
• P5 Submodularity. The same data is less valuable when added to a larger dataset (i.e., law of diminishing returns). The IG on Bayesian model parameters as U satisfies submodularity [Sim et al., 2020].
• P6 Similarity to reference distribution. Dataset D is more valuable when it is similar and representative of the aggregated data DN (or the target task's distribution). Tay et al. [2022] measured similarity using the translated negative maximum mean discrepancy (MMD) between the underlying distributions P and P∗ for the respective datasets

D and DN, i.e., -MMD(P, P∗ ). As this quantity is intractable, Tay et al. [2022] have resorted to its (biased) estimate -̂ MMDb(D,DN ∪ G), where G is a synthetic dataset generated by a generative model trained on DN. Separately, Xu et al. [2021a] have considered the cosine similarity (vector alignment) between (i) the model parameter update from using data from each owner k only and (ii) the aggregated model parameter update across all owners (which uses data from every owner).

• P7 Intra-diversity. Dataset Dk is more valuable and will improve the model's predictive performance if it contains more diverse data points covering a larger region of the input space. Xu et al. [2021b] have measured diversity using the volume (i.e., square root of the determinant) of the left m × m Gram matrix as U (of data with m features). The volume is theoretically shown to positively correlate with the regression performance: A high volume suggests a low mean squared error.
• P8 Replication robustness. Replicated data should not be highly valuable as adding the same data repeatedly to an ML model does not improve its performance significantly. Xu et al. [2021b] have provided an upper bound on their diversity-based performance metric from direct (up to infinite) replication, which does not increase the diversity in data. This principle prevents dishonest owners from gaming the data valuation [Ohrimenko et al., 2019].
• P9 Cost. Data that is costly (to obtain/maintain) should be more valuable to incentivize the data owners' participation [Heckman et al., 2015]. The costs include that of data acquisition, storage, and transport [Agarwal et al., 2019; Devereaux et al., 2016].
• P10 Timeliness. Data that is timely (i.e., up-to-date) is more valuable [Stahl and Vossen, 2016]. Service providers can adjust their market operation strategies using real-time data to provide customized services, for example, to users of the bicycle-sharing services [Zhao et al., 2020]. This strategy is ineffective with outdated data.

3 Data Valuation Strategies, ϕ

We will discuss various data valuation strategies ϕ. ϕ is a more general concept than the performance metric U as ϕ can additionally depend on the datasets DC of others. In other words, ϕk ≜ U(Dk) or UA(Dk) is only one possible specification of ϕ. We will describe several desiderata from existing works and cooperative game theory (CGT) literature for designing ϕ. Importantly, the suitable desiderata to a certain application scenario of data valuation (and the properties of its ingredients) can guide the specific choice of ϕ. For example, fairness is related to how well each owner's value ϕk reflects his contribution and is captured by D4-D6 [Chalkiadakis et al., 2011]. To ensure fairness, we look for a ϕ that satisfies D4-D6 in Table 2. We adopt the notation convention from CGT and let v be the characteristic function that maps each subset of owners/coalition to a value, i.e., v : 2N ↦→ R. For data valuation, we define v(C) ≜ UA(DC) or U(DC).

1 A possible use case is when the data owners cannot agree on the downstream model/task.

• D1 Low computational cost. The total cost of computing ϕ, which depends on the number of evaluations of A and the cost of each evaluation, should be low enough to be tractable/efficient. The number of evaluations may depend on the number n of data owners. When A is expensive and n is large, we will prefer ϕ that can be computed without or with fewer (e.g., O(logn)) evaluations of A. As an example, when millions of data owners [Singhal et al., 2021] contribute their data for the training of a DNN that minimises the empirical risk, we want to avoid retraining on multiple data subsets from scratch.
• D2 Dependence on D. The value ϕk of Dk should depend on other datasets in D [Agarwal et al., 2019]. Formally, owner k's relative valuations to j, (i) ϕk - ϕj and (ii) ϕk/ϕj cannot be simplified into a function g(Dk,Dj,A,U) which is independent of the others' datasets. To illustrate, suppose that the dataset Di of owner i ̸= k, j is updated to include a duplicate of Dj. We expect owner k's relative valuations to increase as j's data is made redundant by Di. (i) may affect the ranking of owners k and j in interpretable ML while (ii) will alter their compensation in CML [Sim et al., 2020; Xu et al., 2021a]. One possible dependency is that more unique data should be more valuable. 2 Formally, if Dj is similar to others' data and Dk is exclusive to owner k, it is possible for ϕk > ϕj even if U(Dk) ≤ U(Dj).
• D3 Clone robustness. If an owner k additionally joins the collaboration as a clone k′ (i.e., Dk′ = Dk), then the value assigned to k (and k′ ) should not increase. Formally, let ϕ′k denote the value of owner k in the clone collaboration setting. We require ϕk > ϕ′k + ϕ′k′ . This will ensure that the owners cannot unfairly obtain a higher

2 Technical whitepaper: Guide To Data Valuation For Data Sharing (link).

valuation [Ohrimenko et al., 2019]. 3

• D4 Uselessness. If including Dk does not improve the performance score for any coalition C with aggregated data DC, then owner k should be valueless: For all k ∈ N, (∀C ⊆ N \ {k} v(C ∪ {k}) = v(C)) ⇒ ϕk = 0.
• D5 Symmetry. If including Dk yields the same improvement as including Dj in the performance score for any coalition C (e.g., Dk = Dj), then their values are equal: For all k, j ∈ N s.t. k ̸= j, (∀C ⊆ N \ {k, j}, v(C ∪ {k}) = v(C ∪ {j})) ⇒ ϕk = ϕj .
• D6 Strict Desirability. 4 Ensuring fairness may extend beyond symmetry to require owner k to have a higher value than owner j if including Dk yields a greater improvement than includingDj to some coalition, but the reverse is not true. For all k, j ∈ N s.t. k ̸= j,

For weak desirability, > is replaced by ≥. Intuitively, strict and weak desirability enforces whom a data owner k must outvalue or cannot be outvalued.

• D7 Group Rationality. 5 The value v(N) of the aggregated dataset equals the total value distributed to the N data owners, ϕN = ∑k∈N ϕk, i.e, v(N) = ϕN. This is important when the absolute value of ϕk is used, e.g., v(N) denotes the revenue (from the collaboration) to be distributed among the data owners according to ϕk [Richardson et al., 2020]. Note v(N) > ϕN implies wastage while v(N) < ϕN is not feasible without external resources. On the other hand, this desideratum may be unnecessary when only the ranking of data points is used (e.g., in interpretable ML applications) or when ϕk is rescaled (e.g., in CML applications where owners share a fixed sum).
• D8 Additivity. The data value computed using the performance metric U = U1 + U2 is the sum of the data value computed using each performance metric independently. Formally, for all k, ϕ(k,D,A,U) = ϕ(k,D,A,U1) + ϕ(k,D,A,U2). As an example, U1 and U2 can be the validation accuracy on two separate validation sets stored on separate servers. Thus, decentralized valuation becomes feasible. This desideratum may be unnecessary when centralized valuation is feasible.
• D9 Lipschitzness. The value of ϕk should not change significantly when Dk changes slightly (e.g., slight perturbation to data point) or when A changes slightly (e.g., change in neural network architecture).

We classify the data valuation strategies ϕ into three categories: leave-one-out (LOO), CGT, and desiderata-based according to (the motivation of) the desiderata that ϕ satisfies.

3 D3 addresses cloning of an owner k and his dataset while P8 addresses replication of data points within any dataset, including Dk.

4 [Maschler and Peleg, 1966].

5 [Jia et al., 2019b]. If U(∅) = 0, then this definition is the efficiency assumption in CGT [Chalkiadakis et al., 2011].

3.1 Leave-one-out (LOO) Based

The LOO based strategy evaluates the value of k's dataset Dk as the change in the performance metric of the output model (or dataset) after leaving out Dk from the overall aggregated data DN. Formally, ϕk ≜ v(N) - v(N \ {k}). LOO satisfies D2-D5 and can be adapted to satisfy D1 and D7. To compute the LOO value [Cook and Weisberg, 1980] for every owner k ∈ N, we need to evaluate v(N \ {k}) an additional n times. This is potentially (computationally) expensive due to the training of A(DN \ Dk) in each v(N \ {k}).

Two fundamental drawbacks of LOO based strategies are that ϕk may be inaccurate and the violation of D6, especially when multiple owners have similar datasets. To illustrate, consider the k-nearest neighbours algorithm as A and the validation accuracy as U. If owner k has a clone j s.t. Dk = Dj , their LOO values ϕk = ϕj = 0 regardless of the validation set as the predictions do not change [Ghorbani and Zou, 2019]. As another example, consider Bayesian linear regression as A and IG (on the regression parameters) as U. For any owner k, when |DN \ Dk| is large, ϕk ≈ 0 as IG is submodular and the increase in IG due to Dk diminishes with more data [Sim et al., 2020]. These drawbacks of LOO prompt the investigation of more principled ways to leave out m > 1 owners' datasets fromDN instead, as in CGT based strategies.

3.2 Cooperative Game Theory (CGT) Based

Several data valuation works adopt well-known CGT solution concepts including (but not limited to) Shapley value (SV) [Shapley, 1953], least core, and Banzhaf index (BI) as ϕk. In these solution concepts, ϕk explicitly depends on k's marginal contribution (MC) to every coalition (subset of owners) C ⊆ N \ {k} or the value of the coalition C ∪ {k}. The MC of k to C is ∆Ck ≜ v(C ∪{k})-v(C) and the value of a coalition including k is v(C ∪ {k}).

Both SV and BI set ϕk = ∑C⊆N\{k} wC∆Ck . In BI, every coalition C ⊆ N \ {k} has an equal weight wC while in SV, the weight wC is the lowest when |C| = n/2 and increases as |C| tends to 1 or n. SV is the unique solution concept to satisfy D4, D5, D7, and D8 simultaneously. This has led to its popularity in existing works (e.g., [Ghorbani and Zou, 2019; Jia et al., 2019b]). SV and BI assume all owners are committed to forming the grand coalition N for collaboration.

In contrast, (egalitarian) least core sets (ϕk)k∈N to the vector with the least l2-norm that minimises the greatest deficit between the characteristic function value and the total payoff to any coalition C, i.e, maxC v(C) - ∑k∈C ϕk [Chalkiadakis et al., 2011]. It is a group notion of fairness and ensures that each C gets its dues and hence would not deviate from the grand coalition [Yan and Procaccia, 2021].

3.3 Desiderata Based

We will group a few Shapley variants to discuss the desiderata they address, and then describe two other strategies with different motivations.

Data Shapley [Ghorbani and Zou, 2019] values are sensitive to the exact choice of DN and do not satisfy stability

6 The techniques that target ∆Ck can be used for BI and LOO based strategies.

Data-valuation specific with Shapley value as ϕk 6

(D9). When a data owner is added or removed, all the data Shapley values must be recomputed. To satisfy D9, Ghorbani et al. [2020] have proposed Distributional Shapley (DShapley) which depends on the underlying distribution instead of the fixed dataset DN. D-Shapley values are stable under slight perturbations to the data points themselves and the underlying data distribution. As similar data will have similar D-Shapley, Ghorbani et al. [2020] have proposed learning a separate regression model to interpolate and predict D-Shapley for unseen data. As a positive result, the cost of predicting the D-Shapley for unseen data does not depend on |DN| or the learning algorithm A but only on the regression model (D1). Moreover, Ohrimenko et al. [2019] have argued that SV does not satisfy D3 and proposed Robust Shapley to satisfy it, but loses D7. Interestingly, Kwon and Zou [2021] have suggested that SV is suboptimal for quantifying the impact of individual datum and proposed losing D7 to assign larger weights for MC to smaller coalitions C. 7 The proposed Beta Shapley is shown to be effective for noisy data identification in interpretable ML applications.

7 Beta Shapley satisfies all the other desiderata of Data Shapley if the weight of every MC is positive.

Bian et al. [2021] have modelled the probability of any coalition C forming out of the power set 2N as an energybased model, and defined ϕk as the optimal decoupled probability that an owner k is in the formed coalition C. They have then approximated ϕk with the m-step Variational Value which satisfies D2, D4, and D5, but not D7 or D8. Interestingly, they have shown that a one-step gradient ascent for maximizing the mean-field objective can recover SV or BI.

Yoon et al. [2020] (DVRL) have used reinforcement learning to automatically learn data values ϕk and improve A during training. They have simultaneously learned a regression model/data valuation estimator (DVE) to pick the 'valuable' data to train A (i.e., trained over randomly sampled batches of data iteratively). Consequently, they have viewed the probabilities that the DVE assigned to k-th datum as its value ϕk. DVRL is scalable to large datasets and complex models (D1) since the computational cost (for data valuation) does not directly depend on |DN| or that of A. It also satisfies D2 but is not guaranteed to satisfy the other desiderata (D4-D9). A practical limitation is that it is w.r.t. individual datum (or singleton datasets |Dk| = 1) and requires n to be large.

4 Open Challenges

4.1 New Ingredients

In Sec. 2, we have seen how the properties of the data and learning algorithm determine the performance metric and steps in the computation of ϕ. Hence, new data and learning algorithm settings should be examined in future work. A few examples are given as open challenges.

Open Challenges. Firstly, if a data owner k contributes noisy and adversarial data, then we may want him to get a lower performance score and data value. However, this new principle may not be satisfied by the current options for U. It can be practically challenging to obtain a clean, noiseless validation set for P1 based metrics. P6 and P7 based performance metrics may contrarily value noisy and adversarial data highly. Future work may propose new performance metrics U that will satisfy the principle and allow CML applications to remove the assumption for high-quality data used in [Tay et al., 2022; Xu et al., 2021b].

Next, data owners (e.g., patients) may require privacy guarantees on their data. If weak guarantees are acceptable, then future work can consider using synthetic data [Tay et al., 2022]. If differential privacy (DP) is needed, then there are a few issues and questions to consider. Firstly, the need to compute the value of multiple coalitions C ⊂ N in ϕ may increase the privacy leakage. This is observed when we directly use DP gradient descent [Abadi et al., 2016] in each training. Next, as DP would reduce the sensitivity of a model's output to data, how would DP affect the MC and its estimation? Finally, what DP learning algorithm Ap and other performance metrics U can provably guarantee the principle that data/model become less valuable when the privacy requirements are higher?

Future work can also propose performance metrics U for existing principles. For example, for P6 similarity-based metrics, the similarity of model parameters or predictions to a reference model (i.e., trained on everyone's data) can be used. Furthermore, other unexplored learning algorithms A (e.g., multi-agent reinforcement learning and meta-learning) may need new performance metrics U and principles.

4.2 NewData Valuation Strategies, ϕ

In Sec. 3, we suggest that in data valuation applications, D7 and D8 may be unnecessary, and their removal will make more strategies satisfactory.

Open challenges. For CGT based strategies, future work should further explore the research direction of Kwon and Zou [2021] to generalize and unify existing strategies, including SV and BI. To guide data valuation users on the right ϕ to use, future work can outline the differences, advantages, and disadvantages of using different CGT solution concepts and the impact on the relative ϕk values of owners.

Moreover, for lower computational cost (D1) and to satisfy other desiderata, future work can explore the research directions of Ghorbani et al. [2020], Wang et al. [2021], and Yoon et al. [2020] and outline how to learn and predict ϕ on datasets with more than one datum.

4.3 NewCost Reduction Techniques

Achieving low computational cost (satisfying D1) is a practical challenge shared by both the LOO and CGT based strategies. We need to compute ∆N\{k} k or v(N \ {k}) (∆Ck or v(C)) for n and 2n times in total to compute ϕk for LOO and CGT based strategies, respectively. We will first outline existing techniques to reduce the computational cost (i.e., summarized in Table 3) and follow up with some open challenges.

Application-agnostic approximations can be used beyond data valuation and often come with theoretical guarantees. SV approximations can reduce the cost of computing ϕk from exponential to polynomial time in n [Fatima et al., 2008; Jia et al., 2019b; Maleki et al., 2013; Mitchell et al., 2022]. Least core relaxations can be computed with (sub)linear time in n [Yan and Procaccia, 2021]. However, these improvements may be insufficient due to the costly training of A such as training DNN on large datasets. Hence, this motivates the need for data valuation-specific techniques to exploit properties of A and U. The techniques may target two different quantities of interest: (i) the value ϕk of owner k, or (ii) his marginal contribution ∆Ck to any coalition C.

For (i), Jia et al. [2019a] have shown that for k-NN learning algorithms, every owner k's SV can be computed exactly in polynomial time. Separately, Jia et al. [2019b] have proven that for uniformly stable learning algorithms, every datum's SV can be approximated by v(N)/n (with theoretical guarantee) and thus, no retraining is needed. Differently, Ghorbani and Zou [2019] have proven that for Lipschitz-stable UA, any datum's D-Shapley value [Ghorbani et al., 2020] can be estimated using a regression model trained only on a fraction of data and their computed D-Shapley. 8 Fatima et al. [2008] and Rozemberczki and Sarkar [2021] have specifically chosen v(·) to give rise to weighted voting games where both ϕk and ∆Ck can be approximated efficiently with theoretical error guarantees. Specifically, they define MC ∆Ck to be 1 iff k is the swing player (i.e., C makes an incorrect classification but C ∪ {k} makes a correct classification), and 0 otherwise.

8 D-Shapley satisfies stability (D9) due to the formulation of ϕ and may hold for a non-stable learning algorithm A.

On the other hand, for (ii), Jia et al. [2019a] have suggested that for incrementally trainable A, 9 we can exactly compute an MC for each of the n owners in the time taken for one evaluation of A. Thus, we can shave off a linear factor of n further from the permutation sampling approximation. TMCShapley [Ghorbani and Zou, 2019] eliminates the cost of retraining A in the later steps of every permutation in permutation sampling by approximating MC ∆Ck with 0 when the performance metric v(S) evaluated on data from S ⊆ C is close to v(N). Jia et al. [2019b] have suggested using the influence function heuristic [Koh and Liang, 2017] to approximate ∆Ck for singleton datasets (i.e., |Dk| = 1) and save a factor of n from the avoided retraining. Lastly, Gradient Shapley [Ghorbani and Zou, 2019] approximates the model trained on multiple passes of the training data with another trained on a single pass and different hyperparameters. This replaces A with another incrementally trainable A′ .

Open Challenges. The challenges focus on data valuationspecific techniques. For (i), by studying CGT literature or analysing UA, future work can identify new properties of A and U that will enable efficient approximation of ϕk with theoretical error guarantees. Existing solutions for (ii) may be unsuitable for some learning algorithms and models (e.g., complex DNNs) as they are not incrementally trainable and do not have convex loss functions. Basu et al. [2021] have pointed out that the influence function heuristics deteriorate significantly for complex DNNs. Also, Koh et al. [2019] have suggested that influence functions are less accurate at measuring group effects, such as of Dk when its size is large. Moreover, TMC-Shapley and Gradient Shapley do not come with theoretical error guarantees. Hence, improvements and new techniques should be proposed, and their approximation errors should be studied theoretically and empirically.

5 Conclusion

This paper presents a technical survey to provide a formal study of data valuation in ML through its ingredients and the corresponding properties (e.g., performance metric and its design principles). We ground the discussion of common desiderata satisfied by existing data valuation strategies on our proposed ingredients and group existing works into LOO, CGT, and desiderata based strategies. Moreover, we outline open research challenges in data valuation. An interesting direction is to use ML to define ϕ and learn and predict the value of data. This survey serves as a technical guide for researchers in this field and a guideline for practitioners of data valuation.

9A is incrementally trainable if we can obtain A(D ∪ Dk) from A(D) with a significantly lower computation cost than retraining from scratch. An example is logistic regression optimized with gradient descent over one epoch only [Ghorbani and Zou, 2019].

Ethical Statement

This technical survey primarily focuses on the algorithmic aspect of data valuation in machine learning and not the regulatory considerations governing the usage of data (e.g., data privacy issues). Hence, practitioners and researchers should keep in mind and closely observe such regulations since the protection of data privacy (especially personal data) is crucial. Furthermore, as briefly described in Sec. 4.1, one possibility is to adopt differential privacy to provide rigorous privacy guarantees. Therefore, future works are encouraged to incorporate formal treatments of data privacy (not limited to differential privacy) into valuation to ensure the proposed methods can be adopted in practice while observing the regulations.

Acknowledgments

This research is supported by the National Research Foundation, Singapore under its Strategic Capability Research Centres Funding Initiative. Any opinions, findings and conclusions, or recommendations expressed in this material are those of the author(s) and do not reflect the views of the National Research Foundation, Singapore.