arXiv subjects

Benny Kimelfeld

Publications and source records attributed to Benny Kimelfeld.

At least 19 recordsLinked to original sources

arXiv2026

Optimal Repairs for Unary Functional Dependencies: Resolving the Case of Updates

If a table violates its required set of functional dependencies (FDs), what is the minimum number of cell changes needed to restore consistency? This fundamental problem, known as finding an optimal update repair (U-repair), is known to admit polynomial-time algorithms only for a small number of specific FD sets. Whether additional tractable cases exist has remained open. The only established hardness result for this problem is due to Kolahi and Lakshmanan (2009); subsequent attempts to prove hardness for additional cases have failed, leaving these cases unresolved. In this work, we make substantial progress on this open problem by completely resolving the case of unary FDs, in which every FD has a single attribute on its left-hand side. We show that every set of unary FDs either falls into one of the previously known tractable classes or makes the problem of finding an optimal U-repair NP-hard.

cs.DS↗

arXiv2026

Neuro-Relational Programs: Unifying Queries and Neural Computation over Structured Data

The conventional approach to deep learning over relational databases applies neural models, such as Graph Neural Networks (GNNs), to a graph representation of the database. Recent approaches instead operate on databases directly, associating tuples with embeddings and extending query mechanisms to jointly process embeddings and relational content. Inspired by these developments, we introduce Neuro-Relational Programs (NRPs), a declarative query language for relational databases whose facts carry numeric vector embeddings. NRPs extend Datalog-style rules with operations that combine, aggregate, and transform embeddings, thereby interleaving relational reasoning and learnable neural components within a single formalism. This yields a general approach to neural computation over relational data: an NRP can be read both as a query plan with trainable components and as a neural architecture with relational structure built in. Natural syntactic fragments of NRPs recover existing architectures and query formalisms. Zero-ary NRPs correspond to non-adaptive query algorithms; monadic NRPs generalize GNN-style message passing and precisely capture Deep Homomorphism Networks, a connection that we extend to frontier-guarded NRPs over databases with row-ids. We characterize the expressive power of unrestricted NRPs with ReLU-FFN transformations by FOCQ, an extension of first-order logic with counting interpreted over real-weighted structures, yielding a precise connection with uniform TC$^0$ over ordered databases. Together, these results establish NRPs as a broad declarative framework for querying and neural computation over relational data.

cs.DB↗

arXiv2026

Incorporating Deep Learning Design in Database Queries

Deep learning over relational databases is conventionally realized by translating data into graph representations and applying graph-based neural networks within external frameworks. This round-trip between the database and external machine learning (ML) systems introduces non-trivial engineering overhead. In effect, these graph neural networks operate on tuple embeddings and manipulate them in ways that capture the interactions induced by relational joins. Given this natural correspondence, there is no fundamental reason why specifying a neural network over relational data should be substantially harder than querying it. We propose an approach that naturally integrates deep learning with database queries. The key idea is to associate each tuple with provenance, represented as a vector embedding with learnable parameters. Queries are lifted to operate jointly on data and embeddings, mapping input relations with embedded tuples to output relations with embedded tuples. This approach provides a declarative foundation for relational deep learning, facilitating integration with database systems, optimization, and wide adoption. We describe RelaNN, a proof-of-concept implementation of this approach built on top of PyTorch and cuDF. We illustrate the utility of RelaNN by implementing various graph-learning models, including graph convolutional networks, heterogeneous graph transformers, hypergraph neural networks and deep homomorphism networks. The simplicity of the programs and their competitive runtime performance demonstrate a concrete path toward making the implementation of state-of-the-art neural networks over databases as simple as writing a query.

cs.DB↗

arXiv2026

Expressive Power of Deep Homomorphism Networks over Relational Databases

The expressive limitations of message-passing Graph Neural Networks (GNNs) have motivated a wide range of more powerful graph learning architectures. We advocate Deep Homomorphism Networks (DHNs) as a model particularly well-suited for learning over relational databases, due to their close connection to important fragments of SQL such as conjunctive queries. We study the precise expressive power of DHNs by relating them to various natural fragments and extensions of first-order logic (FO). For DHNs with max, sum, and mean aggregations, we establish connections to the unary negation fragment (UNFO) and to the extensions of UNFO with counting quantifiers and with ratio quantifiers. We further relate sum-aggregation DHNs to the unary quantifier alternation fragment of FO and to an extension of FO with expressive counting. Through the classical correspondence between FO and SQL, these results also illuminate the relation between DHNs and SQL. They also enable us to study the decidability of two fundamental static analysis problems for DHNs, the emptiness problem and the subsumption problem. Finally, we confirm through experiments that the established differences in expressive power are reflected in the performance on suitable prediction tasks.

cs.DB↗

arXiv2026

Direct Access for Answers to Conjunctive Queries with Aggregation

We study the fine-grained complexity of conjunctive queries with grouping and aggregation. For common aggregate functions (e.g., min, max, count, sum), such a query can be phrased as an ordinary conjunctive query over a database annotated with a suitable commutative semiring. We investigate the ability to evaluate such queries by constructing in loglinear time a data structure that provides logarithmic-time direct access to the answers ordered by a given lexicographic order. This task is nontrivial since the number of answers might be larger than loglinear in the size of the input, so the data structure needs to provide a compact representation of the space of answers. In the absence of aggregation and annotation, past research established a sufficient tractability condition on queries and orders. For queries without self-joins, this condition is not just sufficient, but also necessary (under conventional lower-bound assumptions in fine-grained complexity). We show that all past results continue to hold for annotated databases, assuming that the annotation itself does not participate in the lexicographic order. Yet, past algorithms do not apply to the count-distinct aggregation, which has no efficient representation as a commutative semiring; for this aggregation, we establish the corresponding tractability condition. We then show how the complexity of the problem changes when we include the aggregate and annotation value in the order. We also study the impact of having all relations but one annotated by the multiplicative identity (one), as happens when we translate aggregate queries into semiring annotations, and having a semiring with an idempotent addition, such as the case of min, max, and count-distinct over a logarithmic-size domain.

cs.DB↗

arXiv2026

Database Views as Explanations for Relational Deep Learning

In recent years, there has been significant progress in the development of deep learning models over relational databases, including architectures based on heterogeneous graph neural networks (hetero-GNNs) and heterogeneous graph transformers. In effect, such architectures state how the database records and links (e.g., foreign-key references) translate into a large, complex numerical expression, involving numerous learnable parameters. This complexity makes it hard to explain, in human-understandable terms, how a model uses the available data to arrive at a given prediction. We present a novel framework for explaining machine-learning models over relational databases, where explanations are view definitions that highlight focused parts of the database that mostly contribute to the model's prediction. We establish such global abductive explanations by adapting the classic notion of determinacy by Nash, Segoufin, and Vianu (2010). In addition to tuning the tradeoff between determinacy and conciseness, the framework allows controlling the level of granularity by adopting different fragments of view definitions, such as ones highlighting whole columns, foreign keys between tables, relevant groups of tuples, and so on. We investigate the realization of the framework in the case of hetero-GNNs, and develop a model-specific approach via the notion of learnable masks. For comparison, we propose model-agnostic heuristic baselines and show that our approach is both more efficient and achieves better explanation quality in most cases. Our extensive empirical evaluation on the RelBench collection across diverse domains and record-level tasks demonstrates both the usefulness of our explanations and the efficiency of their generation.

cs.DB↗

arXiv2026

The Importance of Parameters in Ranking Functions

How important is the weight of a given column in determining the ranking of tuples in a table? To address such an explanation question about a ranking function, we investigate the computation of SHAP scores for column weights, adopting a recent framework by Grohe et al.[ICDT'24]. The exact definition of this score depends on three key components: (1) the ranking function in use, (2) an effect function that quantifies the impact of using alternative weights on the ranking, and (3) an underlying weight distribution. We analyze the computational complexity of different instantiations of this framework for a range of fundamental ranking and effect functions, focusing on probabilistically independent finite distributions for individual columns. For the ranking functions, we examine lexicographic orders and score-based orders defined by the summation, minimum, and maximum functions. For the effect functions, we consider global, top-k, and local perspectives: global measures quantify the divergence between the perturbed and original rankings, top-k measures inspect the change in the set of top-k answers, and local measures capture the impact on an individual tuple of interest. Although all cases admit an additive fully polynomial-time randomized approximation scheme (FPRAS), we establish the complexity of exact computation, identifying which cases are solvable in polynomial time and which are #P-hard. We further show that all complexity results, lower bounds and upper bounds, extend to a related task of computing the Shapley value of whole columns (regardless of their weight).

cs.DB↗

arXiv2025

Analyzing Deviations from Monotonic Trends through Database Repair

Datasets often exhibit violations of expected monotonic trends - for example, higher education level correlating with higher average salary, newer homes being more expensive, or diabetes prevalence increasing with age. We address the problem of quantifying how far a dataset deviates from such trends. To this end, we introduce Aggregate Order Dependencies (AODs), an aggregation-centric extension of the previously studied order dependencies. An AOD specifies that the aggregated value of a target attribute (e.g., mean salary) should monotonically increase or decrease with the grouping attribute (e.g., education level). We formulate the AOD repair problem as finding the smallest set of tuples to delete from a table so that the given AOD is satisfied. We analyze the computational complexity of this problem and propose a general algorithmic template for solving it. We instantiate the template for common aggregation functions, introduce optimization techniques that substantially improve the runtime of the template instances, and develop efficient heuristic alternatives. Our experimental study, carried out on both real-world and synthetic datasets, demonstrates the practical efficiency of the algorithms and provides insight into the performance of the heuristics. We also present case studies that uncover and explain unexpected AOD violations using our framework.

cs.DB↗

arXiv2025

Tractability Frontiers of the Shapley Value for Aggregate Conjunctive Queries

In recent years, the Shapley value has emerged as a general game-theoretic measure for assessing the contribution of a tuple to the result of a database query. We study the complexity of calculating the Shapley value of a tuple for an aggregate conjunctive query, which applies an aggregation function to the result of a conjunctive query (CQ) based on a value function that assigns a number to each query answer. Prior work by Livshits, Bertossi, Kimelfeld, and Sebag (2020) established that this task is #P-hard for every nontrivial aggregation function when the query is non-hierarchical with respect to its existential variables, assuming the absence of self-joins. They further showed that this condition precisely characterizes the class of intractable CQs when the aggregate function is sum or count. In addition, they posed as open problems the complexity of other common aggregate functions such as min, max, count-distinct, average, and quantile (including median). Towards the resolution of these problems, we identify for each aggregate function a class of hierarchical CQs where the Shapley value is tractable with every value function, as long as it is local (i.e., determined by the tuples of one relation). We further show that each such class is maximal: for every CQ outside of this class, there is a local (easy-to-compute) value function that makes the Shapley value #P-hard. Interestingly, our results reveal that each aggregate function corresponds to a different generalization of the class of hierarchical CQs from Boolean to non-Boolean queries. In particular, max, min, and count-distinct match the class of CQs that are all-hierarchical (i.e., hierarchical with respect to all variables), and average and quantile match the narrower class of q-hierarchical CQs introduced by Berkholz, Keppeler, and Schweikardt (2017) in the context of the fine-grained complexity of query answering.

cs.DB↗

arXiv2025

Computing Inconsistency Measures Under Differential Privacy

Assessing data quality is crucial to knowing whether and how to use the data for different purposes. Specifically, given a collection of integrity constraints, various ways have been proposed to quantify the inconsistency of a database. Inconsistency measures are particularly important when we wish to assess the quality of private data without revealing sensitive information. We study the estimation of inconsistency measures for a database protected under Differential Privacy (DP). Such estimation is nontrivial since some measures intrinsically query sensitive information, and the computation of others involves functions on underlying sensitive data. Among five inconsistency measures that have been proposed in recent work, we identify that two are intractable in the DP setting. The major challenge for the other three is high sensitivity: adding or removing one tuple from the dataset may significantly affect the outcome. To mitigate that, we model the dataset using a conflict graph and investigate private graph statistics to estimate these measures. The proposed machinery includes adapting graph-projection techniques with parameter selection optimizations on the conflict graph and a DP variant of approximate vertex cover size. We experimentally show that we can effectively compute DP estimates of the three measures on five real-world datasets with denial constraints, where the density of the conflict graphs highly varies.

cs.DB↗

arXiv2025

Using Database Dependencies to Constrain Approval-Based Committee Voting in the Presence of Context

In Approval-Based Committee (ABC) voting, each voter lists the candidates they approve and then a voting rule aggregates the individual approvals into a committee that represents the collective choice of the voters. An extensively studied class of such rules is the class of ABC scoring rules, where each voter contributes to each possible committee a score based on the voter's approvals. We initiate a study of ABC voting in the presence of constraints about the general context surrounding the candidates. Specifically, we consider a framework in which there is a relational database with information about the candidates together with integrity constraints on the relational database extended with a virtual relation representing the committee. For an ABC scoring rule, the goal is to find a committee of maximum score such that all integrity constraints hold in the extended database. We focus on two well-known types of integrity constraints in relational databases: tuple-generating dependencies (TGDs) and denial constraints (DCs). The former can express, for example, desired representations of groups, while the latter can express conflicts among candidates. ABC voting is known to be computationally hard without integrity constraints, except for the case of approval voting where it is tractable. We show that integrity constraints make the problem NP-hard for approval voting, but we also identify certain tractable cases when key constraints are used. We then present an implementation of the framework via a reduction to Mixed Integer Programming (MIP) that supports arbitrary ABC scoring rules, TGDs and DCs. We devise heuristics for optimizing the resulting MIP, and describe an empirical study that illustrates the effectiveness of the optimized MIP over databases in three different domains.

cs.DB↗

arXiv2024

A Unifying Perspective on Succinct Data Representations

Factorized representations (FRs) are a well-known tool to succinctly represent results of join queries and have been originally defined using the named database perspective. We define FRs in the unnamed database perspective and use them to establish several new connections. First, unnamed FRs can be exponentially more succinct than named FRs, but this difference can be alleviated by imposing a disjointness condition on columns. Conversely, named FRs can also be exponentially more succinct than unnamed FRs. Second, unnamed FRs are the same as (i.e., isomorphic to) context-free grammars for languages in which each word has the same length. This tight connection allows us to transfer a wide range of results on context-free grammars to database factorization; of which we offer a selection in the paper. Third, when we generalize unnamed FRs to arbitrary sets of tuples, they become a generalization of \emph{path multiset representations}, a formalism that was recently introduced to succinctly represent sets of paths in the context of graph database query evaluation.

cs.DB↗

arXiv2024

The Importance of Parameters in Database Queries

We propose and study a framework for quantifying the importance of the choices of parameter values to the result of a query over a database. These parameters occur as constants in logical queries, such as conjunctive queries. In our framework, the importance of a parameter is its SHAP score - a popular instantiation of the game-theoretic Shapley value to measure the importance of feature values in machine learning models. We make the case for the rationale of using this score by explaining the intuition behind SHAP, and by showing that we arrive at this score in two different, apparently opposing, approaches to quantifying the contribution of a parameter. The application SHAP requires two components in addition to the query and the database: (a) a probability distribution over the combinations of parameter values, and (b) a utility function that measures the similarity between the result for the original parameters and the result for hypothetical parameters. The main question addressed in the paper is the complexity of calculating the SHAP score for different distributions and similarity measures. In particular, we devise polynomial-time algorithms for the case of full acyclic conjunctive queries for certain (natural) similarity functions. We extend our results to conjunctive queries with parameterized filters (e.g., inequalities between variables and parameters). We also illustrate the application of our results to "why-not" explanations (aiming to explain the absence of a query answer), where we consider the task of quantifying the contribution of query components to the elimination of a non-answer in consideration. Finally, we discuss a simple approximation technique for the case of correlated parameters.

cs.DB↗

arXiv2024

The Cost of Representation by Subset Repairs

Datasets may include errors, and specifically violations of integrity constraints, for various reasons. Standard techniques for ``minimal-cost'' database repairing resolve these violations by aiming for minimum change in the data, and in the process, may sway representations of different sub-populations. For instance, the repair may end up deleting more females than males, or more tuples from a certain age group or race, due to varying levels of inconsistency in different sub-populations. Such repaired data can mislead consumers when used for analytics, and can lead to biased decisions for downstream machine learning tasks. We study the ``cost of representation'' in subset repairs for functional dependencies. In simple terms, we target the question of how many additional tuples have to be deleted if we want to satisfy not only the integrity constraints but also representation constraints for given sub-populations. We study the complexity of this problem and compare it with the complexity of optimal subset repairs without representations. While the problem is NP-hard in general, we give polynomial-time algorithms for special cases, and efficient heuristics for general cases. We perform a suite of experiments that show the effectiveness of our algorithms in computing or approximating the cost of representation.

cs.DB↗

arXiv2024

Repairing Databases over Metric Spaces with Coincidence Constraints

Datasets often contain values that naturally reside in a metric space: numbers, strings, geographical locations, machine-learned embeddings in a Euclidean space, and so on. We study the computational complexity of repairing inconsistent databases that violate integrity constraints, where the database values belong to an underlying metric space. The goal is to update the database values to retain consistency while minimizing the total distance between the original values and the repaired ones. We consider what we refer to as \emph{coincidence constraints}, which include key constraints, inclusion, foreign keys, and generally any restriction on the relationship between the numbers of cells of different labels (attributes) coinciding in a single value, for a fixed attribute set. We begin by showing that the problem is APX-hard for general metric spaces. We then present an algorithm solving the problem optimally for tree metrics, which generalize both the line metric (i.e., where repaired values are numbers) and the discrete metric (i.e., where we simply count the number of changed values). Combining our algorithm for tree metrics and a classic result on probabilistic tree embeddings, we design a (high probability) logarithmic-ratio approximation for general metrics. We also study the variant of the problem where each individual value's allowed change is limited. In this variant, it is already NP-complete to decide the existence of any legal repair for a general metric, and we present a polynomial-time repairing algorithm for the case of a line metric.

cs.DB↗

arXiv2024

SpannerLib: Embedding Declarative Information Extraction in an Imperative Workflow

Document spanners have been proposed as a formal framework for declarative Information Extraction (IE) from text, following IE products from the industry and academia. Over the past decade, the framework has been studied thoroughly in terms of expressive power, complexity, and the ability to naturally combine text analysis with relational querying. This demonstration presents SpannerLib a library for embedding document spanners in Python code. SpannerLib facilitates the development of IE programs by providing an implementation of Spannerlog (Datalog-based documentspanners) that interacts with the Python code in two directions: rules can be embedded inside Python, and they can invoke custom Python code (e.g., calls to ML-based NLP models) via user-defined functions. The demonstration scenarios showcase IE programs, with increasing levels of complexity, within Jupyter Notebook.

cs.DB↗

arXiv2024

Finding Convincing Views to Endorse a Claim

Recent studies investigated the challenge of assessing the strength of a given claim extracted from a dataset, particularly the claim's potential of being misleading and cherry-picked. We focus on claims that compare answers to an aggregate query posed on a view that selects tuples. The strength of a claim amounts to the question of how likely it is that the view is carefully chosen to support the claim, whereas less careful choices would lead to contradictory claims. We embark on the study of the reverse task that offers a complementary angle in the critical assessment of data-based claims: given a claim, find useful supporting views. The goal of this task is twofold. On the one hand, we aim to assist users in finding significant evidence of phenomena of interest. On the other hand, we wish to provide them with machinery to criticize or counter given claims by extracting evidence of opposing statements. To be effective, the supporting sub-population should be significant and defined by a ``natural'' view. We discuss several measures of naturalness and propose ways of extracting the best views under each measure (and combinations thereof). The main challenge is the computational cost, as naïve search is infeasible. We devise anytime algorithms that deploy two main steps: (1) a preliminary construction of a ranked list of attribute combinations that are assessed using fast-to-compute features, and (2) an efficient search for the actual views based on each attribute combination. We present a thorough experimental study that shows the effectiveness of our algorithms in terms of quality and execution cost. We also present a user study to assess the usefulness of the naturalness measures.

cs.DB↗

arXiv2024

Approval-Based Committee Voting under Incomplete Information

We investigate approval-based committee voting with incomplete information about the approval preferences of voters. We consider several models of incompleteness where each voter partitions the set of candidates into approved, disapproved, and unknown candidates, possibly with ordinal preference constraints among candidates in the latter category. This captures scenarios where voters have not evaluated all candidates and/or it is unknown where voters draw the threshold between approved and disapproved candidates. We study the complexity of some fundamental computational problems for a number of classic approval-based committee voting rules including Proportional Approval Voting and Chamberlin-Courant. These problems include determining whether a given set of candidates is a possible or necessary winning committee and whether a given candidate is possibly or necessarily a member of the winning committee. We also consider proportional representation axioms and the problem of deciding whether a given committee is possibly or necessarily representative.

cs.GT↗