In this paper, we introduce a novel approach to computing the contribution of input tuples to the result of the query, quantified by the Banzhaf and Shapley values. In contrast to prior algorithmic work that focuses on Select-Project-Join-Union queries, ours is the first practical approach for queries with aggregates. It relies on two novel optimizations that are essential for its practicality and significantly improve the runtime performance already for queries without aggregates. The first optimization exploits the observation that many input tuples have the same contribution to the query result, so it is enough to compute the contribution of one of them. The second optimization uses the gradient of the query lineage to compute the contributions of all tuples with the same complexity as for one of them. Experiments with a million instances over 3 databases show that our approach achieves up to 3 orders of magnitude runtime improvements over the state-of-the-art for queries without aggregates, and that it is practical for aggregate queries.

Simple, sufficient explanations furnished by short decision lists can be useful for guiding stakeholder actions. Unfortunately, this transparency can come at the expense of the higher accuracy enjoyed by black box methods, like deep nets. To date, practitioners typically either (i) insist on the simpler model, forsaking accuracy; or (ii) insist on maximizing accuracy, settling for post-hoc explanations of dubious faithfulness. In this paper, we propose a hybrid partially interpretable model that represents a compromise between the two extremes. In our setup, each input is first processed by a decision list that can either execute a decision or abstain, handing off authority to the opaque model. The key to optimizing the decision list is to optimally trade off the accuracy of the composite system against coverage (the fraction of the population that receives explanations). We contribute a new principled algorithm for constructing partially interpretable decision lists, providing theoretical guarantees addressing both interpretability and accuracy. As an instance of our result, we prove that when the optimal decision list has length $k$, coverage $c$, and $b$ mistakes, our algorithm will generate a decision list that has length no greater than $4k$, coverage at least $c/2$, and makes at most $4b$ mistakes. Finally, we empirically validate the effectiveness of the new model.

This work presents CaFA, a system for Cost-aware Feasible Attacks for assessing the robustness of neural tabular classifiers against adversarial examples realizable in the problem space, while minimizing adversaries’ effort. To this end, CaFA leverages TabPGD—an algorithm we set forth to generate adversarial perturbations suitable for tabular data— and incorporates integrity constraints automatically mined by state-of-the-art database methods. After producing adversarial examples in the feature space via TabPGD, CaFA projects them on the mined constraints, leading, in turn, to better attack realizability. We tested CaFA with three datasets and two architectures and found, among others, that the constraints we use are of higher quality (measured via soundness and completeness) than ones employed in prior work. Moreover, CaFA achieves higher feasible success rates—i.e., it generates adversarial examples that are often misclassified while satisfying constraints—than prior attacks while simultaneously perturbing few features with lower magnitudes, thus saving effort and improving inconspicuousness. We open-source CaFA,1hoping it will serve as a generic system enabling machine-learning engineers to assess their models’ robustness against realizable attacks, thus advancing deployed models’ trustworthiness.

Quantifying the contribution of database facts to query answers has been studied as means of explanation. The Banzhaf value, originally developed in Game Theory, is a natural measure of fact contribution, yet its efficient computation for select-project-join-union queries is challenging. In this paper, we introduce three algorithms to compute the Banzhaf value of database facts: an exact algorithm, an anytime deterministic approximation algorithm with relative error guarantees, and an algorithm for ranking and top-k. They have three key building blocks: compilation of query lineage into an equivalent function that allows efficient Banzhaf value computation; dynamic programming computation of the Banzhaf values of variables in a Boolean function using the Banzhaf values for constituent functions; and a mechanism to compute efficiently lower and upper bounds on Banzhaf values for any positive DNF function. We complement the algorithms with a dichotomy for the Banzhaf-based ranking problem: given two facts, deciding whether the Banzhaf value of one is greater than of the other is tractable for hierarchical queries and intractable for non-hierarchical queries. We show experimentally that our algorithms significantly outperform exact and approximate algorithms from prior work, most times up to two orders of magnitude. Our algorithms can also cover challenging problem instances that are beyond reach for prior work.

Tabular embedding methods have become increasingly popular due to their effectiveness in improving the results of various tasks, including classic databases tasks and machine learning predictions. However, most current methods treat these embedding models as "black boxes" making it difficult to understand the insights captured by the models. Our research proposes a novel approach to interpret these models, aiming to provide local and global explanations for the original data and detect potential flaws in the embedding models. The proposed solution is appropriate for every tabular embedding algorithm, as it fits the black box view of the embedding model. Furthermore, we propose methods for comparing different embedding models, which can help identify data biases that might impact the models' credibility without the user's knowledge. Our approach is evaluated on multiple datasets and multiple embeddings, demonstrating that our proposed explanations provide valuable insights into the behavior of tabular embedding methods. By making these models more transparent, we believe our research will contribute to the development of more effective and reliable embedding methods for a wide range of applications.

We present TabEE, Tab ular E mbedding E xplanations, a framework designed to generate explanations for interpreting tabular embedding models. Our framework aims to furnish both local and global explanations for the original data, facilitating the detection of potential flaws in embedding models. TabEE is versatile and compatible with any tabular embedding algorithm, as it adheres to the black box perspective of embedding models. The generated explanations also enable comparisons between multiple embedding models. This demonstration illustrates the effectiveness of TabEE in providing interpretable insights into tabular embedding models, contributing to improved model understanding and credibility assessment.

To explain query results, a recent line of work has proposed to leverage the game-theoretic notion of Shapley values to quantify the contribution of each input fact to each result. Despite significant recent breakthroughs improving the complexity of computing Shapley values in query answering, the computation remains quite costly. To this end, we propose an approach that aims at ranking input facts based on their (hidden) Shapley values. Our method utilizes a repository of queries over the same database for which we do store exact Shapley values. Intuitively, some queries bear similarity in the ways they transform data, and consequently in the contribution of database facts to their outputs. In this manner, given a new query and a query result, we can learn and predict the ranking of contributing facts. Our contributions are three-fold. First, we introduce DBShap, a curated dataset of queries and query results, along with the contributing facts and respective Shapley values. Second, we define the task of predicting the ranking of facts contribution w.r.t a query and query result. Finally, we propose a solution for the prediction task based on BERT.

We study the faithfulness of an explanation system to the underlying prediction model. We show that this can be captured by two properties, consistency and sufficiency, and introduce quantitative measures of the extent to which these hold. Interestingly, these measures depend on the test-time data distribution. For a variety of existing explanation systems, such as anchors, we analytically study these quantities. We also provide estimators and sample complexity bounds for empirically determining the faithfulness of black-box explanation systems. Finally, we experimentally validate the new properties and estimators.

Explaining query results is an essential tool for enhancing the transparency and quality of data processing, and has been extensively studied in recent years. In particular, Data Provenance -- the tracking of transformations that data undergoes in query evaluation -- has been shown to be a key component of explanations. A hurdle that remains is that data provenance itself is often too large and complex to be presented in its entirety. To that end, we propose to leverage novel advancements on quantifying and computing the contributions of individual input tuples to query answers, based on the game-theoretic notion of the Shapley value. Our proposed prototype solution, called ShapGraph, combines the global view of explanations through provenance graphs with a local quantification of contributions through Shapley values. The graphical interface allows users to switch between and combine these two views to obtain a deeper understanding of the most influential parts of the database and how they interact to yield query answers.

The Shapley value is a game-theoretic notion for wealth distribution that is nowadays extensively used to explain complex data-intensive computation, for instance, in network analysis or machine learning. Recent theoretical works show that query evaluation over relational databases fits well in this explanation paradigm. Yet, these works fall short of providing practical solutions to the computational challenge inherent to the Shapley computation. We present in this paper two practically effective solutions for computing Shapley values in query answering. We start by establishing a tight theoretical connection to the extensively studied problem of query evaluation over probabilistic databases, which allows us to obtain a polynomial-time algorithm for the class of queries for which probability computation is tractable. We then propose a first practical solution for computing Shapley values that adopts tools from probabilistic query evaluation. In particular, we capture the dependence of query answers on input database facts using Boolean expressions (data provenance), and then transform it, via Knowledge Compilation, into a particular circuit form for which we devise an algorithm for computing the Shapley values. Our second practical solution is a faster yet inexact approach that transforms the provenance to a Conjunctive Normal Form and uses a heuristic to compute the Shapley values. Our experiments on TPC-H and IMDB demonstrate the practical effectiveness of our solutions.

Data repair, i.e., the identification and fix of errors in the data, is a central component of the Data Science cycle. As such, significant research effort has been devoted to automate the repair process. Yet it still requires significant manual labor by the Data Scientists, tweaking and optimizing repair modules (up to 80% of their time, according to surveys).

Many clustering algorithms lead to cluster assignments that are hard to explain, partially because they depend on all the features of the data in a complicated way. To improve interpretability, we consider using a small decision tree to partition a data set into clusters, so that clusters can be characterized in a straightforward manner. We study this problem from a theoretical viewpoint, measuring cluster quality by the k-means and k-medians objectives. In terms of negative results, we show that popular top-down decision tree algorithms may lead to clusterings with arbitrarily large cost, and any clustering based on a tree with k leaves must incur an Omega(log k) approximation factor compared to the optimal clustering. On the positive side, for two means/medians, we show that a single threshold cut can achieve a constant factor approximation, and we give nearly-matching lower bounds; for general k > 2, we design an efficient algorithm that leads to an O(k) approximation to the optimal k-medians and an O(k^2) approximation to the optimal k-means. Prior to our work, no algorithms were known with provable guarantees independent of dimension and input size.

Data repair is a common and crucial step in many frameworks today, as applications may use data from different sources and of different levels of credibility. Thus, this step has been the focus of many works, proposing diverse approaches. To assist users in understanding the output of such data repair algorithms, we propose T-REx, a system for providing data repair explanations through Shapley values. The system is generic and not specific to a given repair algorithm or approach: it treats the algorithm as a black box. Given a specific table cell selected by the user, T-REx employs Shapley values to explain the significance of each constraint and each table cell in the repair of the cell of interest. T-REx then ranks the constraints and table cells according to their importance in the repair of this cell. This explanation allows users to understand the repair process, as well as to act based on this knowledge, to modify the most influencing constraints or the original database.

Machine Learning models are prevalent in critical human-related decision making, such as resume filtering and loan applications. Refused individuals naturally ask what could change the decision, should they reapply. This question is hard for the model owner to answer: first, the model is typically complex and not easily interpretable; second, models may be updated periodically; and last, attributes of the individual seeking approval are apt to change in time. While each of these challenges have been extensively studied in isolation, their conjunction has not. To this end, we propose a novel framework that allows users to devise a plan of action to individuals in presence of Machine Learning classification, where both the ML model and the user properties are expected to change over time. Our technical solution is currently confined to a particular yet important class of models, namely those of tree-based ensembles (Random Forests, Gradient Boosted trees). In this setting it uniquely combines state-of-the-art solutions for single model interpretation, domain adaptation techniques for predicting future models, and constraint databases to represent and query the space of possible actions. We devise efficient algorithms that leverage these foundations in a novel solution, and experimentally show that they are effective in proposing useful and actionable steps leading to the desired classification.

The interpretability of complex Machine Learning models is coming to be a critical social concern, as they are increasingly used in human-related decision-making processes such as resume filtering or loan applications. Individuals receiving an undesired classification are likely to call for an explanation - preferably one that specifies what they should do in order to alter that decision when they reapply in the future. Existing work focuses on a single ML model and a single point in time, whereas in practice, both models and data evolve over time: an explanation for an application rejection in 2018 may be irrelevant in 2019 since in the meantime both the model and the applicant's data can change. To this end, we propose a novel framework that provides users with insights and plans for changing their classification in particular future time points. The solution is based on combining state-of-the-art algorithms for (single) model explanations, ones for predicting future models, and database-style querying of the obtained explanations. We propose to demonstrate the usefulness of our solution in the context of loan applications, and interactively engage the audience in computing and viewing suggestions tailored for applicants based on their unique characteristic.

A main component of many Data Science applications is the invocation of Machine Learning (ML) classifiers. The typical complexity of these classification models makes it difficult to understand the reason for a result, and consequently to assess its trustworthiness and detect errors. We propose a simple generic approach for explaining classifications, by identifying relevant parts of the input whose perturbation would be significant in affecting the classification. In contrast to previous work, our solution makes use of constraints over the data, to guide the search for meaningful explanations in the application domain. Constraints may either be derived from the schema or specified by a domain expert for the purpose of computing explanations. We have implemented the approach for prominent ML models such as Random Forests and Neural Networks. We demonstrate, through examples and experiments, the effectiveness of our solution, and in particular of its novel use of constraints.

Albert Einstein once said "If you can't explain it simply, you don't understand it well enough". This is not only true for physical theories but also highly relevant to the results of data-driven applications. Data science is extensively used in various domains, achieving high performance at different tasks, e.g. question answering, and image recognition. Unfortunately, its results generally remain unexplainable, leaving its users to wonder why a specific result was returned instead of the other. My Ph.D. thesis focuses on providing explanations for data-driven applications. This paper demonstrates approaches for interpretability in two applications: Natural Language Queries, and Machine Learning Classifiers, followed by a discussion of open problems and future work.

Explaining the results of data-intensive computation via provenance has been extensively studied in the literature. We focus here on explaining the output of Machine Learning Classifiers, which are main components of many contemporary Data Science applications. We have developed a simple generic approach for explaining classification results, by looking for constrained perturbations to parts of the input that would have the most significant effect on the classification. Our solution requires white-box access to the model internals and a specification of constraints that define which perturbations are "reasonable" in the application domain; both are typically available to the data scientist.

Natural language (NL) interfaces to databases allow users without technical background to query the database and get the results. Users of such systems may be surprised by the absence of certain expected results. To this end, we propose to demonstrate NLProveNAns, a system that allows non-expert users to view explanations for non-answers of interest. The explanations are shown in an intuitive manner, by highlighting parts of the original NL query that are intuitively "responsible" for the absence of the expected result. Our solution builds upon and combines recent advancements in Natural Language Interfaces to Databases and models for why-not provenance. In particular, the systems can provide explanations in one of two flavors corresponding to two different why-not provenance models: a short explanation based on the frontier picky model, and a detailed explanation based on the why-not polynomial model.

Multiple lines of research have developed Natural Language (NL) interfaces for formulating database queries. We build upon this work, but focus on presenting a highly detailed form of the answers in NL. The answers that we present are importantly based on the provenance of tuples in the query result, detailing not only the results but also their explanations. We develop a novel method for transforming provenance information to NL, by leveraging the original NL query structure. Furthermore, since provenance information is typically large and complex, we present two solutions for its effective presentation as NL text: one that is based on provenance factorization, with novel desiderata relevant to the NL case, and one that is based on summarization. We have implemented our solution in an end-to-end system supporting questions, answers and provenance, all expressed in NL. Our experiments, including a user study, indicate the quality of our solution and its scalability.

We propose to present NLProv: an end-to-end Natural Language (NL) interface for database queries. Previous work has focused on interfaces for specifying NL questions, which are then compiled into queries in a formal language (e.g. SQL). We build upon this work, but focus on presenting a detailed form of the answers in Natural Language. The answers that we present are importantly based on the provenance of tuples in the query result, detailing not only which are the results but also their explanations. We develop a novel method for transforming provenance information to NL, by leveraging the original NL question structure. Furthermore, since provenance information is typically large, we present two solutions for its effective presentation as NL text: one that is based on provenance factorization with novel desiderata relevant to the NL case, and one that is based on summarization.