Current interpretability methods focus on explaining a particular model's decision through present input features. Such methods do not inform the user of the sufficient conditions that alter these decisions when they are not desirable. Contrastive explanations circumvent this problem by providing explanations of the form "If the feature $X>x$, the output $Y$ would be different''. While different approaches are developed to find contrasts; these methods do not all deal with mutability and attainability constraints. In this work, we present a novel approach to locally contrast the prediction of any classifier. Our Contrastive Entropy-based explanation method, CEnt, approximates a model locally by a decision tree to compute entropy information of different feature splits. A graph, G, is then built where contrast nodes are found through a one-to-many shortest path search. Contrastive examples are generated from the shortest path to reflect feature splits that alter model decisions while maintaining lower entropy. We perform local sampling on manifold-like distances computed by variational auto-encoders to reflect data density. CEnt is the first non-gradient-based contrastive method generating diverse counterfactuals that do not necessarily exist in the training data while satisfying immutability (ex. race) and semi-immutability (ex. age can only change in an increasing direction). Empirical evaluation on four real-world numerical datasets demonstrates the ability of CEnt in generating counterfactuals that achieve better proximity rates than existing methods without compromising latency, feasibility, and attainability. We further extend CEnt to imagery data to derive visually appealing and useful contrasts between class labels on MNIST and Fashion MNIST datasets. Finally, we show how CEnt can serve as a tool to detect vulnerabilities of textual classifiers.
相關內容
Ethical principles for algorithms are gaining importance as more and more stakeholders are affected by "high-risk" algorithmic decision-making (ADM) systems. Understanding how these systems work enables stakeholders to make informed decisions and to assess the systems' adherence to ethical values. Explanations are a promising way to create understanding, but current explainable artificial intelligence (XAI) research does not always consider theories on how understanding is formed and evaluated. In this work, we aim to contribute to a better understanding of understanding by conducting a qualitative task-based study with 30 participants, including "users" and "affected stakeholders". We use three explanation modalities (textual, dialogue, and interactive) to explain a "high-risk" ADM system to participants and analyse their responses both inductively and deductively, using the "six facets of understanding" framework by Wiggins & McTighe. Our findings indicate that the "six facets" are a fruitful approach to analysing participants' understanding, highlighting processes such as "empathising" and "self-reflecting" as important parts of understanding. We further introduce the "dialogue" modality as a valid alternative to increase participant engagement in ADM explanations. Our analysis further suggests that individuality in understanding affects participants' perceptions of algorithmic fairness, confirming the link between understanding and ADM assessment that previous studies have outlined. We posit that drawing from theories on learning and understanding like the "six facets" and leveraging explanation modalities can guide XAI research to better suit explanations to learning processes of individuals and consequently enable their assessment of ethical values of ADM systems.
Deep neural networks (DNNs) often perform poorly in the presence of domain shift and category shift. How to upcycle DNNs and adapt them to the target task remains an important open problem. Unsupervised Domain Adaptation (UDA), especially recently proposed Source-free Domain Adaptation (SFDA), has become a promising technology to address this issue. Nevertheless, existing SFDA methods require that the source domain and target domain share the same label space, consequently being only applicable to the vanilla closed-set setting. In this paper, we take one step further and explore the Source-free Universal Domain Adaptation (SF-UniDA). The goal is to identify "known" data samples under both domain and category shift, and reject those "unknown" data samples (not present in source classes), with only the knowledge from standard pre-trained source model. To this end, we introduce an innovative global and local clustering learning technique (GLC). Specifically, we design a novel, adaptive one-vs-all global clustering algorithm to achieve the distinction across different target classes and introduce a local k-NN clustering strategy to alleviate negative transfer. We examine the superiority of our GLC on multiple benchmarks with different category shift scenarios, including partial-set, open-set, and open-partial-set DA. Remarkably, in the most challenging open-partial-set DA scenario, GLC outperforms UMAD by 14.8\% on the VisDA benchmark. The code is available at //github.com/ispc-lab/GLC.
Identifying differences between groups is one of the most important knowledge discovery problems. The procedure, also known as contrast sets mining, is applied in a wide range of areas like medicine, industry, or economics. In the paper we present RuleKit-CS, an algorithm for contrast set mining based on separate and conquer - a well established heuristic for decision rule induction. Multiple passes accompanied with an attribute penalization scheme provide contrast sets describing same examples with different attributes, distinguishing presented approach from the standard separate and conquer. The algorithm was also generalized for regression and survival data allowing identification of contrast sets whose label attribute/survival prognosis is consistent with the label/prognosis for the predefined contrast groups. This feature, not provided by the existing approaches, further extends the usability of RuleKit-CS. Experiments on over 130 data sets from various areas and detailed analysis of selected cases confirmed RuleKit-CS to be a useful tool for discovering differences between defined groups. The algorithm was implemented as a part of the RuleKit suite available at GitHub under GNU AGPL 3 licence (//github.com/adaa-polsl/RuleKit). Keywords: contrast sets, separate and conquer, regression, survival
There is a fundamental limitation in the prediction performance that a machine learning model can achieve due to the inevitable uncertainty of the prediction target. In classification problems, this can be characterized by the Bayes error, which is the best achievable error with any classifier. The Bayes error can be used as a criterion to evaluate classifiers with state-of-the-art performance and can be used to detect test set overfitting. We propose a simple and direct Bayes error estimator, where we just take the mean of the labels that show \emph{uncertainty} of the class assignments. Our flexible approach enables us to perform Bayes error estimation even for weakly supervised data. In contrast to others, our method is model-free and even instance-free. Moreover, it has no hyperparameters and gives a more accurate estimate of the Bayes error than several baselines empirically. Experiments using our method suggest that recently proposed deep networks such as the Vision Transformer may have reached, or is about to reach, the Bayes error for benchmark datasets. Finally, we discuss how we can study the inherent difficulty of the acceptance/rejection decision for scientific articles, by estimating the Bayes error of the ICLR papers from 2017 to 2023.
Numerical predictions of quantities of interest measured within physical systems rely on the use of mathematical models that should be validated, or at best, not invalidated. Model validation usually involves the comparison of experimental data (outputs from the system of interest) and model predictions, both obtained at a specific validation scenario. The design of this validation experiment should be directly relevant to the objective of the model, that of predicting a quantity of interest at a prediction scenario. In this paper, we address two specific issues arising when designing validation experiments. The first issue consists in determining an appropriate validation scenario in cases where the prediction scenario cannot be carried out in a controlled environment. The second issue concerns the selection of observations when the quantity of interest cannot be readily observed. The proposed methodology involves the computation of influence matrices that characterize the response surface of given model functionals. Minimization of the distance between influence matrices allow one for selecting a validation experiment most representative of the prediction scenario. We illustrate our approach on two numerical examples. The first example considers the validation of a simple model based on an ordinary differential equation governing an object in free fall to put in evidence the importance of the choice of the validation experiment. The second numerical experiment focuses on the transport of a pollutant and demonstrates the impact that the choice of the quantity of interest has on the validation experiment to be performed.
Long-tailed classification poses a challenge due to its heavy imbalance in class probabilities and tail-sensitivity risks with asymmetric misprediction costs. Recent attempts have used re-balancing loss and ensemble methods, but they are largely heuristic and depend heavily on empirical results, lacking theoretical explanation. Furthermore, existing methods overlook the decision loss, which characterizes different costs associated with tailed classes. This paper presents a general and principled framework from a Bayesian-decision-theory perspective, which unifies existing techniques including re-balancing and ensemble methods, and provides theoretical justifications for their effectiveness. From this perspective, we derive a novel objective based on the integrated risk and a Bayesian deep-ensemble approach to improve the accuracy of all classes, especially the ``tail". Besides, our framework allows for task-adaptive decision loss which provides provably optimal decisions in varying task scenarios, along with the capability to quantify uncertainty. Finally, We conduct comprehensive experiments, including standard classification, tail-sensitive classification with a new False Head Rate metric, calibration, and ablation studies. Our framework significantly improves the current SOTA even on large-scale real-world datasets like ImageNet.
As data-driven methods are deployed in real-world settings, the processes that generate the observed data will often react to the decisions of the learner. For example, a data source may have some incentive for the algorithm to provide a particular label (e.g. approve a bank loan), and manipulate their features accordingly. Work in strategic classification and decision-dependent distributions seeks to characterize the closed-loop behavior of deploying learning algorithms by explicitly considering the effect of the classifier on the underlying data distribution. More recently, works in performative prediction seek to classify the closed-loop behavior by considering general properties of the mapping from classifier to data distribution, rather than an explicit form. Building on this notion, we analyze repeated risk minimization as the perturbed trajectories of the gradient flows of performative risk minimization. We consider the case where there may be multiple local minimizers of performative risk, motivated by situations where the initial conditions may have significant impact on the long-term behavior of the system. We provide sufficient conditions to characterize the region of attraction for the various equilibria in this settings. Additionally, we introduce the notion of performative alignment, which provides a geometric condition on the convergence of repeated risk minimization to performative risk minimizers.
Learning policies via preference-based reward learning is an increasingly popular method for customizing agent behavior, but has been shown anecdotally to be prone to spurious correlations and reward hacking behaviors. While much prior work focuses on causal confusion in reinforcement learning and behavioral cloning, we focus on a systematic study of causal confusion and reward misidentification when learning from preferences. In particular, we perform a series of sensitivity and ablation analyses on several benchmark domains where rewards learned from preferences achieve minimal test error but fail to generalize to out-of-distribution states -- resulting in poor policy performance when optimized. We find that the presence of non-causal distractor features, noise in the stated preferences, and partial state observability can all exacerbate reward misidentification. We also identify a set of methods with which to interpret misidentified learned rewards. In general, we observe that optimizing misidentified rewards drives the policy off the reward's training distribution, resulting in high predicted (learned) rewards but low true rewards. These findings illuminate the susceptibility of preference learning to reward misidentification and causal confusion -- failure to consider even one of many factors can result in unexpected, undesirable behavior.
This paper focuses on the expected difference in borrower's repayment when there is a change in the lender's credit decisions. Classical estimators overlook the confounding effects and hence the estimation error can be magnificent. As such, we propose another approach to construct the estimators such that the error can be greatly reduced. The proposed estimators are shown to be unbiased, consistent, and robust through a combination of theoretical analysis and numerical testing. Moreover, we compare the power of estimating the causal quantities between the classical estimators and the proposed estimators. The comparison is tested across a wide range of models, including linear regression models, tree-based models, and neural network-based models, under different simulated datasets that exhibit different levels of causality, different degrees of nonlinearity, and different distributional properties. Most importantly, we apply our approaches to a large observational dataset provided by a global technology firm that operates in both the e-commerce and the lending business. We find that the relative reduction of estimation error is strikingly substantial if the causal effects are accounted for correctly.
With the rapid increase of large-scale, real-world datasets, it becomes critical to address the problem of long-tailed data distribution (i.e., a few classes account for most of the data, while most classes are under-represented). Existing solutions typically adopt class re-balancing strategies such as re-sampling and re-weighting based on the number of observations for each class. In this work, we argue that as the number of samples increases, the additional benefit of a newly added data point will diminish. We introduce a novel theoretical framework to measure data overlap by associating with each sample a small neighboring region rather than a single point. The effective number of samples is defined as the volume of samples and can be calculated by a simple formula $(1-\beta^{n})/(1-\beta)$, where $n$ is the number of samples and $\beta \in [0,1)$ is a hyperparameter. We design a re-weighting scheme that uses the effective number of samples for each class to re-balance the loss, thereby yielding a class-balanced loss. Comprehensive experiments are conducted on artificially induced long-tailed CIFAR datasets and large-scale datasets including ImageNet and iNaturalist. Our results show that when trained with the proposed class-balanced loss, the network is able to achieve significant performance gains on long-tailed datasets.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191