Recommenders built upon implicit collaborative filtering are typically trained to distinguish between users' positive and negative preferences. When direct observations of the latter are unavailable, negative training data are constructed with sampling techniques. But since items often exhibit clustering in the latent space, existing methods tend to oversample negatives from dense regions, resulting in homogeneous training data and limited model expressiveness. To address these shortcomings, we propose a novel negative sampler with diversity guarantees. To achieve them, our approach first pairs each positive item of a user with one that they have not yet interacted with; this instance, called hard negative, is chosen as the top-scoring item according to the model. Instead of discarding the remaining highly informative items, we store them in a user-specific cache. Next, our diversity-augmented sampler selects a representative subset of negatives from the cache, ensuring its dissimilarity from the corresponding user's hard negatives. Our generator then combines these items with the hard negatives, replacing them to produce more effective (synthetic) negative training data that are informative and diverse. Experiments show that our method consistently leads to superior recommendation quality without sacrificing computational efficiency.

Recommender systems trained on implicit feedback data rely on negative sampling to distinguish positive items from negative items for each user. Since the majority of positive interactions come from a small group of active users, negative samplers are often impacted by data imbalance, leading them to choose more informative negatives for prominent users while providing less useful ones for users who are not so active. This leads to inactive users being further marginalised in the training process, thus receiving inferior recommendations. In this paper, we conduct a comprehensive empirical study demonstrating that state-of-the-art negative sampling strategies provide more accurate recommendations for active users than for inactive users. We also find that increasing the number of negative samples for each positive item improves the average performance, but the benefit is distributed unequally across user groups, with active users experiencing performance gain while inactive users suffering performance degradation. To address this, we propose a group-specific negative sampling strategy that assigns smaller negative ratios to inactive user groups and larger ratios to active groups. Experiments on eight negative samplers show that our approach improves user-side fairness and performance when compared to a uniform global ratio.

Abstract With the introduction of machine learning in high stakes decision-making, ensuring algorithmic fairness has become an increasingly important task. To this end, many mathematical definitions of fairness have been proposed, and a variety of optimisation techniques have been developed, all designed to maximise a given notion of fairness. Fair solutions, however, tend to rely on the quality of training data, and can be highly sensitive to noise. Recent studies have shown that robustness of many such fairness strategies—i.e., their ability to perform well on unseen data—is not a given and requires careful consideration. To address this challenge, we propose robustness ratio , which is a novel criterion to measure the robustness of diverse fairness optimisation strategies. We support our analysis with multiple extensive experiments on five benchmark fairness data sets, using three prominent fairness strategies, in view of four of the most popular definitions of fairness. Our experiments show that while fairness methods that rely on threshold optimisation (post-processing) mostly outperform other techniques, they are acutely sensitive to noise. This is in contrast to two other methods—correlation remover (pre-processing) and exponentiated gradient descent (in-processing)—which become increasingly fairer as the random noise injected into the data becomes larger. Our findings offer a comprehensive overview of fairness strategies that proves invaluable when tasked with choosing the most suitable method for the task at hand. To the best of our knowledge, we are the first to quantitatively evaluate the robustness of fairness optimisation strategies.

Automated decision-making systems are becoming increasingly ubiquitous, which creates an immediate need for their interpretability and explainability. However, it remains unclear whether users know what insights an explanation offers and, more importantly, what information it lacks. To answer this question we conducted an online study with 200 participants, which allowed us to assess explainees’ ability to realise explicated information – i.e., factual insights conveyed by an explanation – and unspecified information – i.e, insights that are not communicated by an explanation – across four representative explanation types: model architecture, decision surface visualisation, counterfactual explainability and feature importance. Our findings uncover that highly comprehensible explanations, e.g., feature importance and decision surface visualisation, are exceptionally susceptible to misinterpretation since users tend to infer spurious information that is outside of the scope of these explanations. Additionally, while the users gauge their confidence accurately with respect to the information explicated by these explanations, they tend to be overconfident when misinterpreting the explanations. Our work demonstrates that human comprehension can be a double-edged sword since highly accessible explanations may convince users of their truthfulness while possibly leading to various misinterpretations at the same time. Machine learning explanations should therefore carefully navigate the complex relation between their full scope and limitations to maximise understanding and curb misinterpretation. • Users appear ignorant of explanations’ limitations and tend to over-generalise factual insights. • Users exhibit overconfidence when they misinterpret explanations, i.e., invent information that is not communicated by explanations. • Highly comprehensible explanations are more likely to be misinterpreted by users. • Easy-to-understand explanations, while highly comprehensible, tend to be misleading.

Abstract Counterfactual explanations are the de facto standard when tasked with interpreting decisions of (opaque) predictive models. Their generation is often subject to technical and domain-specific constraints that aim to maximise their real-life utility. In addition to considering desiderata pertaining to the counterfactual instance itself, guaranteeing existence of a viable path connecting it with the factual data point has recently gained relevance. While current explainability approaches ensure that the steps of such a journey as well as its destination adhere to selected constraints, they neglect the multiplicity of these counterfactual paths. To address this shortcoming we introduce the novel concept of explanatory multiverse that encompasses all the possible counterfactual journeys. We define it using vector spaces, showing how to navigate, reason about and compare the geometry of counterfactual trajectories found within it. To this end, we overview their spatial properties–such as affinity, branching, divergence and possible future convergence–and propose an all-in-one metric, called opportunity potential, to quantify them. Notably, the explanatory process offered by our method grants explainees more agency by allowing them to select counterfactuals not only based on their absolute differences but also according to the properties of their connecting paths. To demonstrate real-life flexibility, benefit and efficacy of explanatory multiverse we propose its graph-based implementation, which we use for qualitative and quantitative evaluation on six tabular and image data sets.

Abstract Algorithmic recourse suggests actions to individuals who have been adversely affected by automated decision-making, helping them to achieve the desired outcome. Knowing the recourse, however, does not guarantee that users can implement it perfectly, either due to environmental variability or personal choices. Recourse generation should thus anticipate its sub-optimal or noisy implementation. While several approaches construct recourse that is robust to small perturbations – e.g., arising due to its noisy implementation – they assume that the entire recourse is implemented in a single step, thus model the noise as one-off and uniform. But these assumptions are unrealistic since recourse often entails multiple sequential steps, which makes it harder to implement and subject to increasing noise. In this work, we consider recourse under plausible noise that adheres to the local data geometry and accumulates at every step of the way. We frame this problem as a Markov Decision Process and demonstrate that such a distribution of plausible noise satisfies the Markov property. We then propose the RObust SEquential (ROSE) recourse generator for tabular data; our method produces a series of steps leading to the desired outcome even when they are implemented imperfectly. Given plausible modelling of sub-optimal human actions and greater recourse robustness to accumulated uncertainty, ROSE provides users with a high chance of success while maintaining low recourse cost. Empirical evaluation shows that our algorithm effectively navigates the inherent trade-off between recourse robustness and cost while ensuring its sparsity and computational efficiency.

Abstract Interpretable representations are the backbone of many explainers that target black-box predictive systems based on artificial intelligence and machine learning algorithms. They translate the low-level data representation necessary for good predictive performance into high-level human-intelligible concepts used to convey the explanatory insights. Notably, the explanation type and its cognitive complexity are directly controlled by the interpretable representation, tweaking which allows to target a particular audience and use case. However, many explainers built upon interpretable representations overlook their merit and fall back on default solutions that often carry implicit assumptions, thereby degrading the explanatory power and reliability of such techniques. To address this problem, we study properties of interpretable representations that encode presence and absence of human-comprehensible concepts. We demonstrate how they are operationalised for tabular, image and text data; discuss their assumptions, strengths and weaknesses; identify their core building blocks; and scrutinise their configuration and parameterisation. In particular, this in-depth analysis allows us to pinpoint their explanatory properties, desiderata and scope for (malicious) manipulation in the context of tabular data where a linear model is used to quantify the influence of interpretable concepts on a black-box prediction. Our findings lead to a range of recommendations for designing trustworthy interpretable representations; specifically, the benefits of class-aware (supervised) discretisation of tabular data, e.g., with decision trees, and sensitivity of image interpretable representations to segmentation granularity and occlusion colour.

Generating counterfactuals to discover hypothetical predictive scenarios is the de facto standard for explaining machine learning models and their predictions. However, building a counterfactual explainer that is time-efficient, scalable, and model-agnostic, in addition to being compatible with continuous and categorical attributes, remains an open challenge. To complicate matters even more, ensuring that the contrastive instances are optimised for feature sparsity, remain close to the explained instance, and are not drawn from outside of the data manifold, is far from trivial. To address this gap we propose BayCon: a novel counterfactual generator based on probabilistic feature sampling and Bayesian optimisation. Such an approach can combine multiple objectives by employing a surrogate model to guide the counterfactual search. We demonstrate the advantages of our method through a collection of experiments based on six real-life datasets representing three regression tasks and three classification tasks.

Work in Counterfactual Explanations tends to focus on the principle of "the closest possible world" that identifies small changes leading to the desired outcome. In this paper we argue that while this approach might initially seem intuitively appealing it exhibits shortcomings not addressed in the current literature. First, a counterfactual example generated by the state-of-the-art systems is not necessarily representative of the underlying data distribution, and may therefore prescribe unachievable goals (e.g., an unsuccessful life insurance applicant with severe disability may be advised to do more sports). Secondly, the counterfactuals may not be based on a "feasible path" between the current state of the subject and the suggested one, making actionable recourse infeasible (e.g., low-skilled unsuccessful mortgage applicants may be told to double their salary, which may be hard without first increasing their skill level). These two shortcomings may render counterfactual explanations impractical and sometimes outright offensive. To address these two major flaws, first of all, we propose a new line of Counterfactual Explanations research aimed at providing actionable and feasible paths to transform a selected instance into one that meets a certain goal. Secondly, we propose FACE: an algorithmically sound way of uncovering these "feasible paths" based on the shortest path distances defined via density-weighted metrics. Our approach generates counterfactuals that are coherent with the underlying data distribution and supported by the "feasible paths" of change, which are achievable and can be tailored to the problem at hand.

Explanations in machine learning come in many forms, but a consensus regarding their desired properties is still emerging. In our work we collect and organise these explainability desiderata and discuss how they can be used to systematically evaluate properties and quality of an explainable system using the case of class-contrastive counterfactual statements. This leads us to propose a novel method for explaining predictions of a decision tree with counterfactuals. We show that our model-specific approach exploits all the theoretical advantages of counterfactual explanations, hence improves decision tree interpretability by decoupling the quality of the interpretation from the depth and width of the tree.

Machine learning -- the part of artificial intelligence aimed at eliciting knowledge from data and automated decision making without explicit instructions -- is making great strides, with new algorithms being invented every day. These algorithms find myriads of applications, but their ubiquity often comes at the expense of limited interpretability, hidden biases and unexpected vulnerabilities. Whenever one of these factors is a priority, the learning algorithm of choice is often a method considered to be inherently interpretable, e.g. logical models such as decision trees. In my research I challenge this assumption and highlight (quite common) cases when the assumed interpretability fails to deliver. To restore interpretability of logical machine learning models (decision trees and their ensembles in particular) I propose to explain them with class-contrastive counterfactual statements, which are a very common type of explanation in human interactions, well-grounded in social science research. To evaluate transparency of such models I collate explainability desiderata that can be used to systematically assess and compare such methods as an addition to user studies. Given contrastive explanations, I investigate their influence on the model's security, in particular gaming and stealing the model. Finally, I evaluate model fairness, where I am interested in choosing the most fair model among all the models with equal performance.

Machine learning models have become pervasive in our everyday life; they decide on important matters influencing our education, employment and judicial system. Many of these predictive systems are commercial products protected by trade secrets, hence their decision-making is opaque. Therefore, in our research we address interpretability and explainability of predictions made by machine learning models. Our work draws heavily on human explanation research in social sciences: contrastive and exemplar explanations provided through a dialogue. This user-centric design, focusing on a lay audience rather than domain experts, applied to machine learning allows explainees to drive the explanation to suit their needs instead of being served a precooked template.

The prevalence of automated decision making, influencing important aspects of our lives -- e.g., school admission, job market, insurance and banking -- has resulted in increasing pressure from society and regulators to make this process more transparent and ensure its explainability, accountability and fairness. We demonstrate a prototype voice-enabled device, called Glass-Box, which users can question to understand automated decisions and identify the underlying model's biases and errors. Our system explains algorithmic predictions with class-contrastive counterfactual statements (e.g., ``Had a number of conditions been different:...the prediction would change...''), which show a difference in a particular scenario that causes an algorithm to ``change its mind''. Such explanations do not require any prior technical knowledge to understand, hence are suitable for a lay audience, who interact with the system in a natural way -- through an interactive dialogue. We demonstrate the capabilities of the device by allowing users to impersonate a loan applicant who can question the system to understand the automated decision that he received.

The SPHERE project is devoted to advancing eHealth in a smart-home context, and supports full-scale sensing and data analysis to enable a generic healthcare service. We describe, from a data-science perspective, our experience of taking the system out of the laboratory into more than thirty homes in Bristol, UK. We describe the infrastructure and processes that had to be developed along the way, describe how we train and deploy Machine Learning systems in this context, and give a realistic appraisal of the state of the deployed systems.

Understanding data, models and predictions is important for machine learning applications. Due to the limitations of our spatial perception and intuition, analysing high-dimensional data is inherently difficult. Furthermore, black-box models achieving high predictive accuracy are widely used, yet the logic behind their predictions is often opaque. Use of textualisation -- a natural language narrative of selected phenomena -- can tackle these shortcomings. When extended with argumentation theory we could envisage machine learning models and predictions arguing persuasively for their choices.