Abstract Sensor networks let a farmer keep their eye on multiple locations in an agricultural field simultaneously but can be expensive to install, maintain and analyse. Furthermore, sensors often suffer from gaps in the recording process which leads to missing data points or what are essentially ‘blind spots’ in the network structure. To cater for missing values, effective methods for data imputation are essential. In this paper, we use graphs to impute these missing values within sensor networks using a technique called graph signal processing (GSP) applied to soil moisture recordings. Using this method, we simulate network conditions involving missing sensors or inconsistently collected data. This enables farmers to reliably estimate the sensor readings that would have been obtained, thereby increasing the fault tolerance of their agricultural sensor networks. In this work, we are specifically interested in the relative accuracy of data imputation between several graph construction techniques within the GSP framework, both geometric, i.e., dependent on the geographical coordinates, and data-driven techniques, e.g., correlations between the sensor readings. We evaluated seven graph construction techniques, also comparing with a simple mean imputation baseline, for creating edges. By masking sensor values, we identify how accurately sensor values can be inferred. This is done by gradually masking sensors from the network with 1000 random sensor combinations per mask size and then imputing these “missing” sensors. For our experiments, we make use of the Cook Agronomy Farm (CAF) dataset for GSP imputation that contains soil moisture data recorded with 42 sensors. At almost at every timestamp not even once all moisture sensors recorded the data simultaneously, showcasing the value of correct data imputation in these sparse sensor networks. Our results indicate that data-driven graphs, that connect nodes (e.g., sensors) based on the underlying sensor recordings, tend to capture the relationships between sensors most accurately, where the data-driven Gaussian kernel graph (a signal similarity approach) consistently outperforms other graphs on average with 15% improvement across all experiments. Furthermore, compared to a simple baseline, error reduces between 50 and 70% depending on the underlying data. This suggests that the Gaussian kernel graph can function as a solid enhancement in applying GSP when sensors networks are either prone to faults or sparsely placed. Additional analysis showed that the interplay between graph density, signal smoothness and structural connectivity should be balanced for optimal performance.

With their increase in performance, neural network architectures also become more complex, necessitating explainability. Therefore, many new and improved methods are currently emerging, which often generate so-called saliency maps in order to improve interpretability. Those methods are often evaluated by visual expectations, yet this typically leads towards a confirmation bias. Due to a lack of a general metric for explanation quality, non-accessible ground truth data about the model's reasoning and the large amount of involved assumptions, multiple works claim to find flaws in those methods. However, this often leads to unfair comparison metrics. Additionally, the complexity of most datasets (mostly images or text) is often so high, that approximating all possible explanations is not feasible. For those reasons, this paper introduces a test for saliency map evaluation: proposing controlled experiments based on all possible model reasonings over multiple simple logical datasets. Using the contained logical relationships, we aim to understand how different saliency methods handle information in different class discriminative scenarios (e.g., via complementary and redundant information). By introducing multiple new metrics, we analyse propositional logical patterns towards a non-informative attribution score baseline to find deviations of typical expectations. Our results show that saliency methods can encode classification relevant information into the ordering of saliency scores, resulting in unreliable interpretations.

A prominent type of explanation for neural networks are saliency/attribution-based approaches, which highlight the most relevant inputs. Here, recent works suggest sub-optimality of those methods and emphasize the challenge of evaluation. In this paper, we present a new dataset-relative baseline to assess the minimal expectations on saliency scores, leading towards a new any-order interpretation evaluation framework. Using the Global Coherence Representation (GCR), we propose the SimpleGCR as an implementation of this framework acting as a stable minimal performance baseline. It thus enables a reference point for comparing different explainability metrics. We evaluate our proposed approach by applying a set of current saliency methods on the univariate UCR UEA time series datasets, and demonstrate the sub-optimality of those methods in this context.

Abstract The growing number of applications of machine learning and data mining in many domains—from agriculture to business, education, industrial manufacturing, and medicine—gave rise to new requirements for how to inspect and control the learned models. The research domain of explainable artificial intelligence (XAI) has been newly established with a strong focus on methods being applied post-hoc on black-box models. As an alternative, the use of interpretable machine learning methods has been considered—where the learned models are white-box ones. Black-box models can be characterized as representing implicit knowledge—typically resulting from statistical and neural approaches of machine learning, while white-box models are explicit representations of knowledge—typically resulting from rule-learning approaches. In this introduction to the special issue on ‘Explainable and Interpretable Machine Learning and Data Mining’ we propose to bring together both perspectives, pointing out commonalities and discussing possibilities to integrate them.

In recent years, industrial fault detection has become more data-driven due to advancements in automated data analysis using Deep Learning (DL). These techniques facilitate meaningful feature extraction, e.g., in time series data retrieved from sensors, which is typically of complex nature. This enables effective fault detection and prognostics, which increases efficiency and productivity of industrial equipment. However, the optimal settings for these DL architectures are generally use-case specific.

Contemporary machine learning techniques are capable of extracting complex structure from data in a way that complements or exceeds manual examination, yet, as is welldocumented, many of these techniques suffer from a lack of interpretability. This paper extends previous work on explainable and interpretable machine learning, in particular on the ‘Knowledge-Augmented Clusters (KnAC)’ approach, allowing human users to benefit from uninterpretable ‘black box’ models to extract structure from datasets by clustering and to make this better understandable. One of the key functions of KnAC is to relate expert-annotated clusters to clusters that have been identified by a machine learning method, and then provide a comprehensible explanation, thus clarifying the relationships that KnAC discovered. Our novel contribution in this paper is to examine the usefulness of subgroup discovery as a way to generate comprehensible explanations within KnAC, and to compare this to the existing approach based on the XAI algorithm Anchors through a detailed evaluation. We find that the approach using subgroup discovery performs equally or better in our extensive experimentation testing this on six different datasets.

This paper presents an iterative, hybrid neuro-symbolic approach for anomaly detection and complex fault diagnosis, enabling knowledge-based (symbolic) methods to complement (neural) machine learning methods and vice versa. We demonstrate an instantiation of this novel diagnosis system with applicability in a practically relevant real-world context, specifically the automotive domain. Explainability is indispensable for diagnosis and arises naturally in the system through the specific interplay of neural and symbolic methods. The presented architecture can be considered as a blueprint which is generally transferable to various diagnostic problems and domains.

Transformer models have shown significant advances recently based on the general concept of Attention — to focus on specifically important and relevant parts of the input data. However, methods for enhancing their interpretability and explainability are still lacking. This is the problem which we tackle in this paper, to make Multi-Headed Attention more interpretable and explainable for time series classification. We present a method for constructing global coherence representations from Multi-Headed Attention of Transformer architectures. Accordingly, we present abstraction and interpretation methods, leading to intuitive visualizations of the respective attention patterns. We evaluate our proposed approach and the presented methods on several datasets demonstrating their efficacy.

Abstract Finding communities that are not only relatively densely connected in a graph but that also show similar characteristics based on attribute information has drawn strong attention in the last years. There exists already a remarkable body of work that attempts to find communities in vertex-attributed graphs that are relatively homogeneous with respect to attribute values. Yet, it is scattered through different research fields and most of those publications fail to make the connection. In this paper, we identify important characteristics of the different approaches and place them into three broad categories: those that selectdescriptive attributes, related to clustering approaches, those that enumerateattribute-value combinations, related to pattern mining techniques, and those that identify conditional attribute weights, allowing for post-processing. We point out that the large majority of these techniques treat the same problem in terms of attribute representation, and are therefore interchangeable to a certain degree. In addition, different authors have found very similar algorithmic solutions to their respective problem.

Explainable AI has received a lot of attention over the past decade, with the proposal of many methods explaining black box classifiers such as neural networks. Despite the ubiquity of recommender systems in the digital world, only few researchers have attempted to explain their functioning, whereas it raises e.g., ethical issues. Indeed, recommender systems direct user choices to a large extent and their impact is important as they give access to only a small part of the range of items (e.g., products and/or services), as the submerged part of the iceberg. Consequently, they limit access to other resources. The potentially negative effects of these systems have been pointed out as phenomena like echo chambers and winner-take-all effects, because the internal logic of these systems is to likely enclose the consumer in a "dej́ a vu" loop. Therefore, it is crucial to provide explanations' of such recommender systems and to identify the user data that led the system to make a specific recommendation. This makes it possible to evaluate recommender systems not only regarding their efficiency (i.e., their capability to recommend an item that was actually chosen by the user), but also w.r.t. the diversity, relevance and timeliness of the active data used to make the recommendation. In this paper, we propose a deep analysis of 7 state-of-the-art models learnt on 6 datasets based on the identification of the items or the sequences of items actively used by the models. The proposed method, which is based on subgroup discovery with different pattern languages (i.e., itemsets and sequences), provides interpretable explanations of the recommendations - useful to compare different models and explain the reasons behind the recommendation to the user.

This paper proposes a mixed-initiative feature engineering approach using explicit knowledge captured in a knowledge graph complemented by a novel interactive visualization method. Using the explicitly captured relations and dependencies between concepts and their properties, feature engineering is enabled in a semi-automatic way. Furthermore, the results (and decisions) obtained throughout the process can be utilized for refining the features and the knowledge graph. Analytical requirements can then be conveniently captured for feature engineering -- enabling integrated semantics-driven data analysis and machine learning.

We present a new method for detecting descriptive community patterns capturing exceptional (sequential) link trails. For that, we provide a novel problem formalization: We model sequential data as first-order Markov chain models, mapped to an attributed weighted network represented as a graph. Then, we detect subgraphs (communities) using exceptional model mining techniques: We target subsets of sequential transitions between nodes that are exceptional in that sense that they either conform strongly to a specific reference or show significant deviations, estimated by a quality measure. In particular, such a community is described by a community pattern composed of descriptive features (of the attributed graph) covering the respective community. We present a comprehensive modeling approach and discuss results of a case study analyzing data from two real-world social networks.

This paper focuses on the analysis of socio-spatial data, i. e., user-performance relations at a distributed event. We consider the data as a bimodal network (i. e., model it as a bipartite graph), and investigate its structural characteristics towards a social network. We focus on plans of the participants (expressed by preferences) and their fulfilment, and propose measures for matching preference and reality. We specifically analyse behavioural patterns w.r.t. distinct user and performance groups. We utilise real-world data collected at the Lange Nacht der Musik (Long Night of Music) 2013 in Munich.

Today, many people spend a lot of time online. Their social interactions captured in online social networks are an important part of the overall personal social profile, in addition to interactions taking place offline. This paper investigates whether relations captured by online social networks can be used as a proxy for the relations in offline social networks, such as networks of human face-to-face (F2F) proximity and coauthorship networks. Particularly, the paper focuses on interactions of computer scientists in online settings (homepages, social networks profiles and connections) and offline settings (scientific collaboration, face-to-face communications during the conferences). We focus on quantitative studies and investigate the structural similarities and correlations of the induced networks; in addition, we analyze implications between networks. Finally, we provide a qualitative user analysis to find characteristics of good and bad proxies.

In recent years the importance of user interactions has been recognized in a variety of research contexts. There is a variety of algorithms for modeling these in social graphs; in particular, we distinguish static and dynamic relations. In contrast to static graphs in which the networks do not change over time, the underlying relation is changing frequently in various contexts. This should be reflected by a time dependent social neighborhood of users. In this paper, we present a new and intuitive visualization concept for the histories of user interactions. We derive association rules and visualize these using heatmaps. We demonstrate the impact of the presented approach by several examples utilizing real-world data -- using the well known twitter dump of 2009.

Understanding the process of link creation is rather important for link prediction in social networks. Therefore, this paper analyzes contact structures in networks of face-to-face spatial proximity, and presents new insights on the dynamic and static contact behavior in such real world networks. We focus on face-to-face contact networks collected at different conferences using the social conference guidance system Conferator. Specifically, we investigate the strength of ties and its connection to triadic closures in face-to-face proximity networks. Furthermore, we analyze the predictability of all, new and recurring links at different points of time during the conference. In addition, we consider network dynamics for the prediction of new links.

This paper presents an approach for rapid knowledge capture using subgroup-discovery techniques. The method enables the acquisition of scoring rules - a knowledge representation that is easy to understand and to maintain. Furthermore, the method features an incremental refinement step that can be applied for fine-tuning of the learned relations. We provide a case study demonstrating the applicability of the presented method using a knowledge base from the biological domain.