Gorjan Radevski, Kiril Gashteovski, Shahbaz Syed, Christopher Malon, Sebastien Nicolas, Chia-Chien Hung, Timo Sztyler, Verena Heußer, Wiem Ben Rim, Masafumi Enomoto, Kunihiro Takeoka, Masafumi Oyamada, Goran Glavaš, Carolin Lawrence. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Explanations for AI should aid human users, yet this ultimate goal remains under-explored.This paper aims to bridge this gap by investigating the specific explanatory needs of human users in the context of Knowledge Graph Completion (KGC) systems.In contrast to the prevailing approaches that primarily focus on mathematical theories, we recognize the potential limitations of explanations that may end up being overly complex or nonsensical for users.Through in-depth user interviews, we gain valuable insights into the types of KGC explanations users seek.Building upon these insights, we introduce GradPath, 1 a novel path-based explanation method designed to meet humancentric explainability constraints and enhance plausibility.Additionally, GradPath harnesses the gradients of the trained KGC model to maintain a certain level of faithfulness.We verify the effectiveness of GradPath through well-designed human-centric evaluations.The results confirm that our method provides explanations that users consider more plausible than previous ones.

Explaining the predictions of neural black-box models is an important problem, especially when such models are used in applications where user trust is crucial. Estimating the influence of training examples on a learned neural model's behavior allows us to identify training examples most responsible for a given prediction and, therefore, to faithfully explain the output of a black-box model. The most generally applicable existing method is based on influence functions, which scale poorly for larger sample sizes and models. We propose gradient rollback, a general approach for influence estimation, applicable to neural models where each parameter update step during gradient descent touches a smaller number of parameters, even if the overall number of parameters is large. Neural matrix factorization models trained with gradient descent are part of this model class. These models are popular and have found a wide range of applications in industry. Especially knowledge graph embedding methods, which belong to this class, are used extensively. We show that gradient rollback is highly efficient at both training and test time. Moreover, we show theoretically that the difference between gradient rollback's influence approximation and the true influence on a model's behavior is smaller than known bounds on the stability of stochastic gradient descent. This establishes that gradient rollback is robustly estimating example influence. We also conduct experiments which show that gradient rollback provides faithful explanations for knowledge base completion and recommender datasets. An implementation and an appendix are available.

Recognition of activities of daily living (ADLs) is an enabling technology for several ubiquitous computing applications. Most activity recognition systems rely on supervised learning to extract activity models from labeled datasets. A problem with that approach is the acquisition of comprehensive activity datasets, which is an expensive task. The problem is particularly challenging when focusing on complex ADLs characterized by large variability of execution. Moreover, several activity recognition systems are limited to offline recognition, while many applications claim for online activity recognition. In this paper, we propose POLARIS, a framework for unsupervised activity recognition. POLARIS can recognize complex ADLs exploiting the semantics of activities, context data, and sensors. Through ontological reasoning, our algorithm derives semantic correlations among activities and sensor events. By matching observed events with semantic correlations, a statistical reasoner formulates initial hypotheses about the occurred activities. Those hypotheses are refined through probabilistic reasoning, exploiting semantic constraints derived from the ontology. Our system supports online recognition, thanks to a novel segmentation algorithm. Extensive experiments with real-world datasets show that the accuracy of our unsupervised method is comparable to the one of supervised approaches. Moreover, the online version of our system achieves essentially the same accuracy of the offline version.

Due to the emerging popularity of pervasive healthcare applications, tools for monitoring activities in smart homes are gaining momentum. Existing methods mainly rely on supervised learning algorithms for recognizing activities based on sensor data. A key issue with those approaches is the acquisition of comprehensive training sets of activities. Indeed, that task incurs significant costs in terms of manual labeling effort; moreover, labeling by external observers violates the individual's privacy. For these reasons, there is an increasing interest in unsupervised activity recognition methods. A popular approach relies on knowledge-based models expressed by ontologies of activities, environment and sensors. Unfortunately, those models require significant knowledge engineering efforts, and are often limited to a specific application. In this paper, we address the issues of existing methods by proposing a novel hybrid approach. Our intuition is that a generic knowledge-based model of activities can be refined to target specific individuals and environments by collaboratively acquiring feedback from inhabitants. Specifically, we propose a collaborative active learning method to refine correlations among sensor events and activity types that are initially extracted from a high-level ontology. Generic correlations are personalized to each target smart-home considering the similarity between the feedback target and the feedback provider in terms of environment and inhabitant's profiles. Moreover, thanks to this method, new sensors installed in the home are seamlessly integrated in the recognition framework. In order to reduce the burden of providing feedback, we also propose a technique to carefully select the conditions that trigger a feedback request. We conducted experiments with a real-world dataset and a generic ontology of activities. Results show that our hybrid method outperforms state-of-the-art supervised and unsupervised activity recognition techniques while triggering an acceptable number of feedback queries.

Ambient Assisted Living using mobile device sensors is an active area of research in pervasive computing. Multiple approaches have shown that wearable sensors perform very well and distinguish falls reliably from Activities of Daily Living. However, these systems are tested in a controlled environment and are optimized for a given set of sensor types, sensor positions, and subjects. In this work, we propose a self-adaptive pervasive fall detection approach that is robust to the heterogeneity of real life situations. Therefore, we combine sensor data of four publicly available datasets, covering about 100 subjects, 5 devices, and 3 sensor placements. In a comprehensive evaluation, we show that our system is not only robust regarding the different dimensions of heterogeneity, but also adapts autonomously to spontaneous changes in the sensor's position at runtime.

Human activity recognition using wearable devices is an active area of research in pervasive computing. In our work, we address the problem of reducing the effort for training and adapting activity recognition approaches to a specific person. We focus on the problem of cross-subjects based recognition models and introduce an approach that considers physical characteristics. Further, to adapt such a model to the behavior of a new user, we present a personalization approach that relies on online and active machine learning. In this context, we use online random forest as a classifier to continuously adapt the model without keeping the already seen data available and an active learning approach that uses user-feedback for adapting the model while minimizing the effort for the new user. We test our approaches on a real world data set that covers 15 participants, 8 common activities, and 7 different on-body device positions. We show that our cross-subjects based approach performs constantly +3% better than the standard approach. Further, the personalized cross-subjects models, gained through user-feedback, recognize dynamic activities with an F-measure of 87% where the user has significantly less effort than collecting and labeling data.

Recognition of activities of daily living (ADLs) is an enabling technology for several ubiquitous computing applications. In this field, most activity recognition systems rely on supervised learning methods to extract activity models from labeled datasets. An inherent problem of that approach consists in the acquisition of comprehensive activity datasets, which is expensive and may violate individuals' privacy. The problem is particularly challenging when focusing on complex ADLs, which are characterized by large intra- and inter-personal variability of execution. In this paper, we propose an unsupervised method to recognize complex ADLs exploiting the semantics of activities, context data, and sensing devices. Through ontological reasoning, we derive semantic correlations among activities and sensor events. By matching observed sensor events with semantic correlations, a statistical reasoner formulates initial hypotheses about the occurred activities. Those hypotheses are refined through probabilistic reasoning, exploiting semantic constraints derived from the ontology. Extensive experiments with real-world datasets show that the accuracy of our unsupervised method is comparable to the one of state of the art supervised approaches.

Human activity recognition using mobile device sensors is an active area of research in pervasive computing. In our work, we aim at implementing activity recognition approaches that are suitable for real life situations. This paper focuses on the problem of recognizing the on-body position of the mobile device which in a real world setting is not known a priori. We present a new real world data set that has been collected from 15 participants for 8 common activities were they carried 7 wearable devices in different positions. Further, we introduce a device localization method that uses random forest classifiers to predict the device position based on acceleration data. We perform the most complete experiment in on-body device location that includes all relevant device positions for the recognition of a variety of different activities. We show that the method outperforms other approaches achieving an F-Measure of 89% across different positions. We also show that the detection of the device position consistently improves the result of activity recognition for common activities.

We present an approach and mobile application for the interactive exploration and search of geo-located social media entities from different, distributed data providers on the web. When querying the providers, the returned results typically have some overlap. In addition, one has no guarantee that the providers reply within a given time interval. Thus, in order to provide users with geo-located entities in their vicinity in a timely manner, we need to take the asynchronous nature of the data providers' replies into account. Our novel on-the-fly entity resolution engine starts the entity resolution once it retrieves the first responses. It incrementally extends the entity resolution model when more responses arrive. Entities are propagated to the client once the resolution engine has processed them for the first time. Resolution results produced at a later point in time are sent as updates to the client and improve earlier, incomplete results. Our experiments show a matching precision of 95% and scalability of the on-the-fly entity resolution w.r.t. the number of resources being simultaneously processed.

Commercial apps for nearby search on mobile phones such as Qype, AroundMe, Foursquare, or Wikitude have gained huge popularity among smartphone users. Understanding the way how people use and interact with such applications is fundamental for improving the functionality and the user interface design. In our two-step field study, we developed and evaluated mobEx, a mobile app for faceted exploration of social media data on Android phones. mobEx unifies the data sources of related commercial applications in the market by retrieving information from various providers. The goal of our study was to find out, if the subjects understood the metaphor of a time-wheel as novel user interface feature for finding and exploring places and events and how they use it. In addition, mobEx offers a grid-based navigation menu and a list-based navigation menu for exploring the data. Here, we were interested in gaining some qualitative insights about which type of navigation approach the users prefer when they can choose between them. We have collected qualitative user feedback via questionnaires. We also conducted a quantitative user study, where we evaluated user-generated logging data over a period of three weeks with a group of 18 participants. Our results show that the time-wheel can serve as an intuitive way to explore timedependent resources such as events. In addition, it seems that the grid-based navigation approach is the preferable choice when exploring large spaces of faceted data.