Positional Encodings (PEs) are essential for injecting structural information into Graph Neural Networks (GNNs), particularly Graph Transformers, yet their empirical impact remains insufficiently understood. We introduce a unified benchmarking framework that decouples PEs from architectural choices, enabling a fair comparison across 8 GNN and Transformer models, 9 PEs, and 10 synthetic and real-world datasets. Across more than 500 model-PE-dataset configurations, we find that commonly used expressiveness proxies, including Weisfeiler-Lehman distinguishability, do not reliably predict downstream performance. In particular, highly expressive PEs frequently fail to improve, and can even degrade performance on real-world tasks. At the same time, we identify several simple and previously overlooked model-PE combinations that match or outperform recent state-of-the-art methods. Our results demonstrate the strong task-dependence of PEs and underscore the need for empirical validation beyond theoretical expressiveness. To support reproducible research, we release an open-source benchmarking framework for evaluating PEs for graph learning tasks.

In lossy image compression, models face the challenge of hallucinating details or generating out-of-distribution samples due to the information bottleneck. Sometimes, hallucinations are necessary to generate in-distribution samples, but the optimal level depends on image content, as humans notice small changes that affect meaning. We propose ConHa, a compression method that dynamically balances hallucination levels based on content. We train a model to predict user preferences on detail and hallucination levels and use this prediction to adjust the perceptual weight in the reconstruction loss. To evaluate our modes performance, we gathered 1,531 comparisons from 40 participants across three bitrates, see Figure 1. Our model outperforms both Hyperprior [1] and HiFiC [2], achieving a middle ground. The performance improvement over HiFiC holds true at all bitrates. When the perceptual loss weight is fixed to the dataset mean, and not predicted for each image, performance drops to near HiFiC levels. This shows the importance of conditioning the perceptual loss weight on the image for optimal performance. ConHa selectively hallucinates details based on content, improving realism and outperforming all baseline models.

We study the evaluation of graph explanation methods. The state of the art to evaluate explanation methods is to first train a GNN, then generate explanations, and finally compare those explanations with the ground truth. We show five pitfalls that sabotage this pipeline because the GNN does not use the ground-truth edges. Thus, the explanation method cannot detect the ground truth. We propose three novel benchmarks: (i) pattern detection, (ii) community detection, and (iii) handling negative evidence and gradient saturation. In a re-evaluation of state-of-the-art explanation methods, we show paths for improving existing methods and highlight further paths for GNN explanation research.

Ever growing music collections ask for novel ways of organization. The traditional browsing of folder hierarchies or search by title and album tends to be insufficient to maintain an overview of a collection of orders of thousands of tracks. Methods based on song similarity offer an alternative to keyword-based search. In this work we propose to use a high-dimensional map of the "world of music" as a data structure for music retrieval and exploration of personal collections. Our approach does not require expensive analysis of audio signals and scales to hundreds of thousands of tracks. The techniques presented in this work can be used in a variety of applications, ranging from automatic DJs to file sharing on mobile devices. As a concrete example, we have developed a Web-application that allows users to visualize and navigate through their music collections and create playlists by specifying trajectories.

The virtual world is comprised of data items related to each other in a variety of contexts. Often such relations can be represented as graphs that evolve over time. Examples include social networks, co-authorship graphs, and the world-wide-web. Attempts to model these graphs have introduced the notions of hierarchies and layers, which correspond to taxonomies of the underlying objects, and reasons for object relations, respectively. In this paper we explore these concepts in the process of mining such naturallygrown networks. Based on two sample graphs, we present some evidence that the current models well fit real world networks and provide concrete applications of these findings. In particular, we show how hierarchies can be used for greedy routing and how separation of layers can be used as a preprocessing step to implement a location estimation application.