Conference Agenda (All times are shown in Eastern Daylight Time)

This study investigates how different music discovery approaches influence users’ perceptions of playlist diversity and satisfaction, with a particular focus on both objective and subjective measures of musical diversity. A within-subjects experiment involving 144 Spotify users compared two systems: a user-driven seed-based search and the algorithm-driven Discover Weekly. Objective diversity was calculated through genre, artist, and Spotify audio features, while subjective diversity was measured using participants’ self-reported ratings.

Results from structural equation modeling and t-tests revealed that Discover Weekly consistently produced more diverse playlists across all objective measures. However, only genre and artist diversity significantly influenced users’ perceived diversity; sonic diversity had a limited perceptual impact. Moreover, subjective diversity, rather than objective diversity, showed a stronger association with overall playlist satisfaction. These findings suggest that users’ evaluations are shaped more by perceived categorical diversity than by measurable acoustic variance.

Although familiarity influenced individual track preference, its effect on overall playlist satisfaction was weaker, implying the role of other mediating factors such as novelty or user traits. Overall, our findings reaffirm the need for exploration-oriented measures that go beyond accuracy. Specifically, they underscore the importance of diversity in evaluating playlist-based music recommendations.

The exponential growth of data across diverse domains highlights the need for efficient methods in discovering relevant datasets. Traditional search engines such as Google, have served as the go-to tools for this purpose. Recent advancements in large language models (LLMs) such as ChatGPT and Microsoft Copilot have sparked interest in their potential to serve as alternatives for data discovery. While these models are primarily designed for conversational interactions, their capabilities in information retrieval and dataset discovery are becoming areas of active exploration. In this work, we present a mixed-method study that investigates the difference in user experience when using Google and Microsoft Copilot to search for datasets. This study aims to uncover the strengths and limitations of LLMs in data discovery, offering insights into their potential as alternatives or complements to traditional tools.

During public health events, social media platforms serve as key channels for disseminating multimodal information especially from government departments. These data are crucial for enhancing public understanding and emergency preparedness. This study proposes a novel framework for constructing and learning representations from social heterogeneous information networks (SHINs) based on multimodal fusion and an Enhanced-HGCN (Hyperbolic Graph Convolutional Network) model. Approximately 74,403 flu-related microblog posts were collected, from which multimodal features were extracted to construct a heterogeneous network linking users, posts, and topics. Furthermore, the Enhanced-HGCN model with a two-layer graph convolution structure is proposed to learn node embeddings in the SHINs. Experimental results show that our approach significantly outperforms other baseline models including in clustering performance. This research validates the feasibility of multimodal SHINs construction and the effectiveness of the Enhanced-HGCN, providing a foundation for future applications such as knowledge recommendation and cross-platform information collaboration.