Sho Tsugawa

Yuto Tamura, Sho Tsugawa, Kohei Watabe

Information diffusion on social media platforms is often assumed to occur primarily through explicit social connections, such as follower or friend ties. However, information frequently propagates beyond these observable ties -- through external websites, search engines, or algorithmic recommendations -- creating implicit links. How the presence of implicit links affects the diffusion process remains unclear. In this study, we investigate the characteristics of implicit links on Twitter using four large-scale datasets. Our analysis reveals that users who are farther from the original source in the social network are more likely to engage in diffusion via implicit links. Although implicit links contribute less to the overall diffusion volume than explicit links, they play a distinct role in disseminating content across diverse and topologically distant communities. We further examine the user attributes associated with the formation of implicit links and show that these features are unevenly distributed across the network and exhibit moderate levels of homophily and monophily. Together, these findings demonstrate that implicit links exert a meaningful influence on information diffusion and highlight the importance of incorporating them into models of diffusion and social influence.

Yuzuki Arai, Sho Tsugawa

The emergence of large language models (LLMs) such as ChatGPT and Claude presents new challenges for philosophy of language, particularly regarding the nature of linguistic meaning and representation. While LLMs have traditionally been understood through distributional semantics, this paper explores Robert Brandom's inferential semantics as an alternative foundational framework for understanding these systems. We examine how key features of inferential semantics -- including its anti-representationalist stance, logical expressivism, and quasi-compositional approach -- align with the architectural and functional characteristics of Transformer-based LLMs. Through analysis of the ISA (Inference, Substitution, Anaphora) approach, we demonstrate that LLMs exhibit fundamentally anti-representationalist properties in their processing of language. We further develop a consensus theory of truth appropriate for LLMs, grounded in their interactive and normative dimensions through mechanisms like RLHF. While acknowledging significant tensions between inferentialism's philosophical commitments and LLMs' sub-symbolic processing, this paper argues that inferential semantics provides valuable insights into how LLMs generate meaning without reference to external world representations. Our analysis suggests that LLMs may challenge traditional assumptions in philosophy of language, including strict compositionality and semantic externalism, though further empirical investigation is needed to fully substantiate these theoretical claims.

Takahiro Yokoyama, Yoshiki Sato, Sho Tsugawa et al.

A graph is a very common and powerful data structure used for modeling communication and social networks. Models that generate graphs with arbitrary features are important basic technologies in repeated simulations of networks and prediction of topology changes. Although existing generative models for graphs are useful for providing graphs similar to real-world graphs, graph generation models with tunable features have been less explored in the field. Previously, we have proposed GraphTune, a generative model for graphs that continuously tune specific graph features of generated graphs while maintaining most of the features of a given graph dataset. However, the tuning accuracy of graph features in GraphTune has not been sufficient for practical applications. In this paper, we propose a method to improve the accuracy of GraphTune by adding a new mechanism to feed back errors of graph features of generated graphs and by training them alternately and independently. Experiments on a real-world graph dataset showed that the features in the generated graphs are accurately tuned compared with conventional models.

Kohei Watabe, Shohei Nakazawa, Yoshiki Sato et al.

Generative models for graphs have been actively studied for decades, and they have a wide range of applications. Recently, learning-based graph generation that reproduces real-world graphs has been attracting the attention of many researchers. Although several generative models that utilize modern machine learning technologies have been proposed, conditional generation of general graphs has been less explored in the field. In this paper, we propose a generative model that allows us to tune the value of a global-level structural feature as a condition. Our model, called GraphTune, makes it possible to tune the value of any structural feature of generated graphs using Long Short Term Memory (LSTM) and a Conditional Variational AutoEncoder (CVAE). We performed comparative evaluations of GraphTune and conventional models on a real graph dataset. The evaluations show that GraphTune makes it possible to more clearly tune the value of a global-level structural feature better than conventional models.

Shohei Nakazawa, Yoshiki Sato, Kenji Nakagawa et al.

With the development of graph applications, generative models for graphs have been more crucial. Classically, stochastic models that generate graphs with a pre-defined probability of edges and nodes have been studied. Recently, some models that reproduce the structural features of graphs by learning from actual graph data using machine learning have been studied. However, in these conventional studies based on machine learning, structural features of graphs can be learned from data, but it is not possible to tune features and generate graphs with specific features. In this paper, we propose a generative model that can tune specific features, while learning structural features of a graph from data. With a dataset of graphs with various features generated by a stochastic model, we confirm that our model can generate a graph with specific features.