DATE2025.04.07 #Press Releases

Generative Models Using Tensor Networks

Disclaimer: machine translated by DeepL which may contain errors.

-- Correlation structure emerges from stock price rise and fall patterns

Summary of Presentation

Kenji Harada, Assistant Professor at the Graduate School of Informatics, Kyoto University; Tsuyoshi Okubo, Project Associate Professor at The School of Science, UTokyo; and Naoki Kawashima, Professor at The Institute for Solid State Physics (ISSP) , UTokyo, have proposed a novel construction method for generative models based on tensor networks (TN) and demonstrated its effectiveness.

In most cases, generative models are based on neural networks, and research on optimizing network structures has not yet progressed significantly. In this study, the researchers considered a Born machine, which utilizes the correspondence between wavefunctions represented by tree-type TNs and probability distributions. They proposed a generative model construction method, Adaptive Tensor Tree (ATT), that optimizes network structures and verified its effectiveness.

As a concrete example, when constructing a generative model from stock price movement pattern data using ATT, it was observed that, as learning progressed, the correlation between stock names was naturally reflected in the network structure. Since ATT enables the construction of generative models for any sample, it is expected to be useful as a framework for uncovering various correlation structures that were previously difficult to capture and for developing new AI models.

The research results were published online in the international journal " Machine Learning: Science and Technology " on April 1, 2025.

Figure: A tree structure generated by a Born machine based on Adaptive Tensor Tree (ATT) learning about 10 years of S&P 500 price movement patterns. Each node represents a company, with colors indicating different industries. Although industry information was not provided to ATT, it was able to "discover" that companies within the same industry tend to be closely related.

Related Links

The University of Kyoto, The Institute for Solid State Physics (ISSP), UTokyo

Journals

Journal name	Machine Learning: Science and Technology
Title of paper	Tensor tree learns hidden relational structures in data to construct generative models