New deep learning framework solves the cold-start problem

Recommender systems suggest potentially relevant content by evaluating user preferences and are essential in reducing information overload. However, when users join a new online platform, recommendation systems often struggle to understand their preferences. With no prior interactions in the new environment, these ‘cold-start’ users are difficult to serve accurately. One promising solution is cross-domain recommendation (CDR), which transfers knowledge about a user’s tastes from one domain to another.

However, many existing cross-domain systems rely heavily on a user’s highly rated items while ignoring low ratings. In fact, dislikes can be just as informative as likes. Low scores on certain topics signal boundaries that help to define the true interests of these users. Yet conventional models tend to treat all feedback as uniformly positive or focus only on high ratings when constructing the bridge between domains.

To address this limitation, Associate Professor Keiko Ono from Doshisha University, Japan, along with Dr. Yusuke Shimizu and Dr. Takuya Futagami from the same university, have proposed a new framework. This framework, Deep User Preference Gating Transfer for CDR (DUPGT-CDR), can extract high- and low-rating interaction vectors from the source domain, generate corresponding transformation vectors, and adaptively fuse them via a gating network. “The research was inspired by the shortcomings of current CDR systems. Although I believe low-rating input is just as important for capturing a person's genuine preferences, traditional models relied mostly on high-rating feedback and largely ignored low-rating information. As a result, they often failed to correctly interpret cases in which a user gave an unusually high rating to an item that would typically be rated low. By depending only on high‑rating items, these models could not accurately capture a user’s true preference structure. To address this limitation, we set out to build a system that enables seamless and personalized cross‑domain recommendations while effectively incorporating both positive and negative feedback. This led us to develop a more flexible and comprehensive CDR method,” mentions Dr. Ono, while talking about the motivation behind this study. The details of their research were published in Volume 14 of the journal IEEE Access on February 16, 2026.

The framework comprises four major steps. The key innovation lies in separating high ratings (4–5) and low ratings (1–3) and encoding them independently. Instead of merging all past interactions into a single representation, the model constructs distinct feature vectors for positive and negative feedback. Once the vectors are generated, a gating network adaptively fuses the heterogeneous signals. Lastly, a personalized bridge function produces a user- and item-specific target-domain embedding.

The researchers evaluated DUPGT-CDR using three cross-domain tasks derived from the Amazon-5cores review dataset, including Book-to-Music, Book-to-Movie, and Movie-to-Music transfers. They compared the new approach with established models, such as EMCDR, PTUPCDR, and MIMNET, across varying proportions of overlapping users.

Across all settings, DUPGT-CDR consistently achieved lower prediction errors. The model reduced the mean absolute error by up to 17.9% and the root mean square error by up to 20.9% compared to existing models. Interestingly, simply adding low-rating data to earlier models often degraded their performance. Without a mechanism to manage conflicting signals, negative feedback can introduce noise. In contrast, the gating strategy in DUPGT-CDR allowed the system to exploit low ratings effectively, improving both convergence speed and final accuracy.

“The technology offers more precise product recommendations in commerce, more engaging content in entertainment, and more personalized learning resources in education. Moreover, its impact extends far beyond these domains. By seamlessly transferring user preferences from one area to another—such as from books to movies or from shopping to music streaming—the system can also enhance a wide range of online experiences, including communication platforms, social interactions, and community-based services. As a result, consumers can enjoy a seamless, consistent, and deeply customized experience across even the most distinct kinds of online services,” explains Dr. Ono.

This study can have a huge influence on the consumer experience of digital services. Over the next few years, this framework can lay the foundation for building highly integrated personalization in commerce, entertainment, and education. Users can see seamless suggestions that take into account their preferences and behaviors across a variety of platforms, minimizing search effort and facilitating more intuitive product, content, and learning opportunity discovery. As personalized systems develop further, this technology can help make daily decisions more effective and build a cohesive, consistently helpful digital environment that is customized to each person's needs.

About Associate Professor Keiko Ono from Doshisha University, Japan
Dr. Keiko Ono is an Associate Professor at the Department of Intelligent Information Engineering and Sciences, Doshisha University, Japan. She received her Ph.D. degree in engineering from the same university in 2007. Her research areas include soft computing, intelligent robotics, and intelligent informatics. She has contributed to more than 90 highly cited research articles to date. She is a member of the Illuminating Engineering Society of North America, the Architectural Institute of Japan, and the Institute of Electrical Engineers of Japan.

Funding information
This work was supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI (grant number: JP21K12097).

Media contact:
Organization for Research Initiatives & Development
Doshisha University
Kyotanabe, Kyoto 610-0394, Japan
E-mail: jt-ura@mail.doshisha.ac.jp

END