What is it about?
Online platforms recommend products, videos, and music based on what they think you'll like — but most recommendation systems look at only part of the picture. TriCAM is a model that looks at everything about an item at once: its images, its written description, and how users interact with it. First, it blends the visual and text information using a technique called cross-attention, so the model understands how they relate. Second, it learns what "similar" really means by comparing groups of three items at a time — pulling similar items closer together and pushing different ones apart. Third, it maps each user's interests as a network and predicts what they'll enjoy next. The result is more accurate, more relevant recommendations.
Featured Image
Photo by Growtika on Unsplash
Why is it important?
Most recommendation systems struggle with new or less popular items because there's little user history to learn from. TriCAM handles this well because it can judge an item from its picture and description, not just from clicks. That matters for shoppers, who get better suggestions; for sellers, whose new products actually get discovered; and for platforms, whose catalogues grow every day. As online catalogues keep expanding, recommendation systems that understand content — not just behaviour — are the way forward.
Perspectives
This work began during my M.Eng research at the University of Electronic Science and Technology of China, under the supervision of Prof. Wei Zeng. What drew me to it was a simple frustration: recommendation systems often know what people clicked, but not what things actually are. A product's image and description carry real meaning, and I wanted to build a system that could use that meaning — especially for new items that have no click history yet. TriCAM is the result of that idea, developed with a great team of co-authors. For me, it's also a bridge: the same thinking about how machines learn from messy, multi-source data now shapes my work in cybersecurity, where detecting threats means making sense of many signals at once. I'm proud of this paper, and I hope it helps others building recommendation systems that understand content, not just behaviour.
Aaron Felix Opoku Boateng
University of Guelph
Read the Original
This page is a summary of: Modality Fusion with Triplet Contrastive Learning and Cross-Attention for Multimodal Recommendation, ACM Transactions on Recommender Systems, October 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3856983.
You can read the full text:
Contributors
The following have contributed to this page
