Hokkaido University: residual connections skew AI recommendations

Recommendation AI behind e-commerce and video-streaming "recommended for you" features typically lines up browsing and purchase history in time order and predicts the next item. In the representative models studied, the system combines information from multiple items in the history and bases its prediction on the information corresponding to the last-viewed item. The Hokkaido University group examined how strongly five representative recommendation models depended on that last piece of information, using nine datasets related to e-commerce and music-streaming services offered by companies.

Looking inside the models, the researchers found that residual connections — a mechanism that adds pre-processing information back into processed data and is widely used to stabilize AI training — were involved. Combined with the last-item-based prediction method, residual connections made the influence of the most recent product stand out. When the team weakened the effect of residual connections, the influence of the last-viewed information weakened, but recommendation accuracy declined as well.

Professor Miki Haseyama said the findings could lead to designing and improving recommendation technology that reflects a user's interests across their entire browsing and purchasing history. The work was presented at the international RecSys 2026 conference held in the US in September.