2026 Volume 77 Issue 2 Pages 72-84
In recent years, recommendation systems have been widely adopted on e-commerce sites, and similar platforms. When user ratings are available, a common approach is to recommend items in descending order of predicted ratings that have been calculated by a model. In particular, it has been shown that utilizing review texts can improve prediction accuracy in such methods. A representative example is Transformational Neural Networks (TransNet). TransNet leverages the ratings and review texts for items that the user has purchased or viewed in the past in order to predict ratings for items that the user has not yet purchased. In this process, the prediction of the ratings is formulated as a regression problem, and the model is trained using loss functions for continuous values, such as squared error. This approach effectively treats ratings as though they are on an interval scale, with equal spacing between rating levels. However, in reality, the ratings only convey the relative order, and there is not actually consistent spacing, making it more appropriate to handle these as an ordinal scale. Motivated by this, the present research proposes a recommendation model that builds on TransNet while reflecting the characteristics of ratings as an ordinal scale. Specifically, rating predictions are formulated as a multi-class classification problem with inclusive relationships among the classes. This allows the model to flexibly learn the intervals between different rating levels, which is expected to improve prediction accuracy. Finally, experiments using multiple real-world datasets demonstrate the effectiveness of the proposed method.
