Graph-enhanced Knowledge Transfer Learning for Fashion Sequential Recommendation

Fashion sequential recommendation enhanced fashion item recommendations by leveraging user preferences from historical trends. In this paper, we propose an approach termed Graph-enhanced Knowledge Transfer Learning (GKTL) to achieve a ‘teacher-student’ alignment in output, by designing random group sampling to learn from teacher knowledge distilled at the prediction level. For this, we design a Similarity-driven Loss Adjustment method to address the bottleneck of the Multilayer Perceptron based model and distinguish contrasting but similar samples. Compared with previous work, GKTL utilizes knowledge from the teacher model, reduces the model complexity, and improves the prediction accuracy, while still maintains the ability to differentiate negative samples. Experimental results show that GKTL outperforms the current state-of-the-art methods on both Amazon-fashion and Alibaba-iFashion sub-datasets.