Matching Recommendations based on Siamese Network and Metric Learning

Clothing matching is always making customers in trouble because it is both time-consuming and challenging. To solve the issue of "What outfit matches this piece of shirt?", recommendation tasks have attracted attentions from scholars. Most existing recommendation systems focus on the similarity between items or user's interest in items. A few studies pay attention to matching recommendations which are based on the visual features of items. Identifying and understanding relationships between items is the base of the matching recommendation tasks. A distance between items can be defined to measure the relationships. Therefore, we explore a novel recommendation framework to integrate visual features with category features based on Siamese architecture and metric learning. The distance between items can be learnt in target space (matching space) rather than in the input space (original space). In the matching space, the matching items are close to each other while the mismatching items are far away. Compared with the baselines, our methods show great superiority which demonstrates that the matching space defined in our framework is more suitable for learning the matching relationship between items. The results also show that training in both visual features and category features performs better than that of training in only one of them. The method can be even better if giving more negative samples for an item.