Product Similarity Language Model Trained on Oliveyoung User Behavior Data: Semantic Similarity Modeling Beyond Traditional Attribute-Based Recommendation

The existing one-hot encoding-based similarity calculation method had a limitation: it assigned equal weight to all attributes, causing similarity to be overestimated when low-importance attributes matched.

To address this, Oliveyoung designed new labels using users' exploratory behavior logs. By defining products that multiple users commonly clicked on within the same search keyword as having a similar relationship, they turned users' actual comparison perspectives into data.

For modeling, they adopted the Sentence Transformer architecture along with MultipleNegativesRankingLoss. This design enables efficient training using only similarity relationship data, without the need for separate negative (dissimilar) data.

Based on semantic understanding of text data such as product names and attributes, this model reflects the varying importance of attributes across different categories, enabling more sophisticated similar product recommendations.