Solving the Cold Start Problem in Search Re-ranking with Embedding Stabilization: A Case Study of LINE Part Time Jobs

At Tech-Verse 2026, LY Corporation shared how LINE Part Time Jobs stabilized recommendation embeddings for search reranking.

The team moved away from using raw two-tower embeddings directly and added a postprocessing step with low-rank SVD and orthogonal Procrustes.

Implemented in Spark, the method keeps daily retrained embeddings comparable for real-time ranking.

This reduces cold start issues and helps avoid feature-model version mismatches in production.