• Recall + Ranking
  • Recall: Candidate Generation
    • Coviewed items
      • A pair of items that have been frequently viewed together in the same browsing session by multiple users
      • Filter out covered from a different item category to ensure similarity.
    • Title similarity
      • Using ElasticSearch indexing scheme (TF/IDF) give higher quality than previous approaches Recommending Similar Items in Large-scale Online MarketplacesRecommending Similar Items in Large-scale Online Marketplaces
        
        Offline process, generates long-term cluster based on product’s title
        
        e.g.
        
        c1 = {nike, air-max, white, gray, running}
        
        Nike white gray 7 m... which used Locality Sensitive Hashing (LSH)
  • Ranking
    • Pointwise learning to rank problem
    • Binary classification,whether purchased (positive class) vs. non-clicked (negative class)
    • Sampling
      • KL divergence to get a quantitative measure of the overlap of the probability distributions for 2 classes
      • The non- clicked / purchased strategy is optimal for class separation. Reflect…
        • Non-clicked show absolutely no user interest
        • Purchased indicate complete user intention
• Baseline = Linear model with manually adjusted weights based on domain exports

Results

• CTR +3.0%
• PTR (Purchase-Through-Rate) +6.6%
• Revenue +6.0%

Reference

Y. M. Brovman et al., “Optimizing Similar Item Recommendations in a Semi-structured Marketplace to Maximize Conversion,” in Proceedings of the 10th ACM Conference on Recommender Systems, Boston Massachusetts USA, Sep. 2016, pp. 199–202. doi: 10.1145/2959100.2959166.

#recommender-system/ecommerce #recommender-system