Treating Cold Start in Product Search by Priors

New products in e-commerce platforms suffer from cold start, both in recommendation and search. In this study, we present experiments to deal with cold start in search by predicting priors for behavioral features in learning to rank set up. The offline results show that our technique generates priors for behavioral features which closely track posterior values. The online A/B test on 140MM queries shows that treatment with priors improves new products impressions and increased customers engagement pointing to their relevance and quality.