Probabilistic field mapping for product search

Abstract

Online shopping has shown a rapid growth in the last few years. Robust search systems are arguably fundamental to e-commerce sites. Most importantly, sites should have smart retrieval systems to present optimized results that could best satisfy customers purchase intent. To address the demand for such systems we adapted retrieval approaches based on a generative language modeling framework, representing products as semi-structured documents. We present and experimentally compare three alternative ranking functions which make use of different prior estimates. The first method is static field weighting approach relying on field’s individual performance taking nDCG as an effectiveness measure. Two other methods dynamically assign term-field weights according to the distribution of terms in field’s collection. These retrieval functions infers from user search keywords the most likely matching product property probabilistically. The methods differ as one of them considers a uniform field prior whereas the other utilizes performance based prior. The methods were evaluated in relatively new evaluation methodology that evaluated ranking systems when real customer were doing online shopping at toy webshop ‘regiojatek.hu’ : Living labs. In the experiment the lab present an interleaved result, based on Team draft interleaving, from production site and our experimental rankings to customers. The Lab employ an evaluation metric “outcome” and we applied outcome measure to compare our methods and to interpret our results. Our results show that both term-specific mapping methods outperformed the static weight assignment approach. In addition results also suggest that estimating field mapping priors based on historical clicks does not outperform the setting where the priors are uniformly distributed. Furthermore,we also discovered that a trec-style evaluation carried out deeming historical clicks as relevance indicators had ordered the methods inversely in relation to Living labs. This has possible implication that Living labs evaluation platform are essential in IR tasks.