Improving precision of e-commerce search results to generate value for customers and business
Jens Kürsten and Arne Vogt • Back to Haystack Europe 2019
View the slides and video of this talk.
This talk will show how we approached the problem of improving the precision of our product search at otto.de, specifically for business relevant queries. We will share our version of product search relevance improvement process. It consists of four phases:
- Query analysis, clustering and data discovery
- Problem discovery and hypothesis generation
- PoC implementation, offline metric evaluation and candidate selection
- Onsite testing
Based on this blueprint, we will be looking into the specifics of the search traffic at otto.de. The main part of this presentation will focus on adapting the product selection process based on query logs and other interaction data. The novelty of this approach is its focus on improving the matching of the results instead of learning how to rank products according to features of queries. The idea has several implications. First, more precise selections will decrease the complexity of the textual scoring function, because the selected items will have many textual features in common. Second, there are business relevant features in almost every e-commerce product ranking scenario that can be employed as ranking feature once matching items are as precise as possible. Third, the homogeneity of the results might be counter-productive in search sessions that are discovery-oriented.
The presentation will detail the general idea of our approach. It is based on the observation that query- or context-specific patterns with specific groups of products appear frequently in our search interaction logs. These products share attributes that can be used during the matching phase of the retrieval process in order to produce homogeneous item selections. Another useful side effect of these precise selections is its impact on facetting. Items in homogeneous product selections will tend to have similar attributes. This will support our users during attribute-based refinement and help them in identifying the products they have in mind when they are discovering our product catalog.”
Jens is the Technical Lead of the search team at OTTO. Before he started to work on e-commerce search in 2014, he graduated from Chemnitz University in Germany with his PhD thesis "A generic approach to component-level evaluation in information retrieval". Besides search core technology - his interests are in the inter- connected domains of natural language processing, computer linguistics, and machine learning
Arne is the Business Designer in OTTOs search team, focusing on natural language processing and search optimization. He has over ten years of experience in e-commerce search.