Search to Search recommendations (Collaborative Synonym and Spell corrections)

Sadat Anwar and Matthieu Pons • Back to Haystack Europe 2019

We developed an item-item collaborative recommendation algorithm derived from markov chains and co-occurrences in order to take advantage of anonymous sessions data as implicit feedback. This approach can be implemented without external libraries and is very straightforward to understand and reason with. In addition the model only takes a few minutes to train on CPU-only machines for millions of sessions. We would like to present the method, its advantages but also it’s caveats.

Using this approach we imagined a search-query to search-query recommendation system that has numerous potential use cases. From spell-correction to synonym detection.

We used this model to power spell-check and query rewriting on reBuy.de with interesting results.

Sadat holds an MS in Embedded Systems from the University of Stuttgart, and has worked in the field of Search and IR for about 3 years. At reBuy, he helped evolve the search team into a more data-focused team with a special emphasis on relevance tuning and ranking. He also helped build a product recommendation system while at reBuy. Currently, he is at Delivery Hero helping build a central search system that works across different countries and is used by millions of users in numerous languages. Outside of work, he organises the 'Search Technology Meetup' in Berlin and enjoys discussing about NLP, Machine Learning and Photography.

Matthieu Pons holds MS in computer science and network communications. He started working in a French startup developing multi-agent systems for solving industrial problems. Later in 2016 he created a startup delivering hyperlocal news powered by Elasticsearch. Today he’s a backend engineer at reBuy working on the web team as well as developing and maintaining the recommendation system. He is also co-organizing the search technology meet-up in Berlin out of passion for search and fields related to NLP.