<img height="1" width="1" style="display:none" src="https://q.quora.com/_/ad/ad9bc6b5b2de42fb9d7cd993ebb80066/pixel?tag=ViewContent&amp;noscript=1">
Our history

Creating a framework where 1+1>2

In 2011, Loop54’s founders; Mike Odin, a Mathematicial Statistican, and Software Engineer, Joel Kall attempted to apply some of Netflix's award-winning recommendation algorithms to a similar Swedish video streaming service.

They quickly realised that Netflix's Collaborative Filtering algorithm, which requires many users with a lot of interactions that intersect each other, would be out of reach for a Swedish business with much less traffic and a data sparsity problem.

Although it did not work out as intended, the project had ignited an idea: to bring the concept of ‘content discovery’, which the video on-demand industry had been raving about for years, to e-commerce, and to do it without big data. After nearly two years in R&D, the team had successfully built a product discovery algorithm for e-commerce, but instead of a recommender system, they applied the concept to search.

One of their greatest strengths as a team was their completely different backgrounds - Mike from Mathematical Statistics and Joel from Computer Engineering. While the competition approached the problem of on-site search as engineers, by focusing on efficient database indexing and query alteration, Mike and Joel took a more mathematical approach. They modeled relations between products using different techniques ranging from Graph Theory to complex Neural Networks.

In 2012, the duo recruited Robin Mellstrand as CEO and were ready to officially found Loop54.

Their next step was to scale the team and:

  • combine their individual expertise to create an enterprise-level application with state of the art machine learning models.  
  • build an organisation where the mathematicians can work and optimize independently from the backend dev team.
  • build a platform from scatch that supports the more exploratory process of mathematical modeling, machine learning and rapid innovation, while retaining a solid code base.
  • create isolated code in a way that lets the mathematicians plug and play algorithms and models into the system without affecting anything else. 

The evolution of our algorithms

 

Algorithm version Context model Description

Topic Modeling.png

Topic Model
(2012)

Topic Modelling 

Determined underlying structure by clustering words that are used in the same context. Created initial ability to find Related Results.

 Batman New.pngBatman
(Q2 2013)

Behaviour Augmented Topic Model and Association Network (Topic Modelling) 

Added the inherent product structure to reinforce topic models. Improved contextual word clustering. Added behaviour-based learning, faceting, initial personalisation.  

 Coffee Bean Logo.pngBean
(Q1 2015)

 

Behaviour Enhanced Association Network (Graph Theory)

Created new ways of representing words/strings. Made it easier to configure and troubleshoot engine. Improved reliability of learning capabilities. Reduced algorithm training time by 90%.

GOLEM.png

Golem
(Q3 2016)

 

Generalized Organisation by Layered Expanding Maps
(Neural Networks)

Creates an alternative representation of the product catalogue structure - analogous to the structure of a well merchandised physical store. Increased conversion rate by 8% compared to Bean.

   Language model  
Language_model.png

Language Model
(Q1 2015)

 Natural Language Processing (NLP)

Gives more efficient data structure and allows us to develop NLP techniques in a more flexible and scalable way. Greatly improved spelling corrections, compounded words, etc.

 

Team.png

Learn more about about us