Past Meetup

Recommender systems, product attribute discovery, encoding cyclical variables

This Meetup is past

49 people went

Location image of event venue

Details

Join us on Wednesday 20 June for our first meetup on retail data science with 3 wonderful speakers!

This event is sponsored by Shop Direct (https://www.shopdirect.com/) who kindly provide a fabulous venue in central London (Victoria station), pizzas and refreshments.

*Please let us know your full name and the full name of your guests, as reception needs it to let you in. If you haven't already, you can send a message to Ariadna or Laurie (organisers).*

---

Agenda:

18:30 - doors open
19:00 - introduction
19:15 - first talk
19:45 - second (short) talk
20:00 - break, networking, food and drinks
20:15 - third talk
20:45 - moving to the pub *The Victoria* across the street

---

Talks:

---

"Deep Learning Product Attributes to Boost Product Discovery and Conversion for Very and Littlewoods", George Cushen

Abstract coming soon

George Cushen is a data scientist at Shop Direct. George focusses on computer vision, deep learning, and augmented reality challenges and holds a PhD from the University of Southampton. In his spare time, he enjoys CrossFit and maintaining open source projects, sometimes simultaneously!

https://georgecushen.com/

@GeorgeCushen / https://twitter.com/GeorgeCushen

---

"The importance of feature representation in machine learning models", Jonny Brooks

Using the right features is vital for building effective machine learning models. For many features such as hour of the day many people transform this feature using one hot encoding. However, this is not always the best way to represent time. Hour of the day, like many other features, is cyclical and hence encoding them in a cyclical manner can improve the performance of ML models. In this talk, I will elaborate on examples where representing features in a more informative way can lead to improvements in model predictions and clustering. I will also discuss the generalisation of feature representation in lower cardinality dimensions known as entity embeddings. The most common example being word embeddings which has lead to a revolution in the field of natural language understanding. Finally, I will discuss how embeddings can be learned and used in a retail setting to improve the customer based models

---

"Recommendation systems for fashion in the pre-deep learning era and techniques that might still be useful", Javier Rodriguez Zaurin

A few years ago I joined a start-up working in the fashion space. One of the first problems I had to address was related to finding the adequate shoe images to show on our site. From there, we started a very entertaining journey into computer vision and NLP, mostly solving unsupervised problems. Due to the nature of those problems Deep Learning was not always an option, but moreover, back then, packages like Keras and Tensorflow were not as accessible or easy to use as they are today. During that time we developed a series of solutions that I believe would still be valuable in the industry today. During this presentation I will describe some of the techniques we used with special emphasis on those that I think are still applicable. I will frame most of my talk within the context of building a recommendation system and I will also briefly mention some aspects one should consider when these algorithms need to go into production.

---

Venue: Shop Direct, 111 Buckingham Palace Road, London, SW1W 0DT.
map: https://goo.gl/maps/YS9FjvwScp92