Leveraging transfer learning for Outfit Matching in children's fashion, using deep learning
About Babyshop Group
E-commerce is at the heart of Babyshop Group where premium childrens’ fashion is retailed on a multitude of websites (eg. lekmer.se, babyshop.com, alexandalexa.com, melijoe.com). Babyshop Group has had continuous growth the past years, currently selling and shipping internationally with a revenue of more than 1.5 billion SEK. In order to sustain such growth, Babyshop Group will continue to work in a data driven environment, which you hopefully can be a part of! We at Babyshop work with machine learning (ML) on a daily basis, whether it being exploratory data analysis and modelling in a Jupyter Notebook or building ML-pipelines integrating cloud solutions for personalization, demand prediction, recommender systems, learning-to-rank and more. If this is something that you want to develop within and you find the thesis proposal interesting, apply! By the way, at our office by Karlaplan you will find lots of tasty coffee, tea and inspiring individuals.
At Babyshop Group, there are hundreds of thousands of articles for children’s fashion, which makes the matching of outfits by experienced fashion experts extremely time-intensive, if not infeasible, to do by hand. Because of this, a need for a support system in curating stylish outfits is large. In literature, a multitude of previous work has been done in regards to outfit matching, most notably Bettaney Et Al. Here, multiple deep learning models have been trained on tens of thousands of outfits to predict whether a collection of items, where each item consists of an image, text description and categorical features, constitute a good outfit or not, which has proven highly successful. Moreover, other related work where the corresponding datasets are published aswell are Vasileva Et Al, Xu Et Al, Song Et Al and Jagadeesh Et Al However, all of these datasets are of adult fashion, with very little (or none) emphasis on childrens’ fashion.
Recently, transfer learning has proved very successful to transfer knowledge learned within one domain to another, most prominently in regards to object-detection in images, using YoloV3, as well as natural language processing using BERT and GPT. Even within the specific domain of fashion, transfer learning has been used successfully to transfer knowledge between one subdomain and another, see iMaterialist. Because of this advancement in transfer learning, ML-practitioners with access to limited sets of labeled data can successfully leverage these transfer learning frameworks that are trained on immense datasets.
Adult and children’s fashion is, from a human point of view, very similar. However, the extent to which neural networks generalize across the two subdomains is currently unknown. The question then arises, is it possible to combine the pre-trained models from previously mentioned outfit matching models that are trained on almost exclusively adults’ fashion, together with the learnings from transfer-learning frameworks - in order to successfully perform outfit matching within childrens’ fashion?
We are looking one or a pair of two students for this thesis, and for a good fit we believe it to be preferable if you have the following qualifications:
What we offer
What we unfortunately do not offer
Starting, early September till late December / early January (depending on your semester).
Contact: Marcus Svensson, Data Scientist at Babyshop Group firstname.lastname@example.org
Last application date is 31/8, but continuous selection is carried out.
 Joseph Redmon and Ali Farhadi, YOLOv3: An Incremental Improvement
 Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova, BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
 Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever, Improving Language Understanding by Generative Pre-Training