This thesis shows how NLP Deep Learning methods, trained on user interactions sequences at H&M website, can be used to model user behavior and create personalized recommendations. We performed multiple experiments to prove how an ordered user history helps the model learn. Both item and user representations proved themselves to play an important role in our model’s performance. New models won at performance but also saw different patterns on recommended items, recommending less popular items, and more expensive than our baseline model.
We believe it is the powerful representation learning and the ability to capture order within sequences that are responsible for the performance improvements. Multiple lines of work have been opened, the usage of cross market data could increase the learned representations as well as helping the model generalize. The removal of user identifiers is an interesting work as well, replacing them with user features instead could remove the cold-start problem and add more information to the model.
The fashion industry presents a new paradigm for recommender systems, high fast-phased trends and user-brand engagement. Companies in fashion do not only have the objective to recommend items users will buy, but also promote a trendy-fashion image of themselves to engage with the user and project the image of leading company in fashion. Fashion companies compete between themselves to create new trends and increase their customer group, metrics promoting diversity and better user engagement like the ones presented in this thesis will help compete in the market.
To sum up, our work has taken the research of companies like Alibaba, Amazon, and Google to the second-biggest worldwide fashion retailer and proved it useful. We have a new niche were our item’s life frequency is much lower than theirs, and it is our duty to continue the work on researching recommender systems capable of learning high fast-phased trends