Jonathon Riley

Personalised recommendations are now an expected service on any e-commerce platform. When inventory of items is deep and sales history for commonly bundled items is available, this task is trivial. However, this is not the case in user-generated inventories, where products are considered unique and thus once purchased cannot be re-purchased. To overcome this hurdle, a bespoke recommendation engine was developed. Given a user's most recent product actions (i.e. product "likes", saved products, messaging history, etc), a recommendation engine was developed. By generating a graph of product and user nodes, linked by these product actions and weighted by their implied level of engagement, a comprehensive list of users with similar interests is generated. Using this list, recommendations can be made based on previously unseen items from similar users' recent history. This recommendation engine was A/B tested and once released into production in full, drove millions in revenue for the client.

In e-commerce platforms with user-generated inventories, like Depop, Etsy and eBay, product descriptions are often open text fields. This provides a route for sellers to "game" the system, by providing misleading descriptions or including unnecessary and in-accurate information, such as brand names that are not relevant to the item being sold, in an attempt to appear in a wider range of searches. To overcome this challenge, a natural language processing (NLP) model was developed to detect, isolate and filter key words so as to improve the search results. By training a custom named entity recognition (NER) model using the Spacy library, a model to detect brand names, sizes, colours, product quality, and, most importantly, spam details (i.e. irrelevant hashtags, etc), was developed. This model was then applied to product descriptions, along with business logic, to prevent sellers from manipulating search results. Through A/B testing, the improved search results were shown to significantly increase click-through rate.