�鶹Լ��

Recommendations are one way machine learning has a significant influence on our lives. We experience them in most digital products and services, narrowing our near-infinite pool of choice to things that are the most relevant to us. They help us decide what products to buy, what movies to watch, what articles to read, and even whom to date. Recommender systems learn from data on users’ past interactions with items to predict what items will be relevant to each user in the future. For a streaming service, this data could be watch history, users’ age or location, or metadata from the content, such as genre and lead cast.

Naturally, online providers seek “good” recommendations for their users, but what does this mean? To understand how good a recommender is at recommending content, two things need to be measured: accuracy and diversity. First, accuracy measures how well our predictions match what users go on to watch, and it’s the measure we optimise for when designing production recommenders. However, maximising accuracy is not enough in practice. Indeed, diverse recommendations that cover a large portion of the catalogue and offer items of varying popularity are beneficial for everyone: diversity encourages fair promotion of items and serendipitous discovery of new content for users. It also aligns with the �鶹Լ��’s mission and public purposes. Therefore, we aim to achieve the best accuracy at a sufficient diversity level. However, there is often a trade-off between recommending the most popular content (potentially high accuracy) and recommending diverse content. Also, this trade-off becomes very pronounced in the Cold Start case of new users and items.

Indeed, many successful recommender algorithms are not good at recommending content to new users or recommending new items. This is particularly relevant for products like �鶹Լ�� iPlayer, with new content being published every hour of the day and new users constantly joining the service. This means there is a significant opportunity to improve the dynamism of iPlayer recommendations by mixing mathematical theory with some empirical experimentation.

To do this, I designed a novel machine learning algorithm that is both accurate and diverse for new users. I was supervised by Dr Laura Toni and Edoardo Gruppi from UCL, and Dr Ben Clark and Dr Duncan Walker from the �鶹Լ��. We tested our model on a real-world dataset with millions of user interactions on iPlayer, which included a representative proportion of new and nearly new users. This has resulted in higher accuracy and competitive diversity when compared to other successful recommender models from academic literature!

Our initial data analysis focused on the popularity fluctuations of items on the �鶹Լ�� iPlayer dataset over a monthly period. We saw that popularity varies over time depending on multiple factors like the day of the week, time since release, and item metadata descriptors, like genres. One key insight we found is that, under some simplifying assumptions, recommending the most popular content on the day maximises accuracy for new users. Therefore, the key to handling new users is to forecast the popularity of items. This is hard to achieve, but a reasonable estimate is to recommend content that was popular with users who were new the day before.

We can also extend this popularity method to users who have only recently joined the service, to improve their experience. In this generalised model, we recommend to a group of users the items that were most popular on the previous day among users with the same amount of past item interactions. This popularity method outperforms more complex models for very cold users (new and recently joined). However, a significant issue is that this method is not personalised, meaning that it recommends roughly the same items to all new and recently joined users. This would not be a diverse and practical recommender.

To fix this problem, we used item metadata (such as genres) as a complementary measure of similarity between items in the model. This approach, which we term Metadata Infusion, increases the diversity of the popularity model for cold users by over 1300% whilst only reducing accuracy slightly in comparison (17%). This is an acceptable trade-off.

Considering users who watch more items, aggregate item popularity becomes a less accurate predictor of individual preference. As a result, we applied our metadata-infused popularity model to cold users and used EASE^R – an existing highly performant recommender model – for non-cold users. We also boosted the diversity of EASE^R via Metadata Infusion. This combination of metadata-infused popularity and EASE^R is our proposed Bootstrapped Personalised Popularity (B2P) framework. Compared to EASE^R and other highly performant models that we tested, B2P improves accuracy by over 12% in severely cold datasets at a very competitive diversity level. It’s also a novel and rigorous treatment of Cold Start users and items. Finally, the method of forecasting popularity and the wealth of additional metadata that we can incorporate make it highly flexible. This is just the beginning of our exploration of this exciting problem space!