The performance of a recommendation algorithm varies in different scenarios, and intuitions about what algorithms are best suited to a given scenario can be elusive even to recommender system experts. Furthermore, choosing a single algorithm that performs well across all the scenarios often reduces the effectiveness of the system in specific scenarios. In the domain of scholarly article recommendation, for example, an online evaluation of 33.5M recommendations delivered across multiple applications was performed and it was found that the best performing algorithm in one application (Document embeddings; Click-through rate (CTR): 0.21%) was the worst performing in another (CTR: 0.02%).
Avleen Malhi, a postdoctoral researcher at Aalto University, explored Siamese neural networks (SNN) as a potential solution to this problem during her research visit to Trinity College Dublin, Ireland. Besides contributing to the research field in general, the objective was to improve the recommendations-as-a-service system for digital libraries Mr. DLib.
A Siamese neural network is an artificial neural network, which is typically used for comparing similar instances in different type sets. Siamese neural network selection aims to identify the group of similar algorithms to use for all the instances in a given situation. It does this by calculating the relative absolute error for all algorithms on a particular instance and then ranking the algorithms based on their performance. Finally, relative performance is calculated compared with the best algorithm, which enables the similar performing algorithms to be identified. The key idea is that if an algorithm is the best for a particular instance, the same algorithm can be used for recommendation with similar instances later. Siamese neural networks were applied to algorithm selection on the MovieLens dataset and on scholarly article recommendation. The aim was to use this approach to select the best algorithm for each data instance, and for each recommendation request received.
Avleen’s main collaborator during her 4-week research visit was Joeran Beel, Assistant Professor in Intelligent Systems at the School of Computer Science of Trinity College. Professor Beel is also affiliated with the ADAPT research centre, which is a dynamic research centre in Dublin with experts producing ground-breaking digital content innovations. The visit paved the way for further collaboration and networking within the field of intelligent systems. The next major step will be the submission of a research article in ACM conference RecSys 2020, which is to be held in September next year.