Improving Cold Start Stereotype-Based Recommendation Using Deep Learning

Recommendation engines constitute a key component of many online platforms. One of the most challenging problems facing a recommender system is that of cold start, namely the recommendation of items from the catalogue to a new unknown user, or the recommendation of newly injected content to existing users. It is established that incorporating metadata describing the item or the user leads to better cold start performance. Multiple independent findings point to the value of pre-processing the metadata to generate a new set of coordinates to aid the underlying recommendation algorithm; one of such pre-processing techniques, stereotyped features, have been shown to improve standard recommendation algorithms. Deep learning and complex neural networks have also been widely utilized in recent years in recommender systems, but their application and performance benchmarking in cold start scenarios is still a matter of ongoing research. This article reports on the application of deep learning neural networks to the stereotype driven framework for addressing cold start in recommender systems. We discuss the performance using a range of metrics, covering accuracy, value content of ranked lists but also serendipity and fairness of recommendations, with the latter becoming an important metric and risk factor for the online platform offering the recommendations. Our findings indicate that a multi-layer neural network substantially improves cold start accuracy performance metrics, despite the recommendations displaying worse fairness and serendipity traits. The work discusses for which metrics/scenarios stereotyping features may still be useful also for the class of more sophisticated deep learning recommender systems.