Item cold-start is a pervasive challenge for collaborative filtering (CF) recommender systems. Existing methods often train cold-start models by mapping auxiliary item content, such as images or text descriptions, into the embedding space of a CF model. However, such approaches can be limited by the fundamental information gap between CF signals and content features. In this work, we propose to avoid this limitation with purely content-based modeling of cold items, i.e. without alignment with CF user or item embeddings. We instead frame cold-start prediction in terms of item-item similarity, training a content encoder to project into a latent space where similarity correlates with user preferences. We define our training objective as a sparse generalization of sampled softmax loss with the α-entmax family of activation functions, which allows for sharper estimation of item relevance by zeroing gradients for uninformative negatives. We then describe how this Sampled Entmax for Cold-start (SEMCo) training regime can be extended via knowledge distillation, and show that it outperforms existing cold-start methods and standard sampled softmax in ranking accuracy. We also discuss the advantages of purely content-based modeling, particularly in terms of equity of item outcomes.

The item cold-start problem poses a fundamental challenge for music recommendation: newly added tracks lack the interaction history that collaborative filtering (CF) requires. Existing approaches often address this problem by learning mappings from content features such as audio, text, and metadata to the CF latent space. However, previous works either omit artist information or treat it as just another input modality, missing the fundamental hierarchy of artists and items. Since most new tracks come from artists with previous history available, we frame cold-start track recommendation as 'semi-cold' by leveraging the rich collaborative signal that exists at the artist level. We show that artist-aware methods can more than double Recall and NDCG compared to content-only baselines, and propose ACARec, an attention-based architecture that generates CF embeddings for new tracks by attending over the artist's existing catalog. We show that our approach has notable advantages in predicting user preferences for new tracks, especially for new artist discovery and more accurate estimation of cold item popularity.

Recent research shows that weakly supervised contrastive pre-training holds significant promise in learning improved representations of musical audio. Several such works use metadata (e.g. artist names or genre tags) or consumption data (e.g. playlists or user listening history) for cross-modal supervision in their representation learning pipeline. However, methodological differences inhibit direct comparison of the results reported in these studies. In this work, we implement six of these contrastive pre-training regimes under a common framework based around playlist data, benchmarking against from-scratch and self-supervised baselines and systematically evaluating performance in downstream music tagging and playlist continuation. Furthermore, we combine existing methods to produce a novel hybrid approach which shows consistently strong performance by leveraging multiple data modes. Finally, we demonstrate that mixup data augmentation, previously only used in the self-supervised scenario, also has significant downstream benefits in weakly supervised music representation learning.

Collaborative filtering (CF) recommender systems struggle with making predictions on unseen, or 'cold', items. Systems designed to address this challenge are often trained with supervision from warm CF models in order to leverage collaborative and content information from the available interaction data. However, since they learn to replicate the behavior of CF methods, cold-start models may therefore also learn to imitate their predictive biases. In this paper, we show that cold-start systems can inherit popularity bias, a common cause of recommender system unfairness arising when CF models overfit to more popular items, thereby maximizing user-oriented accuracy but neglecting rarer items. We demonstrate that cold-start recommenders not only mirror the popularity biases of warm models, but are in fact affected more severely: because they cannot infer popularity from interaction data, they instead attempt to estimate it based solely on content features. This leads to significant over-prediction of certain cold items with similar content to popular warm items, even if their ground truth popularity is very low. Through experiments on three multimedia datasets, we analyze the impact of this behavior on three generative cold-start methods. We then describe a simple post-processing bias mitigation method that, by using embedding magnitude as a proxy for predicted popularity, can produce more balanced recommendations with limited harm to user-oriented cold-start accuracy.