Training recommender systems for next-item recommendation often requires unique embeddings to be learned for each item, which may take up most of the trainable parameters for a model. Shared embeddings, such as using content information, can reduce the number of distinct embeddings to be stored in memory. This allows for a more lightweight model; correspondingly, model complexity can be increased due to having fewer embeddings to store in memory. We show the benefit of using shared content-based features ('semantic IDs') in improving recommendation accuracy and diversity, while reducing model size, for two music recommendation datasets, including an online A/B test on a music streaming service.

Next-item recommender systems are often trained using only positive feedback with randomly-sampled negative feedback. We show the benefits of using real negative feedback both as inputs into the user sequence and also as negative targets for training a next-song recommender system for internet radio. In particular, using explicit negative samples during training helps reduce training time by ∼ 60% while also improving test accuracy by 6%; adding user skips as additional inputs also can considerably increase user coverage alongside improving accuracy. We test the impact of using a large number of random negative samples to capture a ‘harder’ one and find that the test accuracy increases with more randomly-sampled negatives, but only to a point. Too many random negatives leads to false negatives that limits the lift, which is still lower than if using true negative feedback. We also find that the test accuracy is fairly robust with respect to the proportion of different feedback types, and compare the learned embeddings for different feedback types.

We present a transformer for music personalization that recommends tracks given a station seed (artist) and improves the accuracy vs. a baseline matrix factorization method by 10%. Adding additional embeddings to capture track and station attributes further improves the accuracy of our recommendations by an additional 1% while also improving recommendation diversity, i.e. mitigating popularity bias. We analyze the learned embeddings and find they learn both explicit attributes provided at training and implicit attributes that may inform listener preferences. We also find that incorporating the station context of user feedback helps the model identify and transfer relevant listener preferences across different genres and artists. This particularly helps with music discovery on new stations.

We apply a transformer using sequential browse history to generate next-item product recommendations. Interpreting the learned item embeddings, we show that the model is able to implicitly learn price, popularity, style and functionality attributes without being explicitly passed these features during training. Our real-life test of this model on Wayfair’s different international stores show mixed results (but overall win). Diagnosing the cause, we identify a useful metric (average number of customers browsing each product) to ensure good model convergence. We also find limitations of using standard metrics like recall and nDCG, which do not correctly account for the positional effects of showing items on the Wayfair website, and empirically determine a more accurate discount factor.