I've just published a practical series about building a music recommendation system using ML.NET and ASP.NET Core.
The goal wasn't just to train a recommendation model, but to explain the complete workflow, from understanding how recommendation systems work to evaluating whether the model is actually good enough to be used in a real application.
The series covers:
šø How recommendation systems work
š How to evaluate a recommendation model
š¤ Training a rock music recommender with ML.NET
š Evaluating the recommender's performance
š Exposing the model through an ASP.NET Core Web API
Articles:
https://devfullstack.net/blog/how-recommendation-systems-work
https://devfullstack.net/blog/how-to-evaluate-a-recommendation-model
https://devfullstack.net/blog/training-a-rock-recommender-with-mlnet
https://devfullstack.net/blog/evaluating-the-rock-recommender
https://devfullstack.net/blog/building-the-rock-recommender-api
Top comments (2)
Nice that you split evaluation into its own dedicated post instead of tacking on an accuracy number at the end, that step is the one most recommender tutorials skip. Genre-based recommenders like rock are a good test case for this too, since a naive precision/recall number can look great while the model is just learning "recommend more rock to rock listeners" rather than anything about specific taste within the genre. Did your evaluation post get into ranking-aware metrics like NDCG or MAP@K, or mostly classification-style accuracy? Curious how ML.NET's built-in eval tooling handles ranked recommendation output versus a straight classifier.
Thanks for your comment! In the evaluation article, I focused mainly on Precision@K, Recall@K, and NDCG, since I wanted to evaluate the quality of the ranked recommendations rather than just rating prediction. I briefly mentioned RMSE and MAE because they're commonly used in recommendation systems, but the project itself emphasizes ranking-based metrics. I completely agree that ranking-aware evaluation is essential for recommender systems.