关于推荐系统准确率计算的技术咨询：稀疏评分矩阵相关疑问

Hey there! Let's break this down step by step since you're just starting out in recommendation systems—totally get the confusion around accuracy metrics and sparse rating matrices.

The term "accuracy" can be a bit ambiguous here, depending on what kind of recommendation task you're working on:

1. Rating prediction tasks (filling in missing values in the user-item matrix) #

You mentioned RMSE and MAE, which are the standard metrics for this task—they measure how close your predicted ratings are to the actual user-given ratings. If you're looking for an "accuracy"-style metric here, you could calculate the percentage of predictions that fall within a small threshold of the true rating (e.g., predictions within ±0.5 of the real score). For example:
Accuracy = (Number of predictions where |predicted - true| ≤ threshold) / Total number of test samples
But note: This isn't as commonly used as RMSE/MAE, since those metrics capture the magnitude of errors more effectively for rating prediction.

2. Top-N recommendation tasks (recommending N items a user might like) #

This is where "accuracy" is often referred to as Precision@k or Recall@k, which are more intuitive for recommendation scenarios:

• Precision@k: The proportion of recommended items in your top-k list that the user actually interacted with (or liked). Calculated as:
  Precision@k = (Number of relevant items in top-k recommendations) / k
• Recall@k: The proportion of the user's actual liked items that appear in your top-k recommendations. Calculated as:
  Recall@k = (Number of relevant items in top-k recommendations) / Total number of items the user liked
  These are the go-to metrics when you're focused on recommending the right items, not predicting exact ratings.

Using 0 to fill missing values in your user-item matrix is a big mistake, and yes, it will absolutely inflate your error metrics. Here's the key reason:

Missing ratings don't mean "the user disliked this item"—they mean "the user never rated this item" (maybe they never saw it, didn't care to rate it, etc.). Assigning a 0 treats these unrated items as explicit negative feedback, which is a false assumption.

Instead, here's how to handle missing values properly:

• For rating prediction: Only use the existing (user, item, rating) triples for training and testing. Split your known ratings into a training set (to train your model) and a test set (to evaluate RMSE/MAE). Ignore the missing values entirely during evaluation—you don't need to predict them to measure your model's performance on actual user ratings.
• For implicit feedback tasks (e.g., clicks, views): If you need negative samples, use negative sampling—randomly select a subset of unrated items as negative examples (instead of treating all unrated items as negative). This avoids penalizing the model for items the user might have liked but never encountered.

Hope this clears things up! Let me know if you want to dive deeper into any of these points.

内容的提问来源于stack exchange，提问作者Jane