I have got a question concerning Recommender Systems and Evaluation Metrics. I tested a few collaborative filtering recommendation algorithms on dataset containing amazon ratings. Here you can see MAE and RMSE for different algorithms tested with their best parameters (best parameters were chosen using k-fold validation, based on MAE).
Here you can see a table with mean precision@5, recall@5 and f1 scores computed for all folds of the same dataset, algorithms and their parameters. (precision and recall are presented in percents %)
I am not sure how to interpret the results. Obviously NMF algorithm is best, based on MAE error. However best precision is achieved for SVD and best recall and F1 is achieved for kNN with user based approach using Pearson similarity. Shouldn't low MAE mean, that precision should be high?