Abstract: An improved collaborative filtering algorithm is proposed to solve the problem that the accuracy of recommended results decreases when the sparse degree of score matrix becomes larger in the alternating least square method． According to the user's rating of various Chaozhou delicious foods，the algorithm combines the interest similarity of other users，and uses the similarity of Chaozhou food attributes as the weight factor to complete the matrix． The experimental results show that the average MAE( Mean Absolute Error) value of the improved algorithm is 0． 583，which effectively improves the recommendation accuracy．

Key words: spark MLlib, food recommendation system, collaborative filtering algorithm, alternating least squares method, matrix completion