Many classifiers cannot get good results facing numerous sample features, bayes algorithms get poor estimate of joint probability distribution with limited samples, effective similarity measure is needed to build a local classifier. Considering these problems, an improved multinomial Naïve Bayes algorithm based on weighted instances (IWIMNB) with cosine similarity is proposed. Taking different attributes contributing differently to the classification decision weight into account this algorithm uses cosine similarity as a metric of the similarity between training and validation instances. Selected training instances weighted by cosine similarity are used to train modified Naïve Bayes model. Results of final experiments show that the average classification accuracy of the proposed IWIMNB algorithm gets significant improvement.