The profitability promoted by Google in its brand new video distribution platform YouTube has attracted an increasing number of users. However, such success has also attracted malicious users, which aim to self-promote their videos or disseminate viruses and malwares. Since YouTube offers limited tools for comment moderation, the spam volume is shockingly increasing which lead owners of famous channels to disable the comments section in their videos. Automatic comment spam filtering on YouTube is a challenge even for established classification methods, since the messages are very short and often rife with slangs, symbols and abbreviations. In this work, we have evaluated several top-performance classification techniques for such purpose. The statistical analysis of results indicate that, with 99.9% of confidence level, decision trees, logistic regression, Bernoulli Naive Bayes, random forests, linear and Gaussian SVMs are statistically equivalent. Based on this, we have also offered the TubeSpam -- an accurate online system to filter comments posted on YouTube.