Web spamming is the deliberate manipulation of search engine indexes to make a page get high ranking than which it deserved considering its true value. Since the evolution of web spam, a new based on machine learning algorithm web spam detection method which has self-learning ability has emerged. Web spam detection is viewed as a binary classification learning problem. Because labeled training examples are fairly expensive to obtain which need the participation of experts in this field and labor costs, how to fully utilize a large number of unlabeled web page examples on the web is a challenge faced by web spam detection. In this paper, we present a web spam detection algorithm according to improve tri-training. It uses a small amount of labeled examples and a large number of unlabeled examples to train classifiers, which can reduce the cost of labeled examples and improve the learning performance. Both web page content features and link features are used in this paper.