A frequent item of a data stream is a data point whose occurrence frequency is above a given threshold. Finding frequent item of data stream has wide applications in various fields, such as network traffic monitor, data stream OLAP and data stream mining, etc. In data stream model, the algorithm can only scan the data in one pass and the available memory space is very limited relative to the volume of a data stream, therefore it is usually unable to find all the accurate frequent items of a data stream. This paper proposes a novel algorithm to find ε-approximate frequent items of a data stream, its space complexity is O(ε^-1) and the processing time for each item is O(1) in average. Moreover, the frequency error bound of the results returned by the proposed algorithm is ε(1-s+ε)N. Among all the existed approaches, this method is the best.