Accurate project cost prediction is an important goal for the software engineering community, but there are some defects in the method to estimate software cost. Collaborative Filtering has been developed in information retrieval researchers successfully which recommends items based on other user's reference in historical data set. Cost estimation based on Collaborative Filtering is researched. Because only numerical data can be handled in traditional collaborative filtering technology, and there are non-numeric numeric data in the attributes that describe the project characteristics, so the different strategies are used to normalize for describing the project's properties. And then the mean values are used for the missing contents. Cosine similarity is used to calculate the similarity between projects. Finnally cost is estimated using the weighted sum of the efforts in k-nearest neighbors. The method is applied in an experimental case to evaluate the effort estimation, and the result shows the accuracy of estimation may arrive to 80%.