In this paper, we propose a transductive learning method for content-based image retrieval: Multiple Random Walk (MRW). Its basic idea is to construct two generative models by means of Markov random walks, one for images relevant to the query concept and the other for the irrelevant ones. The goal is to obtain the likelihood functions of both classes. Firstly, MRW generates two random walks with virtual absorbing boundaries, and uses the absorbing probabilities as the initial estimation of the likelihood functions. Then it refines the two random walks through an EMlike iterative procedure in order to get more accurate estimation of the likelihood functions. Class priors are also obtained in this procedure. Finally, MRW ranks all the unlabeled images in the database according to their posterior probabilities of being relevant. By using both labeled and unlabeled data, MRW can be seen as a transductive learning method, which has been demonstrated to outperform inductive ones by previous research work. Systematic experiments on a general-purpose image database consisting of 5,000 Corel images demonstrate the superiority of MRW over state-of-the-art techniques.