Abstract: In Web index page, recommending links of interest is beneficial for users to access Web resources efficiently. However, users won't spend a lot of time labeling samples and the data provided by them may just indicate whether or not a Web index page contains contents in which they are interested but give no information about which link really meets their interests. Therefore, the problem of link recommendation in Web index page is quite difficult since the training data lacks links' label while prediction for links of interest in a new Web index page is required. This problem is converted to a unique multi-instance learning problem and then solved by the proposed CkNN-ROI algorithm. Experiments show that this algorithm is more effective than other ones on solving this difficult link recommendation problem.