A key issue that determines the effectiveness of the manifold alignment approaches is to discover the correlations between the points sampled from different manifolds. This paper proposes a new idea, which uses the geodesic distances to originally construct the correlations between the points sampled from different manifolds, and then uses the similarities measured by the local geometric structures of the samples to modify the correlations, thus discovering the correlations between the data points sampled from different manifolds more accurately. Further more, the paper proposes a new semi-supervised manifold alignment algorithm,which projects multiple manifold data sets to acommon low-dimensional space by using the known correspondences information and the discovered correlations between the sample points.Compared with the traditional semi-supervised manifold alignment algorithms, the proposed algorithm can find the matching points of different manifold data in the low-dimensional space more accurately when the prior information is not sufficient.Finally,the effectiveness of the proposed algorithm is validated by the experiments on real-world data sets.