This thesis presents a new paradigm for non-rigid 3D shape retrieval, which is also called Bag of Feature Graphs (BoFG). The main idea is to connect only the features on the shape to construct the graphs so that the number of points involved in the computation is greatly reduced. Given a vocabulary of geometric words, the BoFG approach generates a graph that preserves the spatial information among features for each word. The spatial information is weighted by its similarities to each word so that points unlike the word category are eliminated. And the graphs are captured by the affinity matrices of Weighted Heat Kernels (WHK) whose eigenvalues form a shape descriptor. Also, the BoFG approach can supports partial 3D shape retrieval by coupling with graph matching techniques and comparing only sub graphs that represent common parts of the shape. Finally, experiments are conducted and show that the proposed BoFG method is faster to compute and the retrieval performance is also competitive compared with other state-of-the-art methods.