Abstract: Lithology identification is an important foundation and critical part of well logging interpretation. However, the complexity of reservoir properties often leads to unavoidable inconsistency in cross-well lithology distribution and logging responses, impeding the robustness of cross-well lithology identification. This paper proposes several heterogeneous data representation features to reveal the invariant features of local reservoir description. Specifically, local topological information was first represented by employing graph representation technique in both vertical and lateral direction of well logging data. Then, three invariant features, namely structural tensor (ST), local binary pattern (LBP), and Hu moments (Hu), were obtained for robustly representing the local structure information of well logging data. Finally, the multi-view resampling strategy was adopted to address the distributional imbalance of log values and overlap of lithology in the original data domains, and meta-learning was employed to model the nonlinear relationship between the obtained heterogeneous features and target lithology information. Experiments were conducted using actual logging data from several wells in the Qijia depression of the Daqing Oilfield. The results indicate that the cross-well identification accuracy of the proposed method is more than 86%, proving it highly capable of solving the problems of inconsistent distribution of logging values and lithology between wells and unbalanced lithological data.