Supervised machine learning can help to learn topological invariants for topological systems. However, it faces severe limitations when applied to non-Hermitian topological phases. The nonlinear nature of the non-Bloch topological invariant causes a sudden change in non-Hermitian or high-dimensional situations, while the supervised machine learning does not have the ability to grasp the rule at this moment. We have rigorously demonstrated the limitation of supervised machine learning and systematically delineates the valid scale range for effective model predictions through the generalization error bounds. By integrating the hypothesis space and the underlying physical mechanisms, our work establishes a generalized theoretical framework for evaluating the feasibility of supervised machine learning in topological phase identification.