To address the high computational cost associated with molecular identification in spectral lines, we propose a graph-based feature learning framework to determine the molecular composition of spectral lines that incorporates rule-driven and node-level constraints to enhance identification accuracy. The approach begins with a sliding-window technique to extract localized spectral features that are then represented as nodes within a heterogeneous graph. To maintain semantic consistency across chemically diverse molecular species, rule-based constraints are introduced. Additionally, node-level constraints are employed to prevent the random walk process from becoming trapped in local cycles. The random walk mechanism is further refined through the integration of 2 parameters that guide node selection. Specifically, we design a frequency-aware sampling strategy in which transition probabilities are dynamically updated according to the number of times each node is used, combined with a stochastic modulation factor to enhance exploration and reduce sampling bias. Furthermore, a dynamic rule selection mechanism is employed to adaptively guide the random walk process in response to the evolving topological structure of the graph. Finally, the effectiveness of the proposed algorithm for molecular identification is validated using simulated data, achieving a recall rate of 94.63%. The method is then applied to the real G327 region, where molecular spectral lines are first identified and then used as inputs for XCLASS fitting to further verify the performance of the proposed algorithm. Moreover, the spatial map of molecular counts within spectral cubes is generated, demonstrating the capability for detailed chemical structure analysis.