Partial label learning (PLL) is an essential weakly supervised learning method. In PLL, the example’s ground-truth label is unknown and hidden in a candidate label set comprising a subset of the label set. A multi-classifier is trained through a set of examples and candidate labels. The obscurity of the candidate label set makes the PLL challenging. Although high-efficiency graph-based methods without any parameters have been proposed to disambiguate, it is challenging to adapt a single graph structure for various actual data. The current work presents a new multi-graph embedding collaborative disambiguation PLL algorithm (PL-MGECD) to address the mentioned problem. The contributions of the current work are: (1) A unified framework for graph-based PLL is presented for the first time, which combines a least squares regression loss and a graph regularization term with ambiguous label constraints. (2) PL-MGECD adopts various graph structures in partial label learning for the first time and compensates for the lack of single graph representation data by fusing the complementarity of different graph structures. (3) PL-MGECD first introduces a graph structure constructed by candidate label information and employs the candidate tag information to modify the graph structure to compensate for the label disambiguation shortage through feature spatial similarity. (4) An efficient optimization algorithm is proposed. Extensive experiments demonstrate that the proposed PL-MGECD method has a competitive or superior performance over some traditional PLL methods.