Abstract: Only a single feature vector associated with the instance is considered in the traditional multi-label learning method, which predicts the overall label without a difference, thereby ignoring other similar instances and hiding label attributes. That results in such problems as less input spatial feature information, label attributes ignored and poor prediction about the large label space. In order to solve the above problems, the traditional multi-label learning task is converted into the sequence-to-sequence task of multi-label learning. On this basis is proposed Fea2Lab, a novel neural network model of multi-label learning label sequence generation. In this model, the chained feature vector input sequence formed by staggered sequential arrangement instances and similar instances is used to enrich the input spatial feature information. The label attributes are mined to predict the labels in a targeted manner. In the decoding process, the global label information is used to ease the problem that the cascading of the wrong label appears. The experimental results and ablation experiments on multiple data sets show the rationality, feasibility and effectiveness of the transformation task and the Fea2Lab model.