The invention discloses a time series data and random descent method-based network reconstruction method, and aims at solving the technical problem that the reconstruction methods in the prior art are limited by the types and scales of network structures. The method is realized through the following steps of: initializing a weight matrix W of a network; decomposing observed time series data (as shown the specification) into (Nt-1) NS training samples to obtain a training sample set Y; randomly disrupting the sequences of the training samples in the training sample set Y so as to obtain a new training sample set Y'; carrying out weight adjustment on the initialized weight matrix W so as to obtain an adjusted weight matrix W'; and taking the adjusted weight matrix W' as a weight matrix W, randomly disrupting the sequences of the training samples in the training sample set Y to obtain a new training sample Y', carrying out weight adjustment on the weight matrix W to obtain an adjusted weight matrix W', and obtaining a reconstructed network structure matrix after repeatedly executing the step for T times.