To address the issues of insufficient utilization of information between nodes and overfitting in short text classification based on heterogeneous graph neural network, a method for semi-supervised short text classification based on gated double-layer heterogeneous graph attention network(GDHG) is proposed. GDHG consists of two layers: node attention and gated heterogeneous graph attention network. Firstly, different types of node attention coefficients are trained by node attention, and then the node attention coefficient is input into the gated heterogeneous graph attention network to obtain the gated double-layer attention. Secondly, the gated double-layer attention is multiplied by different states of the nodes to acquire the aggregated node features. Finally, the short texts are classified with the softmax function. In the proposed GDHG, the information forgetting mechanism of node attention and gated heterogeneous graph attention network is utilized to aggregate node information. Consequently, the information of neighboring nodes is effectively obtained. And then the hidden information of different neighboring nodes is mined to improve the ability to aggregate information from remote nodes. Experiment on four short text datasets , Twitter, MR, Snippets and AGNews, illustrate the superiority of GDHG.

The slime mould algorithm(SMA) yields low convergence efficiency and is easily dropped into local shackles. To address these shortcomings, an improved slime mould algorithm fused with multiple strategies(MISMA) is proposed.The Halton sequence is introduced to enrich the diversity of the initial population, and consequently the ergodicity and convergence precision of the algorithm are improved.The differential variation idea is employed to modify the global position update equations and enhance the global exploration capability and the continuous optimization performance.The local search strategy combining convergence factor and elite selection mechanism is ameliorated to improve the local development capacity of the algorithm. Thus, the global search performance and local exploitation capability of the algorithm are well balanced. The lens imaging learning strategy based on dynamic boundary is proposed to improve the individual quality and the capability of the algorithm avoiding prematurity and getting rid of local constraints.Numerical simulation experiments of 13 benchmark functions and some CEC2014 test functions show the strong robustness of MISMA.Moreover, the superiority and applicability of MISMA for dealing with practical engineering optimization problems are verified through parameter optimization experiments of the PV module model.