“…Hence, the optimized hyper-parameter combination is not restricted to the existing Orthogonal Array table. Specifically, 10 hyperparameters are optimized whose values are set as follows (based on the Taguchi L32 type B orthogonal array): the number of hops of local neighborhood in LPE module h (2,4), the vector dimension K 1 of the (1,2,4,8) dimension K 2 in NMF (1,2,4,8), dimension c in GCN (8,16,24,32), K GCN in GCN (1,2,4,8), number of neurons in each LSTM layer (8,16,32,64), number of neurons in each FCN layer (8,16,32,64), number of neurons in each attention layer (8,16,32,64), the learning rate (0.0001, 0.0005, 0.001, 0.005), the batch size (20,40,60,80). After tuning, the initial parameters are set to h = 4, K 1 = 4, K 2 = 1, c = 8, K GCN =, 16, 16, 16, 0.0001, 40, respectively.…”