“…To demonstrate the effectiveness of the proposed social network spam filter, we compared its performance with several methods used in previous studies for spam filtering [4][5][6][7][8], namely the single DNN, CNN (convolutional neural network), Naïve Bayes, k-NN (k nearest neighbour), C4.5 decision tree, MLP (multilayer perceptron), SVM (support vector machine), AIRS (artificial immune recognition system), Adaboost M1 with decision stump as base learner, and random forest. The settings of these algorithms were as follows: single DNN (the same setting as for the DNN with ensemble learning); CNN (mini-batch gradient descent algorithm with patch size 5×5 and max pool size 2×2, the remaining parameters were the same as for the DNN); k-NN (k = 3); C4.5 (J48 implementation with the confidence factor of 0.25 and minimum instances per leaf = 2); MLP (backpropagation with {10, 20, 50, 100} units in the hidden layer (50 units worked best), learning rate = 0.1, momentum = 0.2, and iterations = 1000); SVM (sequential minimal optimization algorithm with C = {2 0 , 2 1 , … , 2 6 } (C = 2 2 worked best) and polynomial kernel function); AIRS (AIRS2 parallel algorithm with affinity threshold = 0.2, clonal rate = 10, hyper-mutation rate = 2, k = 3 and stimulation threshold = 0.9); Adaboost M1 with 10 iterations and decision stump as base learner; and 100 random trees were used in random forest.…”