Research On Hot Topic Prediction Based On Echo State Network

With the development of Internet, social network has become the main platform of information accessing, updating, communicating and interacting of users. If the topic from social platforms has been drawn attention and spread across the network, it will become a hot topic. The hot topic will impact on users, sociality and as well as the government. Therefore, this thesis proposes a few novel hot topic forecast models based on ESN (echo state network), which are used to predict the tendency of hot topic.Aiming at the defects existed in the echo state network forecast model, this thesis proposes two kinds of optimization methods, i.e., using artificial fish swarm algorithm and cuckoo search algorithm respectively to optimize the parameters of the dynamical reservoir that is suitable for the current hot topic prediction problem; then training model with the optimized parameters and thus obtaining the prediction results. Meanwhile, an improvement of the cuckoo search algorithm is proposed so as to overcome its shortcoming of the insufficient local search.Echo state network is a new type of recurrent neural network. Because of its simple training algorithm and convenient calculation, the ESN has been widely applied in nonlinear time series prediction problems. However, if the dynamical reservoir parameters are specified according to experience, the model may be not necessarily suitable for the current hot topic prediction problem. So, in this thesis, two kinds of models of the hot topic prediction are proposed aiming at parameters optimizing for the reservoir. The first one is the AFSA-ESN model of hot topic prediction which searches the best parameters for the dynamical reservoir, such as spectral radius and the input unit scale, by using the AFSA (artificial fish swarm algorithm). The second one is the CS-ESN model of hot topic prediction that overcomes the shortcoming of the AFSA which cannot optimize the parameters with different scope at the same time. The thesis applies CS (cuckoo search) algorithm to search the parameters of ESN dynamical reservoir, such as the dynamical reservoir size, spectral radius, sparse degree and input unit scale at the same time, and then train the model for hot topic predication with the global optimal parameters obtained. The Experiment results show that the prediction accuracy of the hot topic prediction based on AFSA-ESN is higher than traditional ESN prediction model, and that the CS-ESN model has higher accuracy and is more efficient than the AFSA-ESN model. Meanwhile, a new method of moving towards the best nest is proposed to improve CS. The method is discarding the worst bird’s nest and replacing the nest with a better location which is near the best nest, so that the nest will convergent to the best nest. The experiment results show that the improved CS algorithm’s convergence and optimization accuracy are better than CS algorithm. In addition, this thesis combines the improved CS algorithm with ESN to build the improved CS-ESN model of hot topic prediction. The experiment results show that improved CS-ESN model has higher prediction accuracy than the CS-ESN model.