Research On Trust Prediction Based On Extrime Learning Machine

With the popularization of the Internet,virtual interaction has become an indispensable activity in life.Just like in the real world,users also play different roles in the virtual world.In many cases,when you don’t know the information of a stranger,you can’t directly judge whether they are credible.You can indirectly judge whether they are credible according to the relevant information of their friends,and then you can conduct transactions or cooperations.Therefore,it is necessary to help judge whether strange users can be trusted or not.At present,most scholars make trust predictions based on a social network and predict the relationship in a social network which needs to combine some factors,such as the number of common neighbors,clustering coefficient and user similarity,etc.However,most social networks have the problem of sparse labels without enough labels to train the model.The prediction results will eventually be inaccurate.Cross-domain prediction methods can be used to solve this problem.At present,there are a few cross-domain trust prediction works.Some scholars built a cross-domain prediction model based on the asymmetric tritraining framework.The classifier adopts BP neural network and conducts experiments on six social networks.Based on the above research background,the existing research methods still have two shortcomings.The first deficiency is that the pseudo-label selection method is not automatic enough,and the second deficiency is that the speed of the classifier is not fast enough.Aiming at these deficiencies,the classifier and pseudo-label selection are improved.The following two works have been proposed:(1)The asymmetric tri-training model is extended to build a model framework.Extreme learning machine is used as a classifier for trust prediction,and the model is saved through a similar transfer learning method.The pseudo-label sample selection adopts the method of adopting the sample if the three classifiers are consistent,and not adopting if they are inconsistent methods.Compared with several existing evaluation methods on six online social networks,the experimental results show that the model is better than other evaluation methods in terms of recall and stability.(2)Based on the first work,the improved model has improved,which combines the tri-training model and the asymmetric tri-training model and uses the tri-training model to select pseudo-labels according to the minority obeys the majority voting mechanism.Experiments test the impact of adding special features,which compare the speed and accuracy of extreme learning machines and other methods with other existing algorithms on data sets.The experiments show that the model is better than other algorithms in terms of recall and stability.