Detecting Trends on Twitter : The Effect of Unsupervised Pre-Training
Sandra Bäckström, Johan Fredin Haslum · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2016
Unsupervised pre-training has recently emerged as a method for initializing super- vised machine learning methods. Foremost it has been applied to artificial neural networks (ANN). Previous work has found unsupervised pre-training to increase accuracy and be an effective method of initialization for ANNs[2]. This report studies the effect of unsupervised pre-training when detecting Twit- ter trends. A Twitter trend is defined as a topic gaining popularity. Previous work has studied several machine learning methods to analyse Twitter trends. However, this thesis studies the efficiency of using a multi-layer percep- tron classifier (MLPC) with and without Bernoulli restricted Boltzmann machine (BRBM) as an unsupervised pre-training method. Two relevant factors studied are the number of hidden layers in the MLPC and the size of the available dataset for training the methods. This thesis has implemented a MLPC that can detect trends at an accuracy of 85%. However, the experiments conducted to test the effect of unsupervised pre-training were inconclusive. No benefit could be concluded when using BRBM pre-training for the Twitter time series data.