Cross-Domain Image Sentiment Analysis

With the advent of the social era,it has become commonplace for people to share their images and to interact with others using comments on the websites or in the apps(e.g.,Flickr).Accurately understanding the sentiments conveyed by the vast volume of available images has been proved highly useful for a variety of applications such as image retrieval and social media analysis.Hence,more and more attention has been paid on image sentiment analysis and cross-domain image sentiment classification is a sub-research topic which full of challenges.For example,there are considerable discrepancies between image contents across domains,and the image features that used to convey sentiments are various.This thesis discusses about the new and hot topic-cross-domain image sentiment classification and comes up with a new methodology.Some research work has been proposed to solve image sentiment classification problem,and most of them are supervised learning with traditional machine learning algorithms on large scale visual features extracted from labeled training set,such as support vector machine(SVM)or Naive Bayes(NB).However,these models cannot work well on cross-domain image sentiment classification problem.The main challenges consist:(1)although it is easy to access millions(or even billions)of images today,the polarity labeling work,which is essential for training sentiment classifiers,can be time-consuming and expensive;(2)there are usually numerous domains in real applications(e.g..,Flickr has 1,187 domains[1]),similar visual features may represent different emotions in different domains,leading to the poor portability of image sentiment classifiers.For example,Figl.2 shows that the image from the face domain and that from the car domain share similar color histograms,but the sentiments conveyed are entirely opposite;(3)the requirement of large volume of training data makes it unrealistic to train domain-specific classifiers for each,and a general classification model cannot work well on every domain.Thus,given a source domain that contains a lot of labeled image data and a target domain that contains lots of unlabeled image data,it is an urgent research topic to come up with a new method to train an image classifier which can work well for target domain.We make a series of statistics and analyses on real datasets crawled from Flickr and propose a weighted co-training method for cross-domain image sentiment classification problem to overcome the drawbacks of the existing work.This method is proposed based on these facts:(1)along with images posted on applications,there are usually textual comments from other people,which have abundant features representing sentiment polarities.And the images and corresponding comments are often consistent in sentiment;(2)besides the huge differences of images across domains,there are some commonalities shared by images and comments between the source domain and the target domain.Thence,when training classifiers,this thesis comes up with weighted co-training method.This method considers not only about images but also comments.It computes the image and text similarities between source domain and target domain as weight of corresponding classifier and updates training dataset iteratively to improve the performances of classification models.To evaluate the performance of weighted co-training,we crawl about 10,000 images and their comments from three domains in Flickr,and label them manually.We perform thorough experiments and compare the proposed method with several baseline methods.The experimental results show that weighted co-training outperforms baselines in terms of standard classification measures including recall,precision,F1 value and accuracy.