Cross Platform Information Retrieval And Event Prediction Of Social Media

The development of social media has revolutionized how people understand and track real-world events.Users can upload and share multi-model information about the events on social media,such as text,image and video.Large amount of information distributes on various online social networks where abundant,wide-coverage and comprehensive information about trending events is available.In this thesis,we propose a framework for cross-platform information retrieval and event prediction of social media.Faced with large amount of information,searching has became a convenient tool for event understanding and tracking.However,most single-network searching methods only aim for single-network single-model information.In addition,most social media is biased which limits the coverage and diversity of single-network searching.In this paper,a novel cross-network framework is proposed to integrate cross-media information of a certain event and provide an immersive searching experience.Because of the multi-network distribution,there exist semantic gaps between heterogeneous social networks.This paper proposes using Hashtag and Tag,which are widely used meta data,to bridge the semantic gaps between social networks.The proposed framework includes four stages: multiple methods are adopted to filter Hashtag and Tag and implement representation,clustering and demonstration on search results.Given a search query,the first stage collects search results with Hashtag or Tag and filters the ones with high quality.The second and the third stage implement topic modeling and clustering on Hashtag and Tag and the last stage organizes the results with hierarchical demonstration.Qualitative and quantitative experiments on a large dataset with lots of queries validate the effectiveness of our framework.Faced with multi-platform information distribution,prediction has attracted attention from researchers.In general,news media provides serious,neutral information on events while social media features in real-time,fast propagating and subjective discussions on events.Based on the complementary information across platforms,this paper proposed the task of cross-network social event topic prediction.Information about social events is collected on news media and social media,and then divided into time slices.With parallel sliding windows,parallel dataset is constructed.Time slices are processed with topic modeling and in this way embedding and topic distributions of slices are generated.With the embedding as input,parallel dataset is used for training Seq2 Seq model as topic distribution prediction model.Corresponding experiments validate the effectiveness of the proposed framework.