Mining Techniques For Online Videos' DanMu Data

As the rapid development of Internet,the access of streaming media services is very convenient,which enables users to watch online videos anywhere and anytime.Nevertheless,as the change of user requirements,users do not satisfy basic functions of online videos,such as watching videos.Instead,they look forward to new functions to meet their new requirements,e.g.,expressing instant feelings and interacting with other users during watching videos.While traditional interaction methods(e.g.,commenting,rating)can not realize these goals,i.e.,communicating with other users and expressing feelings during the video.Recently,a new interaction method for online videos(i.e.,screen-show comments or "DanMu")has been popular in many online video platforms.DanMu is scrolling marquee comments,which are overlaid directly on top of video and synchronized to specific playback time.Meanwhile,by DanMu,viewers can commu-nicate with others and express their owning feelings in real time while watching the video.Besides,video producers can acquire users' feedbacks and suggestions of video segments by the new interaction method,i.e.,DanMu,and then accurately improve the quality of subsequent videos.Nowadays,almost all domestic video websites(e.g.,Iqiyi,Tencent)have adopted DanMu on their platforms.In fact,the new type of interaction method has several unique characteristics(e.g.,obvious multi-bursts phenomenon and stronger herding effect)and tremendous appli-cable value.Meanwhile,the unique nature of DanMu also brings new challenges to research and applications.For example,it makes video's popularity prediction harder.What's more,comparing with the widely spreading in industry,the attention from re-searchers on DanMu is far from enough.As yet,there is not even a data-driven research to explore characteristics and patterns of the new interaction method for videos.On this basis,in this paper,we devise modeling methods by utilizing data mining techniques and combining psychology,business and other interdiscipline subjects,which can pro-vide a comprehensive understanding on DanMu.The contributions of this paper are summarized as follows.First,we comprehensively analyze and quantify the new features of DanMu.This new type of interaction methods between users and online videos posses great values for academic research and business applications.However,there are lack of data-driven research to deeply prospect the unique nature of DanMu.Therefore,we analyze the unique characteristics of DanMu from various perspectives.Specifically,we first illus-trate some unique distributions of the new interaction method by comparing with tradi-tional reviews.Second,we devise a model to quantify herding effect of DanMu.Then,a multi-bursts,detection model is proposed.Next,to explore reasons of the unique na-ture,we design a leading DanMu identification model.Finally,we construct two growth prediction models for DanMu on videos and video segments respectively.Extensive ex-periments on a crawled real-world dataset clearly demonstrate the effectiveness of our proposed models.Second,to well solve the uncertainty from DanMu on video popularity prediction,we propose a probabilistic prediction model by fusing multi-factors.Comparing with traditional online videos,the unique nature of DanMu-enabled videos makes video's popularity prediction accompany with larger uncertainty.On this basis,we construct a multi-factors fusion popularity prediction model.Specifically,we first devise a herd-ing effect factor based on popular videos,popular DanMu and videos,uploading date.Then,we generate an uploader's influence factor and a video quality factor.Along this line,we construct a model that incorporates the herding effect,uploader's influence and video quality for predicting the video popularity.Experimental results exhibit that our proposed model can accurately predict DanMu-enabled online videos' popularity.Finally,we propose a deep mixture model for large-scale image classification by utilizing DanMu's key features.Comparing with traditional comments,DanMu is syn-chronized to specific playback time and contains rich semantic information including subjective semantic(e.g.,happy,angry)and objective semantic(e.g.,actors,environ-ment),which can be used to label DanMu,s corresponding images.However,as each online-video website includes tremendous videos,the number of images extracted from these videos is also extremely huge,which brings a big challenge to current image clas-sification models' efficiency end effectiveness.To deeply explore this problem,we can generalize this problem to the task,i.e.,how to classify large-scale images.Hence,we propose a deep mixture algorithm to support large-scale image classification in this pa-per.First of all,according to image categories' semantic correlations,we propose a fuzzy spectral clustering to construct a two-layer(i.e.,category layer and group layer)ontology.Then,based on the two-layer ontology,each task group is assigned with a basic convolutional neural network.It is noting that the image categories with sim-ilar learning complexities can be assigned into the same task group according to the two-layer ontology,which ensure the separability among different basic deep networks.Finally,we devise a gate network to automatically fuse basic deep networks' diverse outputs for the final image classification predictions.According to experimental results,the proposed deep mixture model has a competitive performance for large-scale image classification by comparing with plenty of baselines.