Mobile User Profiling Based On User Behaviors And Video Reviews

With the development of the mobile Internet,more and more users use data traffic to watch videos.Users produce a lot of behavioral data when watching mobile videos.Analyzing user's attribute,mining user interests and preferences from behavior data,will help network operators and video providers improve service quality.This dissertation proposes a video user profiling construction scheme based on user behavior data and video review text.This scheme does not rely on feature engineering to predict user demographic accurately,and a hierarchical tag system is designed to describe the user's preferences.A co-embedded attention network model is proposed to predict the age and gender of users.The model first learns video viewing vectors and comment vectors from user viewing records and video review document respectively,then combines the two vectors as video features and introduces an attention mechanism to build a co-embedded attention network.Finally,we iput the user's viewing history into the network to predict user's demographic.In the two tasks of gender prediction and age prediction,co-embedded attention network achieved 91.6% and 56.4% F1-scores respectively.A coarse-grained tag acquisition algorithm for video is proposed.The algorithm constructs a hierarchical attention network,learning the weights of words and reviews in the review text;finally predicting the confidence of all tags and determining the first-level tags of the video.In addition,a sample balance method based on document reconstruction is also proposed to solve the problem of unbalanced sample distribution while using more text information.The video coarse-grained tag acquisition algorithm achieved a recall of 78.05% on the tag recommendation task,and its performance is better than the mainstream text classification algorithms.An unsupervised hybrid tag extraction algorithm is proposed.In order to obtain the fine-grained tags of the video,the algorithm uses the TF-IDF algorithm and the improved Text Rank algorithm to extract two sets of tags from video comments,and then uses the voting method to determine the seconde-level tags of the video.The results of the survey experiments show that the tags obtained by the hybrid tag extraction algorithm can more accurately describe the video.Finally,based on the first-level tags and second-level tags,we construct a video hierarchical tag system,and the user's viewing history is used to establish a mobile video user's profiling of interest tags.