Brand-related Information Retrieval In Social Media

Brand as a representation of the comprehensive quality for a commodity,reflects the public evaluation for an enterprise and its products.Brand Tracking is to track the brand-related data from time and space dimension,which provides data support for brand construction.How to tracking a brand has close connection to the type of transmission media.In recent years,the rapid development of social media has made great changes on the way of brand marketing,which has been turning into the experience marketing and word-of-mouth marketing gradually.In this situation,how to explore the brand-related information from social media for tracking brands becomes an important research topic.The key step to utilize such brand-related information is extracting them from massive social media effectively and efficiently.However,the traditional information extraction approaches have some limitations in analyzing the massive social multimedia data.Therefor,the approach of multimedia information retrieval provides a feasible way to solve the problem.In this paper,we discuss the problem of retrieving the brand-related information from social media and solve some key challenges in the steps of feature extraction,index construction and retrieval matching.The deep learning based single-modality brand-related information representation method.Towards the problems that the traditional handcrafted feature is not sufficient to represent the brand-related content,we try to learn a discriminative feature from single modality content by deep learning.We present a novel deep feature extract paradigm,which extracts the local features by deep learning and then generate the global feature by quantization,to enhance the discriminative of deep features.To alleviate the issue of lacking labeled sample for training,we also present a transfer learning based triplet ranking method to fine-tune the deep learning model.The effectiveness of the proposed method is demonstrated by experiments.The relevance inference based multi-modality brand-related information identification framework.Towards the problem that trained model under-fitting in noise social media data,which lead to a bad brand-related identification accuracy.In this paper,we propose a relevance inference based brand-related information identification framework.In this framework,a multi-task learning scheme is adopted to the classification model first,and then some social media elements such as user connection are incorporated to infer the relevance degree between testing samples and target brands.The proposed method alleviate the model under-fitting problem and provides an more stable and accurate identification results.Quality-biased model for brand-related cross-media retrieval.The traditional contentbased information retrieval methods always neglect to measure the quality of candidate samples,which may lead these methods to get trivial results.To solve such problem,we propose a quality-biased model for information retrieval in social media.First,we develop a multi-view embedding framework which maps the multi-modality features into a unified latent space.Further,we employ a quality model for microblog retrieval,which incorporates both quality evaluation and content matching to optimize the ranking of final results.The proposed method is able to improve both the accuracy of retrieval and the results quality.The semantic consistency hashing based efficient brand-related cross-media retrieval.The large amount of semi-structure and unstructured data in social media necessitates the cross-media hashing.However,generating a high quality hashing function by machine learning has to face such two challenges: how to reduce the quantization loss and how to keep the semantic consistency of data.To solve such problems,we propose a supervised discrete hypergraph hashing algorithm.In the proposed algorithm,the semantic consistency is guaranteed from two aspects of label consistency and metric consistency.Moreover,to reduce the quantization loss,the discrete variables are optimized directly in a bit by bit manner.The proposed paradigm is adopted in retrieving large-scale brand-related information task to optimize the retrieval efficiency.Design an brand-related information retrieval framework.According to the specific demand of brand-related information retrieval task,a framework for retrieving brandrelated information from massive social media data is developed,which synthetically utilizes the achievements of each chapter.The effectiveness of proposed framework is analyzed by experiments.