End-to-end Learning for Mining Text and Network Data

[摘要] A wealth of literature studies user behaviors in online communities, e.g., how users respond to information that are spreading over social networks. One way to study user responses is to analyze user-generated text, by identifying attitude towards target topics. Another way is to analyze the information diffusion networks over involved users. Conventional methods require manual encoding of world knowledge, which is ineffective in many cases. Therefore, to push research forward, we design end-to-end deep learning algorithms that learn high-level representations directly from data and optimize for particular tasks, relieving humans from hard coding features or rules, while achieving better performance. Specifically, I study attitude identification in the text mining domain, and important prediction tasks in the network domain. The key roles of text and networks in understanding user behaviors in online communities are not the only reason that we study them together. Compared with other types of data (e.g., image and speech), text and networks are both discrete and thus may share similar challenges and solutions.Attitude identification is conventionally decomposed into two separate subtasks: target detection that identifies whether a given target is mentioned in the text, and polarity classification that classifies the exact sentiment polarity. However, this decomposition fails to capture interactions between subtasks. To remedy the issue, we developed an end-to-end deep learning architecture, with the two subtasks interleaved by a memory network. Moreover, as the learned representations may share the same semantics for some targets, but vary for others, our model also incorporates the interactions among entities.For information networks, we aim to learn the representation of network structures in order to solve many valuable prediction tasks in the network community. An example of prediction tasks is network growth prediction, which assists decision makers in optimizing strategies. Instead of handcrafting features that could lead to severe loss of structural information, we propose to learn graph representations through a deep end-to-end prediction model. By finding ;;signatures;; for graphs, we convert graphs into matrices, where convolutional neural networks could be applied.In additional to topology, information networks are often associated with different sources of information. We specifically consider the task of cascade prediction, where global context, text content on both nodes, and diffusion graphs play important roles for prediction. Conventional methods require manual specification of the interactions among different information sources, which is easy to miss key information. We present a novel, end-to-end deep learning architecture named DeepCas, which first represents a cascade graph as a set of cascade paths that are sampled through random walks. Such a representation not only allows incorporation of the global context, but also bounds the loss of structural information.After modeling the information of global context, we equip DeepCas with the ability to jointly model text and network in a unified framework. We present a gating mechanism to dynamically fuse the structural and textual representations of nodes based on their respective properties. To incorporate the text information associated with both diffusion items and nodes, attention mechanisms are employed over node text based on their interactions with item text.

[发布机构] University of Michigan

[效力级别] Deep learning

[关键词] End-to-end learning;Deep learning;Text mining;Network mining;Computer Science;Science (General);Information and Library Science;Engineering;Science;Social Sciences;Information