Wed, May 10, 2017 @ 01:00 PM - 03:00 PM
PhD Candidate: Hao Wu
Kristina Lerman (chair)
Florenta Teodoridis (external)
Title: Learning Distributed Representations from Network Data and Human Navigation
Time: May 10 (Wed) 1:00-3:00pm
Room: SAL 322
The increasing growth of network data in online social networks and linked documents on the Web, presents challenges for automatic feature generation for data analysis. We study the problem of learning representations from network data, which is of critical importance for real world applications, including document search, personalized recommendation and role discovery. Most existing approaches do not characterize the surrounding network structure that serves as context for each data point, or they cannot scale well to massive data in real world scenarios. We present novel neural network algorithms that learn distributed representations of network data by exploiting network structure and human navigation. The algorithms embed data into a common low-dimensional continuous vector space, which facilitates predictive tasks, such as classification, relational learning and analogy. Efficient optimization and sampling methods improve the scalability of our algorithms.
First, we propose a neural embedding algorithm to learn distributed representations of generic graphs with global context. To capture the local network structure of each data point, we use random walks to sample nodes in a network neighborhood. Our algorithm is scale-invariant and the learned global representations can be used for similarity measurement of networks. We evaluate our model against state-of-the-art methods on node classification, role discovery and analogy tasks.
Second, we present a neural language model for generating text in networked documents. The model can capture both the local context of word sequences and the semantic influence between linked documents. The approach is based on an intuition that authors are influenced by words in the documents they cite and readers usually read the words in paragraphs by referring to those cited concepts or documents. We show improved performance in document classification and link prediction with our model.
Third, the information of how people navigate the network data online provides clues about missing links between cognitively similar concepts. Learning human navigation can also help characterizing human behavior and improving recommendation. We devise another neural network algorithm that accounts for human navigation patterns to learn better representations of text documents. We present empirical results of our algorithm on online news and movie review data, and show its effectiveness on real world applications.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon