AAAI-2013概述（2）

AAAI2013下半部分文章中，重要的有如下四篇（分属四类，所以未做分类）

1. Story Generation with Crowdsourced Plot Graphs

目的：

提出Story Generation Problem的新解决方法。以前的方法都要基于一个事先定义好的domain model，而这里提出的方法是从语料库中学习到domain model。

方法：

1） Plot Graph Learning

Plot graph models the author intended logical flow of events in the virtual world as a set of precedence constraints between plot events.

First, a corpus of narrative examples is acquired.

Second, the events their precedence relations are learned.

Third, mutual exclusions between events are learned.

2） Generate Stories Using Plot Graph Based Story Generation Algorithm

The story generation algorithm samples the space of possible narratives by iteratively adding “executable” events to a story.

2. Ranking Scientific Articles by Exploiting Citations, Authors, Journals, and Time Information

问题：

(1) how to rank articles in the heterogeneous network; 

(2) how to use time information in the dynamic network in order to obtain a better ranking result.

方法：

a graph based ranking method, which utilizes citations, authors, journals/conferences and the publication time information collaboratively.

1） Construct a heterogeneous network which contains three sub-networks (citation network, paper-author network, and paper-journal network).

2） Define a time-aware weight to each edge in the network.

3） Conducts the HITS and PageRank algorithm collaboratively on the nerwork above to generate the ranking list of scientific articles.

3. Sparse Multi-Task Learning for Detecting Influential Nodes in an Implicit Diffusion Network

目的：

Identify influential nodes is a central research topic in information diffusion analysis.

方法：

A multi-task sparse linear influence model(MSLIM)（这篇文章的核心）

1） Topic modeling

Constructed the topics using the Latent Dirichlet Allocation to find topic component of every tweet.

2） Applying MSLIM

Linear influence model(LIM) can model the influence for each node and has been proven to be effective for predicting the future volume for each contagion. MSLIM can simultaneously conduct contagion-sensitive volume prediction and influential node detection in a unified framework.

4. Supervised Coupled Dictionary Learning with Group Structures for Multi-modal Retrieval

问题：

To find a better similarity mapping function across heterogeneous high dimensional features for multi-modal data.

方法：

Introduce coupled dictionary learning into supervised sparse coding for multi-modal and propose a method called Supervised coupled dictionary learning with group structures for Multi-Modal retrieval (SliM).

Main part of this paper is the introduction of the model SliM and its formulation and optimization. In evaluation part, this method is applied to the Wiki Text-Image data which contains text-image pairs from ten different categories.