Paper Review in Computer Science

http://pierre.senellart.com/publications/meng2015discovering.pdf

This paper is to search and discover meta-path. The definition of meta-path is a sequence of node classes. Authors of this paper provides a way to discover candidate paths using greedy-search and node class generation. It could generate meta-path in a more efficient way, but I am suspicious of how the problem itself is meaningful to solve and prove its effectiveness. For example, it looks that if we have select any sequences of a topic, it would be relevant. They could be more relevant than others ? Of course, it can be, but I think it is really dependent on its context, and hard to distinguish against simple algorithm. Its p-value (0.09) from t-test against random selection shows the evidence.

Their community is based on Modularity, which is a property of a network and a specific proposed division of that network into communities. There are many kinds of numerical methods to calculate it fast until O(n log^2 n) in the sparse graph.

You can refer to http://arxiv.org/pdf/cond-mat/0408187.pdf

http://research.microsoft.com/pubs/226627/[AAAI2014]%20DNN%20for%20Graph%20Cut.pdf

The paper is about the possibility exploration of deep learning to graph data. Its pre-training is to do SVD decomposition and try its feature space reduction. The decomposition is from spectral clustering in theory, and the only difference is a regularization term using KL divergence, which is in a form of L1. The only difference causes big improvement when we use the reduced feature space as an input of K-Mean. However, it is not clear about how to train their actual autoencoders, and the regularization is not that new from what researchers tried before. But, it could be regarded as a meaningful trial because it was the first to my knowledge as well. 

http://www.ntu.edu.sg/home/axsun/paper/SIGIR14Petar_CR.pdf

Fine-grained location extractor is important to organize experience of people from the web. Here they introduce a POI inventory extractor and time-aware POI tagger. 

POI inventory is from Foursquare, and tagger is based on CRF. In particular, it considers time-granularities such as 2 hours, 1 day, ...  , and 1week+. Their features (location with temporal property) are extracted by lots of heuristics based on observation. The features are used for training CRF and compared to other alternatives. Its grammatical features for CRF are from Twitter nlp (from UW NLP group)(https://github.com/aritter/twitter_nlp) It contains lot of experience for handling the data, and worthwhile read when you have a plan to make very specific tagger to a certain type of data set. In other words, it looks too much data-specific paper.  But, I like its general approach to data itself in the data exploration part.