ICDM 2014 (Shenzen)

1. Deep learning and computer vision have different areas in image recognition. The formal needs learning based on training. And the latter thinks about only comparision or matching.

 

2. Neural netwok may cost much time for operatjon. And computer vision is same. So you shiuld consider the choice well in each case.

 

3. Health will be big potential area. In this conference, many members and presenters were interedted in applying data mining to that area.

 

4. Tutorial subject: Mobile image search

 

- much data shoud be generated in case of SIFT which lowers network performance and costs large memory capacity. And it could reduce battery life and delay comparison time in a sever side.

 

- So, more compact image descriptor is needed.

 

- Compression rate and accuracy have trade off relationship.

 

- When data is gathered enough, the algorithm operation should be stopped such as searching for local features

 

- Local image featyure means interest point.

Global image feature is histogram of local features.

 

In other words, large amount of local features are replaced with global features.

 

- There are already many papers about data compression of image descriptor.

 

- Compact descriptor for visual search was initiated by MPEG. It is named as MPEG- 7 CDVS(Compact Descriptor Visual Search) Standard

: many features are defined in this format such as colors....

 

- MPEC-7 specifies bitstream of descriptor.

 

- The target images are categorized as textured rigid objecs, landmark print document and etc.

 

- Memory size is reduced from 400mb to 1mb for example in mobile device side such as a samrtphone. It means the memory for image operation.

 

- Currently, MPEC-7 is adopted by Baidu, Tencent and etc.

 

- Feature raw data can be reduced from 512kb to less size (256kb, 10kb or 1kb....). The rate can be decided by each situation and selection.

 

- only necessary local featuers are gathered.

 

- WeChat image platform provides this feature.

 

5. Tutorial subject: online machine learning

 

- Data mining is composed of infrastructure(DBMS, storage..), analytics(algorithm), and application(viewer).

 

- Previous offline machine learning needs new trainig over all. But online learning don't touch previous trained data when adding new data.

 

- one kernel represents one similarity.

 

- Learning means that if there is error like mismatching, the function like pattern shoyld be updated. So, there are continuous updating operation.

 

6. Research session

 

(I can't remember the title.)

 

The author said his system detects the geograpgical place with given music. He stressed his algorithm is not based on classification but regression.

 

Regression vs. Classification.

http://www.simafore.com/blog/bid/62482/2-main-differences-between-classification-and-regression-trees

 

7. Research session

 

(I can't remember the title.)

 

Collective prediction of multiple types of links in heterogeneous information networks.

 

- Link prediction.

 

- Discover kind of hidden link

 

- It is applied to medical area. There are three nodes of germ, desease and drug. You are given known link among nodes. And you can find new hidden link relationship. For example, some desease can be cured by new link relationship.

 

8. Research session

 

(I can't remember the title.)

(I am not sure of understanding exactly.)

 

- Scalable svm-based classification in dynamic graphs

 

- There is limited memory capacity in case of previous SVM.

He suggests a cincept of window by which you can control memory size dynamically.

 

------------------------------------------------

Following is about demo session.

 

1. Health care: Insurance companies and hospitals upload much their experience information amd history tp the system. The system analyze those data and predict each person's future health care costs.

 

I asked who is the customer of this system. The presenter said normal persons coulde be targets but organization such as hopitals or insirance company will be its major targets.

 

2. Route recommendation system

: You inputs your current locatuon and your destination. And you input your starting time and your favorate arriving time. The system provides routes which include famous and many visiting places of the way to the destination.

 

I asked how the system get much necessary data such as places. The presenter said men should input data manually. But it could be uploaded automatically llin future.

 

3. Deep learning framework:

 

4. News summary system: The presenter was Korean. There are so much news amd comments in internet newspaper or social network sites. The system gatheted those information. When users input their favorate keyword which could be current hot issue keyword, the system make summary statements with those gathered data.

 

I asked making statement is not easy work. He said especially Korean statement is more difficult and his system is making korean sentences. He said that there's much progress in case of English.

 

He said that it's illegal to gather korean newspaper's content. So, it could be an issue to make it commercial one.

 

5. Rule discovering system:

A, B, C are attributes which have each their value. Rule is like below.

A -> B

It means that if you know A, you can know B.

 

The system provides rules such as A-> B , B->C based on much data.

 

 

6. Review analysis system:

There are so many reviews and many of them are false comments. The system analyze the review comments and filter bad ones. The system analyze reviwers' locations and comments.. although I can't remember well, there are many things to consider for analysis.

 

7. Energy usage monitoring and analyzing system:

: It seems similar to a smart grid system. The presenter said many of electric amount can be saved. The system gathered the usage and analyze.