Magnostics image-based Search of interesting matrix view for Guided network Exploration (a web-based approach to network information matrix images)

Network, relationship and other data into the adjacency matrix (red represents two nodes is a person, there is a link between), but the resulting matrix will be due to the order of the problem of the different arrangement, in the first will find that because there is a clustered block area and easily divide the data into two parts, Then according to the specific meaning of the data to know the meaning of its representative, in this figure can be seen as two groups.

When the data is analyzed, it is transformed into matrix form, and some matrix reordering algorithm is used to transform the matrix into a specific pattern. Now you want to query the image to find out which well-deformed matrices belong to the same pattern.

The current data volume is particularly large, the data dimension is very many, the style is complex and changeable, for the detection of specific image pattern, it is difficult to use the naked eye to identify a pattern is not a type of pattern. The current matrix-based image recognition image is feature, and there is not a clear standard to assess which image features are suitable for identifying the pattern of which type of image

Two Contribution

6 types of feature for detecting specific pattern

4 scoring criteria for measuring feature test results

A tool that complements feature-based analysis in visual analysis

Three Experiment

• Select feature

The author chooses 27 commonly used feature to describe the image, and defines 3 new feature (medium red as the newly defined feature).

• Construct data

The author chooses the 6 patterns to be probed and adds 4 different ways to combine them.

6 Types of pattern:

4 Types of changes:

A) Variations

Different forms of expression of the same pattern

B) Point Swap

Random exchange of the points, divided into 0, 1, 2, 4, 8, 16, 32 percent of the case, 32% of the case can not distinguish its own pattern (at this time is noise)

C) Index Swap

Randomly swap two or two columns, where there are 0-10 random exchanges.

D) Masking

Add additional points, (0% to 16%)

• Generating vectors

For each matrix and a feature, can be regarded as a vector, and calculate the Euclidean distance between the same feature of the distance between the vector can be considered two patterns very close.

• Standard analysis

Based on the distance of the vectors, 4 metrics are defined by the author.

a) 0?2?0?2?0?2?0?2?0?2?0?2?0?2C1 rating Standard 1

Used to assess whether a feature can distinguish pattern from noise (masking), the darker the color is, the higher the score is 1. A 0 or a fork indicates that it has no practical meaning in this change and does not need to be tested.

b) 0?2?0?2?0?2?0?2?0?2?0?2C2 rating Standard 2

To assess the degree to which a feature distinguishes between the different manifestations of the same pattern, if the distance between the vectors is greater, the explanation can be effectively distinguished.

The image represents the difference in the distance between the matrix of the same pattern and the feature component vector, and the darker the color indicates the closer the distance.

The darker the color, the higher the score of the C2 is 0.5.

c) 0?2?0?2?0?2?0?2?0?2?0?2?0?2C3 rating Standard 3

To assess the degree to which a feature is sensitive to noise after adding noise (point swap, Index swap) to the same pattern, the less sensitive it is, the better the effect.

The horizontal axis indicates the rate of noise addition, and the ordinate indicates the distance from the original pattern vector. The black dots represent the distance between the vector and the original pattern vector of different patterns of pattern, and the red dots indicate the average distance, which indicates that the distance increases and is not unpleasant when the noise is added. So the sensitivity level is very low, the C3 score is better

It can be seen that the distance is growing fast. So the sensitivity is very high, the C3 score is worse

This figure shows that the distance growth is not very fast.

But compared with the first figure, although the sensitivity is not as good as it is, but the change in pattern performance is very fast, that is, the black dots in the graph. So the C3 score is inferior to the first figure, but the C2 score is higher than it

This figure shows that for the increase in noise, the distance is a gradually increasing trend, the slower the trend, the better the noise level.

For C3, the darker the color, the higher the intensity of noise immunity.

d) 0?2?0?2?0?2?0?2?0?2?0?2C4 rating Standard 4

Used to evaluate the ability of a feature to separate the pattern, and C1 to distinguish it from the noise pattern. All vectors 22 are averaged to determine whether the distance between them is far apart from each other and can be differentiated.

Indicates the distance between the pattern, that is, the size of the ability to distinguish the pattern, the higher the overall box to the greater the ability to differentiate, the red feature is very strong, the higher the C4 score.

Five. Experimental summary
In the scoring standard, C1,c4 is the main scoring standard, and the C2,C3 should be weighted and trade-off according to the specific feature meaning characteristics.

The C1 (blue), C2 (red), C3 (brown) of the feature, the more black dots on the C1, C2, or C3, the higher the ranking in feature. The box-selected feature in the figure represents the 6 types of feature selected by the last author

Magnostics image-based Search of interesting matrix view for Guided network Exploration (a web-based approach to network information matrix images)