2.2 Distance on Numeric Data Minkowski Distance

Loading...

来自 University of Illinois at Urbana-Champaign 的课程

Cluster Analysis in Data Mining

149 个评分

University of Illinois at Urbana-Champaign

Cluster Analysis in Data Mining

149 个评分

课程 5（共 6 门，Specialization Data Mining ）

Discover the basic concepts of cluster analysis, and then study a set of typical clustering methodologies, algorithms, and applications. This includes partitioning methods such as k-means, hierarchical methods such as BIRCH, and density-based methods such as DBSCAN/OPTICS. Moreover, learn methods for clustering validation and evaluation of clustering quality. Finally, see examples of cluster analysis in applications.

从本节课中

Module 1

2.1 Basic Concepts: Measuring Similarity between Objects3:23

2.2 Distance on Numeric Data Minkowski Distance7:01

2.3 Proximity Measure for Symetric vs Asymmetric Binary Variables4:55

2.4 Distance between Categorical Attributes Ordinal Attributes and Mixed Types4:04

2.5 Proximity Measure between Two Vectors Cosine Similarity2:54

2.6 Correlation Measures between Two variables Covariance and Correlation Coefficient13:32

与讲师见面

Jiawei Han
Abel Bliss Professor
Department of Computer Science

[SOUND] Now we examine Session 2:

Distance on Numerical Data: Minkowski Distance.

So we first introduced data matrix and dissimilarity matrix, or distance matrix.

Data matrix is referenced in the typical matrix form is we have n data points,

we use n rows.

We have l dimensions, we use l columns to reference this data set.

The distance matrix or dissimilar matrix usually is referenced in

triangular matrix because you have n data points,

we register only the distance between like objects

one versus one or two versus one to look at their distance.

But the distance function usually could be quite different, but

different kinds of variables, like a real,

boolean, categorical, ordinal, ratio, and vector variables.

Their distance definition could be different.

Moreover, sometimes they may have weights associated with those

different variables based on the applications and data semantics.

We are going to introduce this more later.

Then, we look at an example.

Suppose we have four points, four objects in two dimensional space.

Then the data matrix is rampant in this typical form, okay.

Then for dissimilarity matrix, or distance matrix, for Euclidean distance,

you can see the matrix is in this way, like x1, x1, they are identical.

Their distance is 0.

x2, x1, their computation is based on the distance.

This part is two, this distance is three, you take the sum of the square area.

You take square root, you get this value.

Then in general, we define the Minkowski distance of this formula.

It means if we have area dimensions for object i and object j.

Then their distance is defined by taking every dimension to look

at their absolute value of their distance, then to the power of p,

then you sum them up, get the root of p.

Then we get the Minkowski distance.

Such distance that P is the order we usually also call this

distance this LP norm.

The Minkowski distance in general have these properties.

The first property is called positivity.

It means, the distance be equal zero when

they are identical otherwise they are greater in there.

The second property called symmetry means

the distance between I and J, distance between J and I should be identical.

Then the third one called triangular inequality means for

the distance between i and j.

If you go through k, that means you go first to k, then from k to j.

That distance should be no less than directly go from i to j.

Should be greater than or equal to d(i,j).

The distance measure that satisfies these three properties are metric distance.

Notice not all the distance are metric for example set difference is not metric.

Of course in this lecture we mainly discuss metric distance.

Then we look at some special cases of Minkowski distance.

If p = 1, we call L1 norm, they also call Manhattan or

city block distance define this formula.

In particular, if we are dealing with binary vectors we call these

Hamming distance is the number of bits that are different.

Then if p equals two, you can find also L2 norm.

As Euclidean distance defined the typical way like this, and

this vary for the middle one I think for all people, okay?

Then if p goes to infinity means we're dealing with L infinity norm or

L max norm, this is also caused supernum distance.

Is that you find, is limit for p goes for infinity.

In that case, actually the distance is really the maximum

difference between any attribute of the vectors.

For example you can see for F, from 1 to L.

The maximum such absolute value of the distance,

is the distance of L infinity norm or supremum distance.

Let's look at some examples, for the same data sets, we get a four points.

We get two dimensions.

Then we look at the Manhattan distance is just a city block distance.

For example from x2 to x1 you

will go three blocks down then two blocks left.

So the Manhattan distance is 3 plus 2, we get 5, okay.

Of course x1 to x1 itself, is 0.

And you Euclidean distance, as we discussed already, that's the measure.

For Supremum distance, L infinity norm.

You probably can see x2 to x1.

So probably you can see the difference is you first go down 3 blocks for

Manhattan distance, you can see you get a 3 plus 2.

But for Supremum distance you could just see which attribute

would give the maximum distance or maximum difference so

you probably can see the maximum amount is 3 instead of 2 that's why x2 to x1 is 3.

So it's pretty easy to compute, thank you.

[MUSIC]

探索我们的目录

免费加入并获得个性化推荐、更新和优惠。

Coursera 致力于普及全世界最好的教育，它与全球一流大学和机构合作提供在线课程。

© 2018 Coursera Inc. 保留所有权利。

通过 App Store 下载

通过 Google Play 获取