On the application of cluster analysis in work

Estimated reading time ：8min

Reading suggestions ： This paper describes the application of clustering , Some experiences in actual combat are summarized , I hope it helped you .

Solve the pain ： What is clustering ？ What value does it have in data analysis ？ How to cluster ？ What are the advantages and disadvantages of each method ？ I hope you will come to this article with these questions .

00

preface

mention 「 clustering 」, Do you think 「 Birds of a feather flock together 、 Birds of a feather flock together 」 Well , This is the essence of clustering . At work , Cluster analysis is still common , Have you ever encountered the following problems ？

• User partition ： When the product reaches a certain level , It is hoped that users can be divided into sub groups with obvious characteristic attributes , Implement strategies for specific groups .
• Product portfolio ： With the enrichment of the company's products , We hope to divide the products into different combinations according to value and liquidity , And make marketing strategies for different combinations .
• Anti cheating judgment ： The user's operation on the product belongs to normal behavior , And some people seek benefits , Use machine and other methods to brush data , For this kind of cheating , How do we explore it ？

When you encounter problems like the above , Cluster analysis can be used . below , Xiaohuolong will start from the practical application of clustering , Unveil the veil for everyone .

01

What is clustering

Clustering is based on certain rules （ for example ： Distance ）, Divide the data set into different clusters , Make the similarity of individuals in the same cluster as large as possible , The similarity between different clusters should be as small as possible . The performance in the data is , Individuals with similar eigenvalues are more likely to get together , On the contrary, the possibility is small .

Tips： There are many ways to calculate distance , List several common

Euclidean distance

Manhattan distance

Cosine distance

「 clustering 」 There is no prior knowledge 「 Unsupervised algorithm 」, It corresponds to 「 classification 」 It is 「 Supervised algorithms 」. Some students may have some dizziness , For example, it is easy to understand ：

Classification problem

Men must have Adam's apple , Therefore, as long as people with Adam's apple , It's male . among , The Adam's apple is transcendental knowledge , Through this feature , People can be classified into men and women .

Clustering problem

For our ancestors , There is no distinction between men and women , But through features , Some people can be found to have Adam's apple 、 The other part didn't , Therefore, the group with Adam's apple is named male .

Here is a case of clustering in the work , Help people enhance their cognition .

02

Practical application of clustering

stay 「 Strategy Promotion 」 In the direction of , Clustering can cluster the target groups , And expand the potential population . Therefore, the application directions include but are not limited to ： Precise market promotion 、 Expansion of potential target groups, etc . below , Take a case ：

A car company , Before the Spring Festival , For possible target groups 「 Blue collar workers in first tier cities 」 Send wechat ads to push , Help improve the drainage of car enterprises . But after clustering by user behavior , Explore 「 White collar workers in second tier cities 」 Also have the potential to buy a car . therefore , In the process of promotion , Expanded the target group's delivery range , Improve drainage effect .（ The specific application steps are shown in the figure below ）

03

Common clustering methods and their advantages and disadvantages

After introducing the function of clustering , Do you have any friends who want to know , What are the ways of clustering , How to achieve , And the advantages and disadvantages of each method , Now little fire dragon makes a summary for you .

There are many kinds of clustering models , It can be roughly divided in the following three directions ：

1. The processing power of the model ： Ability to handle different distribution shapes ; Ability to handle outliers ; The ability to handle big data .

2. Whether the model needs preset parameters ： Whether it is necessary to provide category quantity, etc .

3. Model requirements for data input ： Does the order of data input affect the model ; Whether the model has requirements for the type of features, etc .

According to the above classification , Clustering models can be divided into the following categories ：

Because of space , Little fire dragon selects three commonly used models to elaborate （ The red part of the picture ）.

1、 A hierarchy based approach - Hierarchical clustering

1. Model principle

Hierarchical clustering consists of two ways ,「 Agglomerative hierarchical clustering 」 and 「 Split hierarchical clustering 」. Agglomerative hierarchical clustering is a common method in hierarchical clustering , The core principle is , The initial assumption is that each individual is a class , Each iteration merges the closest points , When all the points are merged into one class or the stop condition is satisfied , Then terminate the model iteration , It's a bottom-up approach . The corresponding split hierarchical clustering , It iterates in the opposite bottom-up way , Final output .

2. Model flow

With 「 Agglomerative hierarchical clustering 」 For example ：

step 1： Each point as a class , Calculate the distance between two points ;

step 2： Merge the closest points together , Forming new classes , And calculate the centroid of the class ;

step 3： Repeat the first 1、2 step , Until the conditions are met .

3. The advantages and disadvantages of the model

[ advantage ]

• Strong model interpretation ability
• There is no need to set K（ Can be used as K-means Cluster exploration K A priori algorithm ）
• about K-means Non spherical points that you are not good at are handled better

[ shortcoming ]

• Time complexity is high , Slow operation
• Can't solve non convex object distribution

2、 Partition based approach - K-means clustering

1. Model principle

Some of its ideas are similar to 「 Agglomerative hierarchical clustering 」, But before starting the model , You need to input the number of final clusters in advance K, Then initially select several points as the center of mass , Then merge the similar points , And form a new center of mass , The principle of iteration is 「 The inner point of the class is close enough , Distance between class points is far enough 」, Until the number of clusters finally matches .

as for K The choice of , Can pass 「 The law of the elbow 」、「 Profile factor 」 And so on , Not much here , I will write about it in the following articles .

2. Model flow

step 1： Random selection K Objects , As K The initial centroid of a cluster （ because K Is random , therefore K-means The results of each cluster are different ）;

step 2： Objects close to the center of mass , Merge into one class , And iterate out a new center of mass ;

step 3： Repeat the first 1、2 step , Until you are satisfied K Conditions for clusters .

3. The advantages and disadvantages of the model

[ advantage ]

• Low time and space complexity , Run fast

[ shortcoming ]

• Sensitive to the principle of initial centroid
• Sensitive to noise , Will be biased
• It is easy to have local optimal solutions
• Can't solve non convex object distribution

3、 Density based approach - DBSCAN clustering

1. Model principle

Neither of the above two methods can deal with irregular shape clustering , and DBSCAN Density based method can solve the problem well , And it is friendly to noise data . Its core principle is more popular , Is to draw a circle through each dot , Contact with surrounding points , If certain rules are met, they will be included in this category , Until the iteration requirements are met .

The rules are mainly two parameters , One is the maximum radius of a circle （eps）, The other is that the circle should contain at least a few points （MinPts）. Here are DBSCAN The effect picture formed by clustering .

2. Model flow

step 1： From any point p Start ;

step 2： Find and merge p Objects within the diameter range ;

step 3： If p As the core point , And form a cluster ; If p For the boundary point , namely ： It does not meet the requirements of minimum accommodation points , Find the next object point again ;

step 4： Repeat the first 2、3 step , Until all objects are overwritten .

3. The advantages and disadvantages of the model

[ advantage ]

• Satisfy arbitrary shape clustering
• Insensitive to noise

[ shortcoming ]

• The clustering result is directly related to the initial setting value
• Because the value is fixed , Therefore, it is not very friendly to the sparse distribution of different objects

The above is the content sharing of this issue , I hope it can give you a clear understanding of cluster analysis .