Anomaly detection method based on SVM

author |Mahbubul Alam compile |VK source |Towards Data Science

Introduction to single class support vector machines

As an expert or novice in machine learning , You may have heard of support vector machines （SVM）—— A supervised machine learning algorithm often cited and used in classification problems .

Support vector machines use hyperplanes in multidimensional space to separate one class of observations from another . Of course , Support vector machine is used to solve multi class classification problems .

However , Support vector machine is more and more applied to a class of problems , That is, all data belongs to one class . under these circumstances , Algorithms are trained to learn what is “ natural ”, So when a new data is displayed , The algorithm can identify whether it should be normal . without , New data will be marked as exception or exception . To learn more about single class support vector machines , Please check out Roemer Vlasveld This long article of ：http://rvlasveld.github.io/blog/2013/07/12/introduction-to-one-class-support-vector-machines/

The last thing to mention is , If you are familiar with sklearn library , You'll notice that there's an algorithm for what's called “ Novelty testing ” And Design . It works in a similar way to what I described in the anomaly detection using single class support vector machines . in my opinion , It's just the context that determines whether it's called novelty detection or outlier detection or whatever .

Here is Python Simple demonstration of single class support vector machine in programming language . Please note that , I alternate between outliers and outliers .

---

step 1： Import library

For this demonstration , We need three core libraries - For data disputes python and numpy, For model building sklearn And Visualization matlotlib.

#  Import library 
import pandas as pd
from sklearn.svm import OneClassSVM
import matplotlib.pyplot as plt
from numpy import where

step 2： Prepare the data

I use the famous... From online resources Iris Data sets , So you can practice using , You don't have to worry about how to get the data from where .

#  Import data 
data = pd.read_csv("https://raw.githubusercontent.com/uiuc-cse/data-fa14/gh-pages/data/iris.csv")

#  input data 
df = data[["sepal_length", "sepal_width"]]

step 3： Model

It is different from the adjustment of super parameters in other classification algorithms , Single class support vector machines use nu As a super parameter , Used to define which parts of the data should be classified as outliers .nu=0.03 Indicates that the algorithm will 3% Is specified as an outlier .

#  Model parameters 
model = OneClassSVM(kernel = 'rbf', gamma = 0.001, nu = 0.03).fit(df)

step 4： forecast

The predicted dataset will have 1 or -1 value , among -1 Value is the outlier detected by the algorithm .

#  forecast 
y_pred = model.predict(df)
y_pred

step 5： Filter exception

#  Filter outlier index 
outlier_index = where(y_pred == -1) 

#  Filter outliers 
outlier_values = df.iloc[outlier_index]
outlier_values

step 6： Visual exception

#  Visual output 
plt.scatter(data["sepal_length"], df["sepal_width"])
plt.scatter(outlier_values["sepal_length"], outlier_values["sepal_width"], c = "r")

Red data points are outliers

summary

In this paper , I'd like to talk about a class of support vector machines （One-classsvm） Make a brief introduction , It's a form of fraud / abnormal / Machine learning algorithm for anomaly detection .

I showed you some simple steps to build intuition , But of course , A real implementation requires more experimentation to find out what works in a particular environment and Industry , What doesn't work .

Link to the original text ：https://towardsdatascience.com/support-vector-machine-svm-for-anomaly-detection-73a8d676c331

Welcome to join us AI Blog station ： http://panchuang.net/

sklearn Machine learning Chinese official documents ： http://sklearn123.com/

Welcome to pay attention to pan Chuang blog resource summary station ： http://docs.panchuang.net/