An End-to-end Guide on Anomaly Detection

This article was published as a part of the Data Science Blogathon.

Introduction

Anomaly is something that is not normal. Any data point which is placed at a distance from all normal data points is an anomaly. Hence anomalies are also called outliers.

Anomaly detection is also called as deviation detection because anomalous objects have attribute values that are different from all normal data objects.

Everyone loves a mystery and that is what anomaly detection is – spotting the unusual, catching the fraud, discovering the strange activity.

What is Anomaly Detection?

The dictionary definition of an anomaly is something that deviates from what is standard, normal, or expected. In the context of data science, an anomaly is a data point that deviates much from the expected fashion.

Types of Anomalies

• Point anomalies – A sole instance of data is abnormal if it is too distant from the remainder. Use case – Detecting MasterCard fraud based on amount spent.

 

• Contextual anomalies – The abnormality is based on the specific context. We can find this kind of anomaly in time-series based data. Use case – Amount spent on petrol on a daily basis throughout the working days is normal but will be found odd when spent on vacation.

• Collective anomalies – A set of data instances put together helps in detecting anomalies. Use case – Someone attempting to copy data from a remote machine to a local host unexpectedly such an anomaly would be flagged as possible cyber-attack.

Causes of Anomalies

1. Data from different classes

An object may be different because it is of a different class. Cases like credit card theft, Intrusion detection, outcome of disease, abnormal test result are good examples of anomalies occurring and identified using class labels.

2. Natural variation

In a Normal or Gaussian distribution the probability of a data object decreases rapidly. Such objects are considered as anomalies. These are also called as outliers.

3. Data measurement and Collection Errors

These kinds of errors occur when we collect erroneous data or if there is any deviation while measuring data.

Where is Anomaly Detection Used?

• Data Breaches
• Identity Theft
• Manufacturing
• Networking
• Medicine
• Video Surveillance

What is Anomaly Detection Good for?

• To find unusual values of metrics in order to find undetected problems.
• To find changes in an important metric or process so that humans can investigate.
• To reduce the surface area or search space diagnosing a problem that has been detected.
• To reduce the need to calibrate or recalibrate thresholds.
• To augment human intuition and judgment.

Drawbacks

• Cannot provide a root cause analysis although it can help in that.
• Unable to provide yes or no answers if there is an anomaly.
• Inefficient to prove that there is an anomaly in the system just that there is something unusual about the metric that is being observed.
• Incapable to detect actual system faults or failures because a fault is not similar to an anomaly.
• It cannot substitute human judgment and experience.
• Incapable to understand the meaning of metrics.

Anomaly Detection Approaches

• Model-based

In this kind of approach a model of the data is built. The objects that do not fit very well are considered as anomalies. For example let us assume that a model is built of certain data which is in the form of cluster. Then anomalies are those data objects that do not strongly belong to the cluster built. In a Regression Model the anomalies are far from predicted value. However, the problem with such approach is when no training data is present to build a model or there is no statistical distribution of data. Therefore, in such cases we require techniques that do not require a data model to be built.

• Distance-based

This approach is based on the proximities. Consider a 2D or 3D scatter plot all the data objects are in one proximity. But Anomalous objects are away from them.

• Density-based

Density of objects is easy to compute especially if a proximity measure between objects is available. Low density objects are those that relatively distant from neighbours. We call such objects as Anomalies.

There are three basic categories for anomaly detection:

1. Supervised Anomaly Detection: In this supervised learning there must a training set for both data objects and expected anomalous objects. We have to observe that there can be more than one anomalous class.

2. Unsupervised Anomaly Detection: For situations where class labels are not available. We can give a score for each object that shows the degree to which the instance is anomalous. We also can observe that if there are many anomalies present which are similar to each other, then we can group them as normal group or the outliers score is low. So, we can say that for unsupervised anomaly detection to be successful it is must that anomalies are distinct.

3. Semi Supervised Anomaly Detection: Sometimes when there is training data with labelled normal objects and score given, but has no anomalous objects, then we can implement the semi supervised anomaly detection to find the anomalies. We use the normal objects to find the anomalies. But, the difficulty is sometimes it is not easy to find that representative set of normal objects using which we have to find out anomalies.

Some more details of the Anomaly Detection Approaches

The Model-based Approach (also called Statistical Method)

Trends and seasonality are two features of time series metrics that disrupt many models. In fact, they are one of two major reasons why static thresholds break. Trends are continuous increase or decrease in a metrics value. Whereas seasonality reflects periodic patterns that occur in a system, usually rising above a baseline and then decreasing again. Common seasonal periods are hourly, daily, and weekly but your systems may have a seasonal period that is longer or even some combination of different periods.

The Distance-based Approach

Distance-based approach classifies a data point as an anomaly (outlier) if its locality is at sparser area. It is calculated from its distance to all points, distance to nearest neighbor, average distance to k Nearest Neighbor or median distance to k Nearest Neighbor.

The Density-based Approach

Density-based approach determines the density of the region. Density-based clustering uses the localized density of points to decide the clusters, rather than using only the distance between points.

DBSCAN

DBSCAN (Density-based Spatial Clustering of Applications with Noise) is one of the extensively used density-based clustering algorithms which is used in the applications of outlier detection. The basic idea behind DBSCAN is that a cluster has to contain minimum number of points within the specified radius. DBSCAN algorithm uses two parameters, minimum points (“minpts”) and epsilon(“eps”) to perform clustering.

We will take an example to detect outlier using Z score and Interquartile range

First by using Z score

Formula for Z score is (Observation – Mean)/ Standard Deviation

z = (X — μ) / σ

dataset= [11,10,12,14,12,15,14,13,15,102,12,14,17,19,107, 10,13,12,14,12,108,12,11,14,13,15,10,15,12,10,14,13,15,10]

outliers=[]
def detect_outliers(data):

    threshold=3
    mean = np.mean(data)
    std =np.std(data)


    for i in data:
        z_score= (i - mean)/std 
        if np.abs(z_score) > threshold:
            outliers.append(y)
    return outliers

outlier_pt=detect_outliers(dataset)

outlier_pt

[102, 107, 108]

Now by using Interquartile Range which is 75% – 25% values in a given dataset

quartile1, quartile3= np.percentile(dataset,[25,75])

print(quartile1,quartile3)

12.0 15.0

3.0

print(lower_bound_val, upper_bound_val)

17.5, 19.5

Conclusion

Anomaly Detection is a very vast concept; in this guide, we learned about some basic concepts of anomaly detection

• We learnt what anomalies are, their causes and types.
• We also discussed approaches to detect anomalies and how they are used in real world.
• We also saw how they can be used to make better models.
• Anomalies are extremely important to be detected before building a really good model.
• They need to be removed from the data before we can use the data. 
• It helps us to alert ourself of any unwanted event that needs to be detected well in advance to prepare ourselves to face situations.

I hope this guide on Anomaly Detection helps you to gain a fair understanding of how anomaly detection works.

The media shown in this article is not owned by Analytics Vidhya and is used at the Author’s discretion.