Beginning Anomaly Detection Using

Figure 8-18.  Historical sales figures of a product

Chapter 8   praCtiCal Use Cases of anomaly DeteCtion

---

317

•  Investigating the results and feedback analysis as it effects the 

business

•  Operationalizing the model used in the day-to-day activity of the 

business

In particular, we are very interested in how the models are built and in what type 

of models we should be using. The type of anomaly detection algorithm used affects 

pretty much everything that we are trying to get out of this anomaly detection strategy. 

This in turn depends on the type of data available, as well as whether the data is already 

labeled or identified. One of the things that will affect the decision to figure out what 

type of anomaly detection will work best for the specific use case is whether it is a 

point anomaly, contextual anomaly, or a collective anomaly. We are also interested in 

looking at whether the data is an instantaneous snapshot at some point in time or if it 

is continuously evolving or ever-changing, real-time, time series data. Also important 

is whether the specific features or attributes of the data are categorical or numerical, 

nominal, ordinal, binary, discrete, or continuous. It is also very important to know if the 

data is being labeled already or if some sort of a hint is provided as to what this data is, 

since it could steer us in the direction of supervised, semi-supervised, or unsupervised 

algorithms.

While the technologies and algorithms are available to be used, there are several key 

challenges to implementing an anomaly detection approach based on deep learning:

•  It’s hard to integrate AI into existing processes and systems.

•  The technologies and the expertise needed are expensive.

•  Leadership needs to be educated on what AI can and cannot do.

•  AI algorithms are not natively intelligent; rather, they learn by 

analyzing “good” data.

•  There is a need for change in “culture,” especially in large companies.