Anomaly detection with Amazon Lookout for Metrics
This is a guest blog post from Quantiphi, an AWS Advanced Consulting Partner that specializes in artificial intelligence, machine learning, and data and analytics solutions.
We’ve all heard the saying “time is money,” and that’s especially true for the retail industry. In a highly competitive environment where large volumes of data are generated, quick and accurate anomaly detection is critical to smooth business operations and positive customer experiences. However, doing so is easier said than done. Traditional anomaly detection techniques fall short because they can’t efficiently keep up with the growing volume of data. To put this in perspective, one of the largest retailers today collects around 2.5 petabytes of data per hour. Techniques that detect anomalies quickly at this scale and identify their root causes are required for taking effective business decisions.
Traditionally, businesses use dashboards to track metrics or key performance indicators. However, as the number of metrics grow, anomaly identification and remediation using traditional techniques become cumbersome. Therefore, many organizations look at machine learning (ML)-based anomaly detection to overcome this challenge.
In this post, we walk you through a retail example to demonstrate how Amazon Lookout for Metrics, a new ML-based AWS service for anomaly detection, helps accelerate detection and remediation of anomalies.
Anomaly detection in retail transactional data
Let’s dive into an anomaly detection use case we worked on for an online retailer. The retailer generated large amounts of customer transactional data, which serves as a window into their end-customer’s behavior (what products, brands, promotions, and ads they engaged with). For them, quick and accurate anomaly detection in KPIs is imperative for timely remediation, in order to maintain inventory flow and price compliance. As a result, the retailer wanted to ensure that the anomaly insights were actionable and reflected real risks to inventory availability and revenue.
To set up Lookout for Metrics, we first divided the data into regular time intervals. We then set up the detector, specifying the category of every column and the time format of the timestamp, which are mandatory fields. Lookout for Metrics allows us to define up to five measures and five dimensions for continuous monitoring for anomalies. We then trained the detector using historical data and used live data for testing and continuous learning. We uploaded this data to Amazon Simple Storage Service (Amazon S3) regularly, at the time interval specified when setting up the detector.
At each specified time interval, Lookout for Metrics checked for the presence of new data and new anomalies. When it detected an anomaly, it provided two additional insights. First it provided a severity score measuring the magnitude of the anomaly. The severity scores also helped the retailer tune the sensitivity of their model to focus only on the most important events. Second, it revealed a breakdown of the dimensions that contributed to the anomaly, with a percentage contribution from each dimension value, which was useful for determining the appropriate actions to take.
When the retailer applied the detector, we identified a few transactions in which thousands of items were sold at steep discounts. We used insights from the dimensions to quickly learn that these transactions were due to wholesale purchasing and bulk order shipments to large corporations. The retailer was then able to promptly take action to ensure inventory availability for other customers wasn’t impacted.
Comparing Amazon Lookout for Metrics to traditional anomaly detection methods
One of the key benefits of using Lookout for Metrics is a reduction in setup time from days to hours. The setup process is straightforward (see the following diagram). After you define your data source and the metrics you want to track, you can build your first detector model. Lookout for Metrics is then capable of detecting anomalies continuously. In the previous retail example, we detected anomalies in the transactional dataset within 10 minutes of setting up the Lookout for Metrics detector. The process with traditional methods would take 2–3 days. Traditional methods also require highly technical resources and subject matter experts versed in analytics and ML to create and manage custom models. In the case of Lookout for Metrics, the service manages the ML pipeline, allowing you to focus on anomalies, root causes, and actions.
By leveraging a managed service like Lookout for Metrics, anomaly detection is simplified and automated, saving time and effort. Lookout for Metrics helps customers in a wide variety of industries such as retail, ads and marketing, gaming, software and internet, and telecommunications to accurately detect anomalies in their time series data and understand the root cause.
You can use Lookout for Metrics to do the following:
- Detect anomalies in metrics with high accuracy using advanced ML technology, leading to fewer false alarms and missed anomalies, with no ML experience required.
- Get a more holistic view of their business while minimizing disparate alerts. The service groups concurrent anomalies and sends a single alert on related anomalies, which summarizes the what, how, and where.
- Continuously improve accuracy and performance by providing feedback on detected anomalies. Lookout for Metrics incorporates your feedback in real time to learn what is most relevant to your business.
- Prioritize which anomalies to focus on first using the ranked severity score and tune the sensitivity threshold to get alerted on only the relevant anomalies (see the following diagram).
As an Advanced Consulting Partner, Quantiphi comes with immense experience in solving critical business challenges for our customers. We have worked on several anomaly detection engagements. In addition to the retail example discussed in this post, we recently helped an advertisement analytics company improve their offerings by building models to improve the effectiveness of ad campaigns.
With Lookout for Metrics, we aim to expedite the delivery of solutions to detect anomalies in different business processes, thus bringing unprecedented value to our customers.
Amazon Lookout for Metrics (Preview) is available on the AWS Management Console, via the AWS SDKs, and the AWS Command Line Interface (AWS CLI) in the following Regions: US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Ireland), and Asia Pacific (Tokyo). See the AWS Region Table for more details. Request access to the preview today.
Quantiphi is an AWS Advanced Consulting Partner with a deep understanding of artificial intelligence, machine learning, and data and analytics solutions. For more information about Quantiphi, contact them here.
The content and opinions in this post are those of the third-party author and AWS is not responsible for the content or accuracy of this post.
About the Authors
Vibhav Gupta is a Senior Client Solutions Partner at Quantiphi.
Tripurana Rahul is a Senior Business Analyst at Quantiphi.