Advice on anomaly detection algorithm for customer support data

Question

Background

In the customer support department, when customers contact us we categorize the interaction into the issues customers are having. For example, customer X contacted us about an issue “Order was not delivered”, customer Y contacted us about “Package was broken”, and customer Z contacted us about an issue “Inquiring about order status” and so on. There are about 100 different issues that we monitor.

In our reports, we aggregate them with the counts and rank them based on their count. For example top 15 issues in a day or a week or a month.

Number of orders and amount of complaints are correlated meaning If there are more orders, there will be more complaints. If there are fewer orders there will be fewer complaints.

Aim

Aim is to detect the change in rank. For example when an issue jumps from rank 15 to rank 8. We want to understand when this happens (when an issue starts affecting more customers than it usually does) and get alerts of the rising of the issue.

To achieve this, I would like to detect anomalies in time series. My question is what machine learning technique or algorithm do you recommend to detect anomalies in time series data?

Answer 1

I would start with doing this on weekly data. Each category, for each week would be an independent instance. Then the compute features of interest that you mention:

• number of orders
• number of issues
• issue ranking
• etc.

IsolationForest is suitable for this kind of features (possible mix of categorical and numerical).

Since weeks are treated as independently the problem has been converted from a time-series to not-time-series. This is a simpler formulation, hence why it is a good starting point. To introduce a little bit of time modelling, you can add the change over previous week for each feature (in number of issues, orders etc.).