Billing Anomaly Detection with Artificial Intelligence (AI) and Machine Learning (ML)

Until now, the system enabled you to detect anomalies in the bills before bill completion using the Bill Tolerance feature. Now, in addition, Oracle Revenue Management and Billing enables you to use artificial intelligence and machine learning to detect anomalies in the bills and bill segments before the bill completion. The different steps involved in the machine learning process are:

• Data Preparation

• Model Generation

• Data Prediction

These three steps are referred as ML jobs in the system. An ML Job goes through various statuses in its lifecycle. We have shipped the following three batches to execute the above three steps:

• C1-MLDPR - This batch is used to extract the historical bills and their bill segments which are later used to create a machine learning training model. It transforms the historical bills and their bill segments into materialized views.

  You can specify the extraction date range to fetch the bills and bill segments which are created within that date range. We recommend you to extract the historical data of at least last three years. This batch provides an option to extract the data incrementally. If you want to extract the data incrementally, then you need to set the Incremental Data parameter to Y and specify the extraction from date and cutoff days. This option should be used when you want to extract historical data from a particular date till the cutoff date (i.e. system date - cutoff days).

  Ideally, you should use this option when you want to use a static from date and a dynamic end date each time the batch is executed. However, if you do not want to extract the data incrementally, then you need to set the Incremental Data parameter to N and specify the extraction date range (i.e. from and to dates). You can extract the historical bills and their bill segments either separately or together. However, if you have voluminous data, we recommend you to extract the historical bills in the first batch run and then extract the historical bill segments in the next batch run.

  This batch also enables you to filter the dataset based on the bill cycle and division. It is a single-threaded batch. Ideally, you must execute this batch every three months to refresh the dataset of the machine learning training model.

• C1-MLMCR - This batch is used to create a machine learning training model using the historical data extracted through the C1-MLDPR batch. A machine learning training model is a process in which a machine learning (ML) algorithm is fed with sufficient training data to learn and analyze using the dataset. It enables you to process large volumes of data to detect anomalies and test correlations while searching for patterns across the dataset. At present, you can build a model using the following approach:

  • BUILTIN - Used when you want to build a model far ahead in time before the anomaly detection.

  Once a machine learning training model is created, it is stored in the database as a database object. This batch is a single-threaded batch. Ideally, you must execute this batch every three months to create a new machine learning training model with the latest dataset.

• C1-MLPRE - This batch is used to detect anomalies in the current bills and their bills segments based on the patterns observed in the machine learning training model. At a time, you can either detect anomalies in the bills or their bill segments in a single batch run. You must execute this batch after generating bill segments and bills through the billing batches.

  While using AI ML for anomaly detection, you should not enable the Freeze and Complete option for the bill cycle. This is because you will then have the option to make the required corrections based on anomaly detection before freezing and completing the bills and their bill segments. A To Do notification is created to summarize anomaly detection for a given batch code, batch number and batch rerun number.

  While executing the C1-MLPRE batch, you need to specify the bill segment generation or bill completion batch control and its batch run number whose bills or bill segments you want to consider for anomaly detection. You can either specify the batch run number or set the incremental run flag to true during the C1-MLPRE batch execution. This batch is a multi-threaded batch.

At present, the following data mismatch is considered as anomalies by the machine learning process:

• Data mismatch in the bill segments (i.e. SQI or amount mismatch)

• Data mismatch in the bills (i.e. an additional price item is charged in the current bill, a price item (which is billed earlier) is not charged in the current bill, and bill amount mismatch)

To implement this AI ML feature, the following two screens are introduced in this release:

• ML Jobs - Enables you to track the status of different ML jobs.

• Diagnostics Central - Enables you to view and close the detected anomalies in the bills and bill segments. After reviewing the anomaly, you can either take any one of the following action:

  • Fix the anomaly in the bill or bill segment and then close the anomaly. In this case, while closing the anomaly, you must set the Anomaly field to Yes to indicate that it is an anomaly.

  • Ignore the anomaly in the bill or bill segment and then close the anomaly. In this case, while closing the anomaly, you must set the Anomaly field to No to indicate that it is not an anomaly.

Parent Topic: Oracle Revenue Management and Billing Financial Services Business Processes