Challenges with model monitoring
Machine Learning (ML) models predict the world around them which is constantly changing. The unique and complex nature of model behavior and model lifecycle present challenges after the models are deployed.
Cloudera Machine Learning provides you the capability to monitor the performance of the model on two levels: technical performance (latency, throughput, and so on similar to an Application Performance Management), and mathematical performance (is the model predicting correctly, is the model biased, and so on).
- Time series metrics: Metrics measured in-line with model prediction. It can be useful to track the changes in these values over time. It is the finest granular data for the most recent measurement. To improve performance, older data is aggregated to reduce data records and storage.
- Post-prediction metrics: Metrics that are calculated after prediction time, based on ground truth and/or batches (aggregates) of time series metrics. To collect metrics from the models, the Python SDK has been extended to include the following functions that you can use to store different types of metrics:
track_metrics: Tracks the metrics generated by experiments and models.
read_metrics: Reads the metrics already tracked for a deployed model, within a given window of time.
track_delayed_metrics: Tracks metrics that correspond to individual predictions, but aren’t known at the time the prediction is made. The most common instances are ground truth and metrics derived from ground truth such as error metrics.
track_aggregate_metrics: Registers metrics that are not associated with any particular prediction. This function can be used to track metrics accumulated and/or calculated over a longer period of time.
- Tracking accuracy of a model over time
- Tracking drift
Usecase 1: Tracking accuracy of a model over time
Consider the case of a large telco. When a customer service representative takes a call from a customer, a web application presents an estimate of the risk that the customer will churn. The service representative takes this risk into account when evaluating whether to offer promotions.
The web application obtains the risk of churn by calling into a model hosted on Cloudera
Machine Learning (CML). For each prediction thus obtained, the web application records the UUID
into a datastore alongside the customer ID. The prediction itself is tracked in CML using the
At some point in the future, some customers do in fact churn. When a customer churns, they or another customer service representative close their account in a web application. That web application records the churn event, which is ground truth for this example, in a datastore.
An ML engineer who works at the telco wants to continuously evaluate the suitability of the
risk model. To do this, they create a recurring CML job. At each run, the job uses the
read_metrics function to read all the predictions that were tracked in the
last interval. It also reads in recent churn events from the ground truth datastore. It joins
the churn events to the predictions and customer ID’s using the recorded UUID’s, and computes an
Receiver operating characteristic (ROC) metric for the risk model. The ROC is tracked in the
metrics store using the
Use-case 2: Tracking drift
Instead of or in addition to computing ROC, the ML engineer may need to track various types of drift. Drift metrics are especially useful in cases where ground truth is unavailable or is difficult to obtain.
The definition of drift is broad and somewhat nebulous and practical approaches to handling it are evolving, but drift is always about changing distributions. The distribution of the input data seen by the model may change over time and deviate from the distribution in the training dataset, and/or the distribution of the output variable may change, and/or the relationship between input and output may change.
All drift metrics are computed by aggregating batches of predictions in some way. As in the
use case above, batches of predictions can be read into recurring jobs using the
read_metrics function, and the drift metrics computed by the job can be
tracked using the