Sign up for our newsletter and get the latest big data news and analysis.

10 Tips for Building Effective Machine Learning Models

Machine learning continues to gain headway, with more organizations and industries adopting the technology to do things like optimize operations, improve inventory forecasting and anticipate customer demand. Recent research from the McKinsey Global Institute found that total annual external investment in AI was between $8 billion and $12 billion in 2016, with machine learning attracting nearly 60 percent of that investment. What’s more, organizations with senior management support for machine learning and AI initiatives reportedly stand to increase profit margins anywhere from 3 percent to 15 percent.

Despite this momentum, many organizations struggle with simple machine learning best practices and miss out on the benefits as a result. Following are 10 tips for organizations who want to use machine learning more effectively.

1. Look at the data

Preparing a training data set takes time, and mistakes are common. Before you build a model, review your data. This way, you can quickly discern if you have the right data in the correct form. Use PROC Print with OBS=20 in Base SAS®, the Fetch action in SAS® Viya™, and the Head or Tail functions in Python to see and “touch” the observations.

2. Slice the data

There’s usually an underlying substructure in data, so slice your data as you would a pizza, building separate models for each slice. Once you’ve pinpointed a target, build a shallow decision tree and then create separate models for each segment. Keep in mind that clustering algorithms often do not derive as good of segments when you have a target.

3. Use simple models

Building complex models to extract as much as you can from your data is crucial; however, simple models are easier to deploy and make the process of explaining results to key business stakeholders easier. Build simple, white-box models using regression and decision trees, and use a gradient boosting or ensemble model to confirm how your simple models are performing.

4. Detect rare events

Machine learning often requires the use of unbalanced data, so correctly classifying rare events can be difficult. To counteract this, construct a biased training data set by oversampling or undersampling. Your training data will be more balanced, and the higher ratio of events will help your algorithm learn to better isolate the event signal. Another rare event modeling strategy is to use decision processing to place greater weight on correctly classifying the event.

5. Combine lots of models

Data scientists commonly use algorithms like gradient boosting and random forests to automatically build lots of models. However, these models may generalize well, and some algorithms will fit better than others within specific data boundaries. Overcome this hurdle by combining different modeling algorithms. Place more emphasis on the modeling method that has the overall best classification or fit on the validation data, but also consider including two weak classifiers that do a better job of capturing specific spaces of your training data.

6. Deploy your models

Too often, model deployment can take weeks or months, with some models never even getting deployed. To get models into production efficiently, you must determine the business objective; access and manage the data; and develop, validate, deploy and monitor the model. Also, use tools that automatically capture and bind the data preparation logic so data engineers or IT staff have a blueprint for implementation. Also consider creating standardized analytical data marts to foster data replication and reuse.

7. Autotune your models

Before building a machine learning model, algorithm options called hyperparameters need to be assigned. Autotuning can help pinpoint suitable hyperparameters accurately and quickly. Use Grid Search (we recommend using a Latin hypercube to search across the hypermeter space) to autotune your parameters, by searching through a manually specified subset of the hyperparameter’s space, guided by a performance metric.

8. Manage change

Changes or additions to features can cause entire models to fundamentally change. To compensate for this reality, calculate population stability indexes and characteristic monitoring statistics to measure model decay at frequent intervals. Monitor the ROC and lift for classification models, set up jobs to detect model decay, and schedule re-training jobs at specific intervals.

9. Balance generalization

Generalization is the learned model’s ability to fit well with new, unseen data. The key with generalization is maintaining balance, where you shift between models with high bias and those with high variance. So, if you have high variance error, use more data or subset features. If you have high bias error, use more features. Using the right evaluation metric is essential to selecting models that will generalize well on new data.

10. Add features

Training data sets require several example predictor variables to classify or predict a response. In machine learning, the predictor variables are called features and the responses are called labels. Infuse your training data with several readily available features to help derive a better fitting model. For example, try infusing your model with customer feedback data, if what customers are saying is highly predictive. You can also infuse your models with purchase history data, since what customers have bought can be helpful information when building purchase propensity or next best offer models.

About the Author

Wayne Thompson is the Chief Data Scientist at SAS. He is one of the early pioneers of business predictive analytics, and he is a globally recognized presenter, teacher, practitioner and innovator in the field of predictive analytics technology. His current focus initiatives include easy-to-use, self-service tools for business analysts, deep learning and cognitive computing. Wayne received his PhD and MS degrees from the University of Tennessee.

 

Sign up for the free insideBIGDATA newsletter.

Leave a Comment

*

Resource Links: