Senior Data Scientist
If you are a data scientist like me, you may have rolled your eyes more than once at the casual and frequent mention of “machine learning” in advertising. There is far more nuance to this topic than a single phrase can capture. Consumer advertising to the masses boasts ML and “AI” capabilities of cars, healthcare systems, appliances and more, yet these advertisers meet no consumer demand to go deeper on machine learning methodologies. But here in the marketing industry, where machine learning has been a prevalent name drop in just about every business’ sales and marketing materials for the past ten years, the fantastic variety of capabilities is worth delving into. Whether you’re a marketer, agency media buyer, or data scientist, arming yourself with a true knowledge of machine learning capabilities will only improve your campaign outcomes.
A wealth of machine learning algorithms exist; gradient boosted trees, random forests, neural networks, support vector machines and more. As marketing technology and strategies have evolved, so has the application of these underlying algorithms that are powering them. One example is how data scientists are using ensemble methods to improve how marketers engage with customers. While ensembles are not a new concept, they are being used in new ways for acquisition, retention and reactivation efforts.
Each algorithm has its own strengths and weaknesses. For example, logistic regression is often best for linear effects, while tree-based methods are better for capturing interactions. Through years of experience, Alliant’s data scientists advocate that combining, or ensembling, algorithms produces stronger results than any single one on its own. Measuring something several different ways, rather than just once, is likely to decrease variability in the predictions of the models and provide more accurate results.
So what are the different ways that you can combine machine learning models? There are three common approaches:
Averaging: The final predicted probability (score) of a desired action is the average of the predictions (scores) of all the models.
Voting: Each model separately predicts whether each record has a high probability of a desired action (e.g. being in score group 1). The final judgement is the majority vote of these predictions.
Additional Modeling: Model scores from several models are used as predictors in a final model. This allows a stronger model to have a greater say in the final score than when all models are averaged equally.
Once the chosen models are scored, they can be combined to achieve various objectives. In the context of acquisition, you could either focus on refining or expanding your audiences for targeting. Refinement emphasizes improving the quality of prospects, only selecting those that are in the top score groups of every model. Although this narrows the quantity of available prospects, they will be of higher quality. At Alliant, we often use a form of voting to combine models for refinement, identifying the best 3 to 4 models. From there, we will look at the intersection of the top score groups to select the most qualified audiences for our clients’ campaigns.
Expansion on the other hand, puts emphasis on quantity, selecting prospects that fall into any of the top score groups.
Each model will identify quality audiences a bit differently, so with this approach you may be able to find customers that would have otherwise been missed. Whichever approach you choose, you will have increased flexibility in the final output of your solution, compared to just using one model.
Creating and combining collections of models to produce better results than any one model will certainly require investment and a test and learn approach. One thing to not forget: no matter how advanced your workflows, it is beneficial to maintain a human touch. Strategizing with data scientists to determine model objectives, data requirements and success metrics will not only help the modeler understand objectives and nuances, but will result in higher performing campaigns. Not to mention, we have all seen examples of fully autonomous projects gone awry. Many platforms machine learning capabilities often live in a “black box” setting, and with your marketing dollars at stake, a set it and forget it should be avoided. Lastly, if flawed input data is flowing into the system it will be subject to problems. Having a team to monitor and validate results is necessary. At least I hope, otherwise us data experts may be in trouble.
Interested in learning more about how you can partner with Alliant’s data scientists to build custom data solutions? Contact us at any time! Our team has been on an analytic evolution, enabling the data scientists to take predictive modeling to new places and ultimately creating stronger solutions for our partners.