Implementing a machine learning solution in a business can seem like a daunting task, especially when looking at some of the different options that are available. Artificial Intelligence – while a buzzword in many industries – is complicated to actually adopt when it comes to both training models and implementing solutions that are understandable by non-experts.
In this article, I will introduce a new category of machine learning provider, Automated Machine Learning, and discuss how these relatively simple to implement tools can be used to quickly add value to a business.
First, what exactly is meant by Automated Machine Learning, or AutoML?
AutoML providers take on the difficult tasks of creating the framework on which machine learning models are created, allowing users to simply upload their data and then check the results of the automatically generated models. It is, in a word, automatic. This can be compared to a savvy excel user creating a well-designed Excel spreadsheet that only requires data inputs to perform advanced calculations.
For the most part, if you have a basic understanding of a spreadsheet program like Microsoft Excel or Google Sheets, you will have the expertise necessary to take advantage of an AutoML solution.
In the current marketing environment, it is easy to suffer from data overload, so when utilizing artificial intelligence to analyze data, it is important to be clear on the desired outcomes. Some areas where AutoML shines include prediction models (predicting future behavior) and classification models (identifying types of people/things/businesses from).
In a sales organization, for example, an AutoML model could be created to identify which new sales leads were most likely to convert into future sales. In a retail environment, a classification model might help managers identify categories within a customer base to allow the segmentation of customers using a custom-built machine learning model. These customer segments could then be tracked and targeted using different advertising promotions based on their profile and behaviors.
Most AutoML providers make it simple to import data, but the data will need to meet a few basic requirements. Data is often imported from a spreadsheet program, so this is a good starting point for most first-time users.
When creating a data set, it is important to have the following information. First, the data set should have identifiers that label each unique case (or record) in the data set. A case can be any unit of interest but is often a person or a business. For example, a golf course might have a database of each of its members which tracks money spent in the pro-shop, rounds of golf played, guest passes used, and other things of interest to that course.
Each row in the data set will identify a golfer and each column will represent information on some variable of importance. This is all pretty simple. I quickly put together the table below to show what I mean.
In this table, the golfer # is the unique identifier for each case, and then the information in the column and row intersection represents the value for that golfer on that variable.
Chances are, this is already something most students and business professionals are comfortable with, but the key here is to identify key “cases” such as potential customers for a business, as well as key “variables” which can be used to give the model more information about the case as it runs its analysis.
In the above example, it might be useful to predict if a golfer is likely to renew their membership for the next year. In this case, the analysis would be a predictive analysis, and the target variable (predicted variable) would be membership renewal. If this analysis was run a few months prior to membership renewal (after having built the model using this year’s data), it would allow the golf course to target customers which are in danger of exiting, and then apply a marketing intervention to try to retain them.
These are some interesting data problems that an AutoML solution will handle rather easily. First, while zip codes are listed as a number, they really represent discrete categories. As a zip code number increases, it doesn’t mean there is “more” zip code, it just represents a different category. Second, there is a text-based categorical variable for Yes and No. Most AutoML providers are set up to handle this type of information.
An AutoML provider such as Akkio will generate output designed to help you understand your data. An example of Akkio output for churn prediction is included below. Notice the following:
Here we can see the performance of the model when tested against a portion of the data that was held back for this purpose. This model is correct about 80% of the time, and you can see more detail about how well it predicts the “no” and “yes” classes.
While it’s generally true that more accuracy is better - model performance is only ever as good as the underlying data and patterns. In this case, the model is quite useful for identifying customers who are at risk of churn. You can also see that tenure, monthly & total charges, and device protection plans are the biggest predictors of churn.
This is a key step for any marketer. If the time is spent to build a machine learning model, it needs to be used strategically by the business. In the example of the golf course above, it is relatively simple to see a workflow where a machine learning model could be updated seasonally before membership renewal season and then used to target retention spending on existing members. In another business, a model might be trained and used daily, depending on the predictive or the classification needs of that business - the best solutions integrate directly with your data storage solution making it easy to push and pull data from the model.
The key part of the AutoML solution is its user-friendliness and accessibility to non-expert users. Students who are just learning data analytics can use AutoML solutions to further their education and as an entry to the machine learning world, while businesses, which don’t wish to invest significantly in machine learning personnel and resources, can use AutoML cheaply and quickly to implement impactful marketing solutions.
AutoML can do much more than just predictive or classification analysis. Segmentation, for example, is an area where AutoML can be quite useful. In the retail environment, a segmentation model might help managers identify clusters within a customer base to allow the segmentation of customers using a custom-built machine learning model. These customer segments could then be tracked and targeted using different advertising promotions based on their profile and behaviors
Another interesting area where AutoML can help is in text analytics. Akkio’s NLP module can take unstructured text data and turn it into insights that are otherwise hidden in free-form responses. This type of analysis is often used in call center data, social media data, and customer survey data.
Finally, one last interesting application of AutoML is optimization. This type of analysis is often used in marketing mix modeling or other situations where businesses need to understand the optimal combination of inputs to produce a desired output. In this case, the outputs are often financial (ROI, sales, etc.), but they could be non-financial as well.
In a world where data is ubiquitous, the businesses that succeed will be the ones which are able to make use of that data to improve their processes. AutoML provides a powerful tool which can be used by businesses of all sizes to improve their marketing, sales, and operations.
From churn prediction in a call center to retail customer segmentation to social media sentiment analysis, AutoML can be used to improve just about any business process. As the technology continues to develop, the number of potential applications will only grow. Be on the lookout for new and interesting ways to use AutoML in your business.