A Curious and Terrified Marketer’s Start to AI and Predictive Analytics

Summary

One marketer's journey into machine learning and predictive analytics. This blog explores various techniques within predictive analytics highlighting their relevance in marketing. Follow a journey that demystifies the process of executing a propensity model with AI tools, discussing steps from choosing an AI platform to model deployment, emphasizing the accessibility of predictive analytics tools for marketers without extensive statistical knowledge.

By Sarah Threet, Marketing Consultant at Heinz Marketing

In grad school, I played around with sentiment analysis: the contextual mining of text that identifies and extracts subjective information in source material to help marketers understand the social sentiment of their brand. I found it interesting to be able to quantify something so qualitative. One of my favorite aspects of marketing is the combination of psychology and behavioral analysis with formal data. I am someone who typically equally weighs my own decision-making based on a combination of things that are subjective with things that are factual – or “head and heart decision-making” as some may say. 

I am by no means any expert or authority to speak on either predictive analytics nor artificial intelligence. However, I understand that most marketers are in a similar position, so I wanted to write about my process trying to learn the basics of this subject. 

Artificial intelligence is one of those subjects that frankly scares the crap out of me – not in the typical “the robots are going to take my job” sense, but I am fearful of technology moving at a faster pace than the laws in place to keep it ethical. Admittedly, I am someone who worries that people will use this power for evil rather than for good. In the context of marketing, I see how it would be advantageous to understand it better. Humans tend to be naturally fearful of things that they don’t understand, so I am attempting to get over this fear by learning more about it. 

(By the way, I used AI to answer many of my questions around this topic, fact checked the information provided, and then dove deeper into some research.)

First thing’s first, a few definitions:

What is predictive analytics in marketing? 

Predictive analytics in marketing is the use of current and/or historical data with a combination of statistical techniques (data mining, predictive modeling, and machine learning) to predict future events and behavior related to marketing. It can help marketers to optimize their strategy, target the right customers, and improve campaign outcomes.  

I quickly learned that there are many kinds of techniques and models used within predictive analytics – from regression to time series models, clustering algorithms to neural network models, decision trees to classification models… it’s easy to get lost in all of it as interesting as it is.  

So which techniques and models are used most frequently within the context of marketing, and how do they apply to marketing? 

Techniques and marketing examples in predictive analytics: 

Identification Modeling: Involves building models that classify individuals into predefined categories or segments based on specific attributes and characteristics.  

• It can be used for demographic profiling, targeted advertisements lead scoring, segmentation, and retention.  
• Logistic regression, decision trees, and neural networks are commonly used techniques for identification modeling. 

Propensity Modeling: Used to predict the likelihood or propensity (“inclination or natural tendency to behave in a particular way”) to take action or exhibit a particular behavior.  

• It’s often used in marketing to identify potential customers who are most likely to respond to a marketing campaign or make a purchase. Propensity models are typically binary classification models that predict the probability of an event occurring (e.g., making a purchase, clicking on an ad).  
• Logistic regression and gradient boosting machines are commonly used for propensity modeling. 

Customer Segmentation: Involves dividing customers into groups based on similar characteristics or behaviors.  

• This allows marketers to tailor their strategies and messages to specific segments, improving the effectiveness of marketing campaigns.  
• Techniques such as clustering algorithms (e.g., K-Means, hierarchical clustering) and latent class analysis are used for customer segmentation. 

Churn Prediction: Identifies customers who are at risk of leaving or discontinuing their relationship with a business.  

• By predicting churn in advance, marketers can implement targeted retention strategies to prevent customer attrition.  
• Techniques such as logistic regression, decision trees, and survival analysis are commonly used for churn prediction. For example, by fitting a logistic regression model to historical data with known churn outcomes, you can identify which factors are most influential in predicting churn. 

Customer Lifetime Value (CLV) Modeling: Predicts the future value that a customer will generate over their entire relationship with a business.  

• By understanding the potential value of different customer segments, marketers can allocate resources more effectively and prioritize acquisition and retention efforts.  
• Techniques such as regression analysis and machine learning algorithms are used for CLV modeling. 

Recommendation Systems: Analyze customer data to provide personalized recommendations for products or services.  

• These systems help marketers improve cross-selling and upselling opportunities and enhance the customer experience.  
• Techniques such as collaborative filtering, content-based filtering, and matrix factorization are used for recommendation systems. 

Market Basket Analysis: Examines patterns of co-occurrence or association among products purchased by customers.  

• It identifies which products are frequently bought together, enabling marketers to optimize product placement, promotions, and bundling strategies.  
• Association rule learning algorithms such as Apriori and FP-Growth are used for market basket analysis. 

Sentiment Analysis: As mentioned earlier, analyzes text data (e.g., customer reviews, social media posts) to determine the sentiment or opinion expressed towards a product, brand, or topic.  

• It helps marketers understand customer perceptions and sentiment trends, enabling them to tailor messaging and address customer concerns effectively.  
• Natural language processing (NLP) techniques, including machine learning algorithms and lexicon-based approaches, are used for sentiment analysis.  
• I hadn’t realized that sentiment analysis was considered a component of predictive analytics, but it is in the context of forecasting future trends or outcomes based on the analysis of sentiment data. 

As you may have noted, many of these examples can include machine learning algorithms. I wanted to focus in on one and better understand how I could use AI tools to execute on it. Propensity modeling is the closest to how I imagined using predicative analysis in the first place.  

First, propensity modeling can use Gradient Boosting Machine (GMB). For those of you who don’t know what the heck GBM is (like I didn’t): it is a type of machine learning algorithm that belongs to the ensemble learning family, used for both regression and classification tasks. Essentially it builds a series of sequential decision trees with each tree aiming to correct errors made by the previous trees.  

Here is what I’ve learned about how GBM works from a high level: 

1. Base Learner: The algorithm starts by fitting a simple model to the data, often a single decision tree, which serves as the initial prediction. 
2. Residual Calculation: After the first model is trained, the algorithm calculates the residuals, which represent the difference between the predicted values and the actual values in the training data. 
3. Sequential Learning: The subsequent models are trained to predict the residuals instead of the actual target values. Each new model focuses on capturing the patterns or relationships in the data not captured by the previous models. 
4. Ensemble Building: The predictions from all the models are combined to produce the final prediction. In regression tasks, the final prediction is the sum of the predictions from all the individual models. In classification tasks, it’s often the result of averaging the predicted probabilities or using voting. 

The key idea behind gradient boosting is to iteratively improve the model by focusing on the errors made by previous models. By sequentially adding new models that complement the strengths of the existing ones, GBM tends to produce highly accurate predictions. 

Some popular implementations of Gradient Boosting Machines include: 

• Gradient Boosting Classifier (GBC): Used for classification tasks, where the goal is to predict the categorical class labels of data points. In sentiment analysis, these tasks determine the sentiment of a piece of text, classifying it as either positive, negative, or neutral. 
• Gradient Boosting Regressor (GBR): Used for regression tasks, where the goal is to predict a continuous numerical value rather than a categorical label. In sentiment analysis, the numerical score represents the intensity or strength of sentiment expressed in a piece of text. This is often referred to as sentiment intensity prediction. 

In plain terms, this makes a lot of sense. Getting into the nitty gritty of it becomes much more complicated. Unfortunately, I currently do not have access to a tool for personal testing, and even if I did, I’d have to brush up on statistics, and maybe learn some coding, right? A tool like DataRobot can supposedly support this kind of analysis, and they do offer a free trial, so maybe I’ll follow up with a second blog where I try my hand at this. In the meantime… 

Executing a propensity model with GBM 

On a high level, if you wanted to build a propensity model with the objective of learning which customers will engage most with your marketing campaign, these are the steps you’d take to execute. Please be mindful that I have not tried these steps yet myself, but will play with these exact steps when I try to execute for my next blog: 

1. Choose an AI Platform or Tool: Select an AI platform or tool that provides pre-built models and user-friendly interfaces for building predictive models. In this case, I might try my hand with a tool like DataRobot which offers a simplified interface and automates workflows for building predictive models without extensive statistical knowledge. 
2. Data Preparation: Upload your historical marketing data. The platform may offer tools for data preprocessing, such as handling missing values and encoding categorical variables, which can be done with minimal manual intervention. 
3. Model Selection: Choose the pre-built GBM model from the platform’s model repository. These models are often optimized and ready to use out-of-the-box, eliminating the need for manual model selection or hyperparameter tuning. 
4. Data Splitting and Training: Utilize the platform’s automated data splitting functionality to divide your data into training and testing sets. Train the GBM model on the training data with a few clicks, without the need to specify complex parameters or algorithms. 
5. Model Evaluation: The platform will automatically evaluate the trained model’s performance using appropriate evaluation metrics. It will provide intuitive visualizations and summaries of the model’s performance, making it easy to interpret the results. 
6. Propensity Scoring and Customer Ranking: Once the model is trained and evaluated, use the platform’s prediction capabilities to generate propensity scores for each customer in your dataset. The platform may offer features for ranking customers based on their propensity scores automatically. 
7. Threshold Selection: With guidance from the platform’s recommendations or best practices, select an appropriate threshold for classifying customers as responders or non-responders. 
8. Model Deployment: Deploy the trained propensity model into production directly from the platform. The platform will handle the deployment process making the model accessible for use in targeting marketing campaigns. 
9. Monitoring and Iteration: Monitor the model’s performance over time using the platform’s monitoring tools. If necessary, retrain the model periodically with new data to maintain its accuracy and effectiveness. 

Conclusion – for now… 

Supposedly, by leveraging AI platforms or tools that have simplified workflows and automation capabilities, you can execute the process of creating a propensity model with a GBM without much need for personal statistical knowledge – that’s cool if accurate! These platforms streamline the process, enabling users to focus on leveraging data-driven insights for effective marketing decision-making. 

Shoot, if it’s as easy as my blog makes it sound, then any average marketer with an inclination for data analysis can do this. I’m sure it’s not that easy, but maybe it is! If tools like DataRobot offer free trials, then I will follow up in another blog testing the waters. Nevertheless, this brief dive into AI and predictive analysis has me interested to learn more, even if it means getting over some humps that might not come as intuitively for me. In that way, AI feels more helpful than just scary. 

Please let me know if you find this blog useful! While not as detailed about the how-to, maybe it also helped you get over the initial knowledge gap by reading about my learning process.