Machine Learning Use Cases in Marketing
simMachines supports use cases for marketing spanning customer lifecycle predictions, dynamic predictive customer segmentation, customer experience management, sales forecasting, trending and analysis.
Churn Prevention
Problem
Large international telecommunications provider needed to reduce churn for pre-paid phone cards.
Solution
simMachines deployed dynamic predictive segmentation to enable proactive churn prevention.
- simMachines created predictions of who was likely to defect based on past defectors. Every prediction’s Why factors are different.
- We then created clusters of similar predictions. Each cluster was assigned a corresponding action.
- In one case of this example, the prediction is based on a group of customers with a family plan who were recently divorced with children no longer in the household.
- The corresponding action for this prediction is: “Offer these customers an individual plan.”
Result
30% reduction in churn amongst contacted customers
Sales Forecasting and Customer Lifetime Value
Problem
Specialty retailer needed to calculate the future sales and lifetime value of customers.
Challenges
Client wanted a technology solution that could learn and adjust over time, as well as accurately predict and classify customers based on limited data.
Solution
The solution predicted future purchase volume and net income, as well as classified the customer base into groups based on value and the associated “Why” factors that define them.
Result
Predictions with 95% accuracy compared to ground truth demonstrate that the algorithm can provide an automated calculation of a new or existing customer’s future purchase volume, net profit and LTV.
Machine Learning Fraud Detection
Problem
Top 3 financial institution wanted to speed up ability to implement ML fraud detection solutions for e-commerce clients, enable continuous learning, and expose the factors driving the fraud.
Challenges
ML methods did not continuously learn or expose the Why factors. Increased competition was creating a need to upgrade speed and service value to client.
Solution
Ensemble approach using gradient boosting and similarity-based machine learning.
Result
70% increase over current fraud detection performance. Time to implement reduced from 6 weeks to 2 weeks. Continuous learning enabled and the Why factors reported/available behind every prediction.
Shopper Reactivation
Problem
Leading specialty apparel brand retailer needed to improve precision of targeting prior customers for reactivation to improve campaign efficiency and sales.
Challenges
Client needed a modeling tool that could produce models more accurately and quickly than their existing solution.
Solution
Propensity modeling application generated grid of multiple models in days and ran selected highest performing algorithm in minutes on full US population producing predictive segmented outputs for campaign execution.
Result
50%+ improvement over existing model, and the WHY allows for heightened understanding of key drivers to affect treatment and offer.
Dynamic Predictive Customer Segmentation
Problem
E-commerce pet supplement provider wants to shift purchase volume from Amazon to direct, online channels.
Challenges
Client had limited analytic resources and need to generate look-a-like prospecting segments based on existing customers that could be activated in a direct to consumer campaign quickly and effectively.
Solution
The solution generated dynamic predictive prospect segments on a 3rd party file based on existing customers, generating hundreds of segments that were narrowed to five representing 65% of the target population within 2 weeks. Weighted predictive buying factors associated with each segment enabled tailored creative, messaging and offers through email and digital channels.
Result
Dynamic predictive segments immediately achieved higher than industry average click through rates in first DTC campaign ever run and is foundational to future DTC strategy for two brands.
Digital Identity Resolution
Problem
Large scale DMP needed to improve linkages across devices associated with single users.
Challenges
Client had millions of unconnected devices it needed to accurately connect for improving campaign reach and frequency objectives.
Solution
The solution developed used similarity search to analyze a high volume of records quickly to pair two or more devices together with a high degree of accuracy quickly.
Result
The solution was able to comb through over a quintillion pairing combinations in 48 hours and pair over 800,000 out of 1 million device records together with ~95% accuracy compared to known ground truth. Daily updates can be run on a growing volume of records in under 2 hours.
One-to-One AB Testing Measurement
Problem
Large scale agency interested in more efficiently matching test and control groups with greater precision and understand ad effect at a one to one vs. stratified sample level for client measurement projects.
Challenges
Current statistical approaches result in higher level aggregations of test vs. control comparisons to generate conclusions regarding test vs. control ad effect at a group level.
Solution
The solution applies similarity-based machine learning methods to match test and control populations enabling campaign drop outs from the test group to be automatically matched and removed from the control group. For analysis, individual one to one twin pairs enable the clustering of positive vs. negative ad effect, with the most predictive weighted factors associated with cluster revealed in weighted order of importance. This customer centric approach supports the automated selection of similar audiences to receive the same successful ad treatment.
Result
The solution provides a simple user interface to create and test pairing accuracy and then generate one to one ad effect cluster analysis to inform campaign adjustments in under 8 hours per project with a high degree of accuracy. Look-a-like audiences can be identified and output in minutes.
B-to-B Intelligent Lead Modeling
Problem
Large scale telecom provider that offering multiple services to businesses needed new solution that could quickly predict likely to purchase / upgrade with detailed explanations to inform sales reps as to Why.
Challenges
No current state solution existed to accomplish this with the required precision and explainability.
Solution
The solution leveraged explainable machine learning methods to predict likely to purchase and upgrade business prospects and customers as well as reveal the weighted factors associated with the predictive purchase behavior. The top factors can be pushed into lead sheet applications used by sales reps in the field.
Result
Predictive explainable machine learning model performs with ~70% precision compared to ground truth, while providing Why factors in rank order of importance to inform the sales team’s messaging and offers.
Automated Online Product Purchase Classification
Problem
Large scale business intelligence platform provider wanted to be able to automate the classification of millions of products sold online to pre-set categories.
Challenges
The client was manually assigning products to categories since the products had a wide range of naming conventions ranging from 2 to over 25 words including numbers.
Solution
The solution initially trained a classifier on known ground truth and was able to automate product labeling. Then for never before seen products the solution leveraged unsupervised clustering to group similar products together and allow users to assign labels. The classifier was then re-trained with the update product category assignments. This process enables the client to work through over 100 million un-categorized products for reporting.
Result
96% accuracy in precision was achieved for the classification model within 1 week. Unsupervised clustering was created in an additional 1 week’s time enabling the model to be updated and assign product category’s to products at a high rate of speed.
Automated Contextual Content Segmentation
Problem
Large scale digital ad tech provider needed to generate contextual-content based segments from web pages faster on a global basis in multiple languages to ingest into audience segmentation products.
Challenges
The client was limited in their current rules based system in terms of scale and was looking for machine learning technologies that could handle their scale, speed, content clustering and labeling needs globally.
Solution
The solution applied unsupervised clustering on a sample of bid stream optimization data that ingests tens of thousands of online bid events per second. The clustering was able to group like URL pages together for easy labeling and the level of granularity desired, which can be adjusted by the user. A simply user friendly interface optimizes the cluster labeling. A classifier was built to automate the assignment of URL pages to known content categories.
Result
Clustering was able to support completely accurate assignments quickly for content segment labeling within two weeks in multiple languages. The classifier is able to achieve ~95% accuracy in assigning pages to content categories to eliminate personnel time. The solution increases the value of audience segmentation products the client sells to advertisers.
Survey Modeling
Problem
Market research firm needed to segment survey responses in a variety of ways to uncover valuable insights for client product and marketing planning.
Challenges
The client was leveraging analysis tools that made clustering of survey responses cumbersome and limited in its insights.
Solution
The solution leveraged unsupervised clustering to group together similar survey responses based on questions answered as well as question level views of responses for analysis.
Result
The application is able to reveal, in order of importance, the most important factors associated with each survey response, overall and by question, to speed the time to insight and the depth of insight the analysis team has at its disposal for client projects.
Automated Customer Segment Assignment
Problem
Large scale subscriber-based telecom provider assigns customers to segments based on survey responses but wasn’t happy with its level of accuracy and speed.
Challenges
The client was leveraging a wide range of in-house regression modeling and rule- based tools to assign survey responders to its customer segmentation architecture, finding it difficult to achieve the results it wanted efficiently.
Solution
The solution leveraged similarity based supervised clustering to assign responders to pre-existing segment categories.
Result
In hours the classifier was able to outperform existing models generated in-house over several months. The automation achieved enables the solution to continuously learn and reveal Why it assigned a customer to a segment – based on the most important characteristics associated with the prediction.
Media Spend Forecasting
Problem
Client needed to automate media spend forecasts to eliminate significant manual labor costs. Given the explainability factors of simMachines, the client was interested in leveraging our technology.
Challenges
Data is fairly limited for forecasting and requires significant labor to produce.
Solution
Similarity based machine learning enabled the successful integration of consumer spend data combined with previous media spend historical data to forecast next 12 month spend and reveal key indicators of spend.
Result
Client evaluating with clients the delivery of competitor spend with new competitive forecasts by product for roll-out.
Call Prioritization Scoring
Problem
An online health insurance organization needed to automate lead call prioritization based on likely to convert and LTV calculations with greater precision.
Challenges
The client had open source models running for likely to convert, but didn’t have LTV or the ability to leverage explainable prediction factors to drive personalization of offer.
Solution
The solution included two explainable machine learning algorithms, one for conversion and one for LTV that were able to improve precision and provide explainability at a transaction level.
Result
Models matched neural net performance on likely to convert but provided significant LTV lift as well as transaction level weighted factors for prioritizing call routing to the most important leads to increase conversion.
Dynamic Predictive Audience Propensities
Problem
Marketing services partner with thousands of global clients needed to automate and upgrade propensity models for data products and client data projects to a rate of hundreds per week from ~10 per week.
Challenges
The client was faced with regression modeling limitations that wouldn’t enable the broad creation of propensity models at the scale needed.
Solution
Propensity modeling application runs in 6 minutes on full US population per model and produces segmented outputs in 5 clicks
Result
Cost reductions are significant for the client in reducing model creation time as well as run time.
Dynamic Predictive Customer Segmentation
Problem
Leading agency is dealing with escalating data science resource costs, while needing to drive customer segmentation projects for large scale clients without increasing their prices, faster and with greater throughput.
Challenges
The client was using a two step K-means clustering and predictive modeling process that took months to complete and was too slow and lacked detailed resolution for micro-segmentation needs.
Solution
The application of dynamic predictive segmentation is able to run the same micro-segmentation in single pass prediction and clustering in hours, enabling analysts to achieve reduced costs and greater resolution.
Result
Time to market for segmentation projects is reduced by 90%, and reduces hours spent by over 60%, saving the client significant project costs. Additionally, micro segments are generated at whatever degree of resolution the client needs. Direct comparison between methods demonstrated that the same micro-segments could be automatically generated at far lower cost and also inform creative, message and offer with specific Why factors associated with each segment.
Customer Acquisition Credit Worthiness
Problem
An international bank client provides loans to small businesses. Our client needed a custom, predictive engine that would help quickly determine the credit worthiness of a small business owner.
Challenges
Many systems and methods have been implemented to determine if a person is likely to pay back a loan or if they are capable of owning a credit card. The problem with these methodologies is that they are based on general models and theories that may not closely match the reality of a specific financial institution.
Solution
simMachines implemented an algorithm that could achieve the following:
- Predict the type of new customer being contacted: A++, A+, A, B, C, D
- Predict the number of times an operator would have to call the customer in order to collect payments
Result
simMachines successfully created and deployed the solution in less than two weeks.
Audience Digital Measurement
Problem
How do you analyze a massive stream of internet data? DataPulse is a joint venture between JAS Global Advisors and simMachines formed around combining unique data assets for internet traffic monitoring with simMachines similarity-based machine learning technology.
Challenges
Ninety-five percent of all web analytics is based on data harvested through websites controlled by a client. By adding code, tracking cookies or pixels, customer behavior is tracked across the website. These valuable insights are designed to optimize traffic within the website and push conversion. The challenge for broader promotion response and consumption analytics, however, is to establish a pattern across different websites and channels.
Solution
DataPulse’s Market Intelligence Suite can cover any individual or group of domain names, making it ideal to compare traffic within industries, competitors or affiliates. Augmented with our insights, driven by similarity based machine learning, DataPulse can provide unmatched analytical insights and details about Internet presence and activity for your industry, your brand, or even your competitors.
Result
With insights being generated in a real-time manner, time specific results drive the Market Intelligence suite. Zooming in on instances, or taking a long-term view, allows for trending and predictive analysis. Data Pulse traffic has a proven, strong correlation to Alexa Web Site Traffic Statistics but is available in real-time.
Sales, Product Usage and Retention Forecasting
Problem
Top 3 retailer needs to be able to forecast product level sales, usage, and ongoing retention of customers who have purchased the product starting with the high end electronics category and associated VR products.
Challenges
SKU level analysis and forecasting requires machine learning technology to handle the volume of data and analyses/forecasts
Solution
Similarity based machine learning enabled the trending of monthly product users by those predicted to churn vs. stay loyal to the product at a very granular level of machine generated customer segments.
Result
Client prioritizing SKU level products and data requirements to enable optimal forecasting.
E-Commerce Fraud Detection
Problem
Top 3 financial institution wanted to speed up ability to implement ML fraud detection solutions for e-commerce clients, enable continuous learning, and expose the factors driving the fraud.
Challenges
ML methods did not continuously learn or expose the Why factors. Increased competition was creating a need to upgrade speed and service value to client.
Solution
Ensemble approach using gradient boosting and similarity-based machine learning.
Result
70% increase over current fraud detection performance. Time to implement reduced from 6 weeks to 2 weeks. Continuous learning enabled and the Why factors reported/available behind every prediction.