Machine Learning for Marketing: From Separate Models to One Platform
In a talk at the Data Science Innovators Conference 2023, Meta data science manager Jen Wang described how machine learning supports marketing decisions. She contrasted per-campaign models with a centralized, modular platform that reduces audience overlap and improves budget allocation.
By The AI Enterprise Magazine Team4 min read
Key takeaways
- Tie marketing to outcomes. Data science defines metrics such as sales, return on expense and customer lifetime value.
- Per-campaign models don't scale. They cause audience overlap, inconsistent measurement and poor adaptability.
- Think in layers. A clustering model feeds a treatment model, with feedback loops, developed as parallel modules.
- Test everything. A/B testing and causal inference decide whether a new approach replaces the baseline.
- Collaboration is required. Business and technology teams must work together at each step.
In a talk at the Data Science Innovators Conference 2023, "From Data to Dollars: Using Machine Learning to Drive Marketing Success," Jen Wang, a data science manager at Meta, explained how data science supports marketing and how the machine learning setup should change as a company grows. Wang opened with her own experience of an ad on Facebook, noting that Facebook explains why a user sees an ad: advertiser choices and the user's activity. The goal of marketing, she said, is to show the right people the right ads at the right time.
How does data science help marketing?
Wang said marketing teams make many decisions, and data science can inform each one. Data scientists help define metrics and measurement frameworks that tie marketing to business outcomes, such as sales, return on expense and customer lifetime value. They analyze historical data to predict customer behavior and value, which supports better resource allocation. They also cluster customers by shared characteristics and behaviors, which allows more personalized campaigns, and they help automate marketing operations.
She described a cycle in which marketers and data scientists begin with the current business as usual and its limits. They analyze past data for trends and then build models for segmentation, personalization, optimization and attribution. A/B tests and causal inference check whether a new approach beats the current one, and if it does, it becomes the new baseline. Wang stressed that each step needs close collaboration between business and technology teams.
How do company size and culture shape the approach?
Wang said a more data-driven culture and flatter structure enable faster iteration, while centralized marketing and data science teams can bring coordination and consistent messaging. There is no single right answer, she said; the aim is a solution that fits the company's objectives and culture.
What does a small-company setup look like, and where does it break down?
For a company just starting in-house data science, Wang described building separate models for each marketing program, organized around the marketing funnel from awareness to consideration to purchase. Marketing defines which audience is eligible for each program, and data scientists build models such as brand targeting, email response and retargeting models. Smaller companies, she said, often lean on managed services because they lack in-house machine learning operations expertise.
That approach is easy to start and gives quick results, she said, but its weaknesses show as a company grows:
- Audience overlap. Separate models can target the same people repeatedly with different ads.
- Inconsistent standards. Different teams measure differently, making apples-to-apples comparisons hard and risking poor budget allocation.
- Isolation. Little cross-team communication leads to inefficiency.
- Poor adaptability. Solutions built for specific situations are hard to update, for instance when privacy regulation changes.
What does a scalable, modular platform look like?
For larger organizations, Wang proposed translating business questions into overarching machine learning themes instead of building per program. In her framing, marketing programs share three questions: "how we categorize our customers," what treatment each customer should receive given what is known about them, and how to measure outcomes and keep improving.
A centralized platform treats all marketing programs as treatment options and compares the expected outcome of each for a given customer. Wang described layered models: a customer clustering model that feeds a treatment model, plus models that send outputs back as feedback. Modules can be developed in parallel, she said, which reduces overlap, makes performance comparisons consistent and helps allocate budget.
She added that most companies she knew still rely heavily on human effort to choose campaigns, channels, messages and creatives. Even with newer data science methods, she said, strong integration between business and technology remains necessary.
What risks and practices did Wang flag?
Wang warned about the need for high-quality data, privacy and ethics, and models that are interpretable and unbiased. Her practical advice, drawn from her experience and her network of marketing data scientists, included:
- Develop a clear strategy that aligns technology capability with business objectives and KPIs.
- Understand the target audience before launching a campaign, and prioritize personalization.
- Invest in marketing technology that automates operations, and put data governance in place for accuracy, security and compliance.
- Encourage regular communication between business and technology teams.
- Keep a measurement system with A/B and multivariable testing.
What came up in Q&A?
An attendee said organizations need to create a data science mindset first, and Wang said, "I totally agree with that." Asked about pricing, she said some companies use price elasticity models to set prices that generate more net profit. She added that marketing mix modeling is a more traditional method she had not used much personally. It suits full-funnel planning and budgeting, she said, but granular daily decisions about which customers to target call for more advanced machine learning.
Where can you hear more about AI and data in marketing?
Marketing and customer data strategy are recurring topics for enterprise data leaders at COLLIDE 2027 on Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center in Atlanta. Details are at /events/collide/, and registration is coming soon.
Frequently asked questions
How does machine learning help marketing?
Wang said it helps define metrics, predict customer behavior and value, cluster customers for personalized campaigns, and automate operations.
Why do per-campaign machine learning models stop working as a company grows?
They can target the same audience repeatedly, use inconsistent standards that distort budget decisions, and are hard to update when conditions such as privacy rules change.
What is a modular marketing ML platform?
A centralized system that frames marketing around shared questions, with layered models such as customer clustering and treatment selection developed in parallel.
What risks should marketing data science teams watch?
Wang cited data quality, privacy and ethics, and interpretable, unbiased models.
This article is based on remarks by Jen Wang, Data Science Manager, Meta at Data Science Innovators Conference 2023 on April 12, 2023. More coverage: More AI Connect Events.
Filed under: AI Use Cases & Applications, AI Strategy & ROI
