Choosing Between Discriminative vs Generative Models
Choosing Between Discriminative vs Generative Models
The differences have significant downstream implications
The differences between discriminative and generative models have significant downstream implications for your data and AI strategy.
Today I want to provide you some analogies to quickly illustrate the differences between these types of models, then I’ll explain how these differences impact your strategic decision-making process, and how you should go about implementing your model choices.
Spoiler Alert: This content was meant to be included within the pages of my upcoming Wiley book on data and AI strategies for growth. I decided to cut it from my manuscript and share it directly with Convergence members here, but if you’d like to be notified when the book becomes available, be sure to jot your name down on this list.
Illustrating with examples
I’ll start off with a simple example. You know how Gmail quickly separates spam from non-spam emails? What it’s doing is “classifying” or “segregating” the email data into two buckets. A discriminative model does exactly that. It identifies patterns that differentiate one type of data from another. When you’re training a discriminative model, you’re teaching it to understand or discover patterns that are distinct to each data type. So, a good model here is one that can detect spam mail from a mile away and throw it in the bin.
When trained well, your generative model can actually create a text that resembles a spam email. Woohoo! 🥳
Generative models are different. Instead of just identifying the differences between the data types, they’ll actually attempt to “understand” the inherent characteristics of spam and non-spam data. This is analogous to profiling. These models could be used to clue in on what goes into creating a spam or a non-spam email and - with each iteration - they’ll get more accurate and more precise. So, when trained well, your generative model can actually create a text that resembles a spam email. Woohoo! 🥳
Choosing between model types
When choosing between discriminative vs generative models, you need to closely examine the specific objectives and requirements of the data strategy you’re building. For instance, if the strategy is focused on identifying trends in customer behaviors or segmenting markets, discriminative models might be more appropriate. Conversely, if the strategy involves innovating new products or simulating data for stress testing, then generative models would be more suitable.
Let’s say you have all your data resources in one place, and you want to build a forecasting or predictive feature atop them. If you wanted to predict whether a customer will churn, or forecast sales for the next quarter, then the discriminative model would be a better choice because it will help you predict, forecast, cluster, or classify based on your requirement.
If, however, you’re looking to add a new feature but you don’t have the data resources in place support it, then you’ll typically start by gathering data. In this data-gathering phase, you can start off by using a LLM + RAG to “generate” synthetic data that you can use for the pilot phase while you put your data-gathering mechanism or pipeline in place. Here, you’ll be using what we refer to as the generative model. In fact, there are third-party companies that solely work on helping you generate data through simple queries or prompts.
Tip: When you’re using LLMs to generate data, be sure to provide detailed prompts to help you generate close to real-world data.
Action steps for implementing your model choices
Now that you understand the fundamental differences between discriminative and generative models and how they fit into differing strategic needs, the next step is to evaluate your current projects and data initiatives. To do that, start by asking yourself:
Which projects could benefit from more precise classification or prediction? Consider using discriminative models for these to improve accuracy and efficiency.
Where might you innovate or create with data? For projects needing innovation or simulation, look into employing generative models to foster creativity and extend your data capabilities.
Assess your data readiness: Do you have the necessary data to support these models? If not, it might be time to explore synthetic data generation or to bolster your data collection strategies.
Consult with experts: If you're unsure about the best approach, consider reaching out to an advanced data science consultant who can provide you with the insights and guidance you need to tailor to your specific circumstances.
By actively applying these considerations, you can more effectively align your AI strategy with your business objectives to gain a clear competitive edge. Keep in mind, the right model will be able to support your current needs and adapt to future challenges and opportunities.
I hope this post was helpful! And, if you’d like to be notified when book launch festivities begin, be sure to jot your name down on this list.
Warm regards,
Lillian Pierson
PS. If you’re looking for marketing strategy and leadership support with a proven track record of driving breakthrough growth for B2B tech startups and consultancies, you’re in the right place. Over the last decade, I’ve supported the growth of 30% of Fortune 10 companies, and more tech startups than you can shake a stick at. I stay very busy, but I’m currently able to accommodate a handful of select new clients. Visit this page to learn more about how I can help you and to book a time for us to speak directly.
PPS. If you liked this email, go ahead and forward it to a friend!
Discover untapped profits in your marketing efforts today!
Tired of guesswork and inefficiencies in your marketing strategy?
Make sure to download my B2B Marketing KPI Scorecard & Pipeline Tracker before I move it behind a paywall.
It’s packed with insights and processes specifically designed to help B2B tech companies to boost marketing ROI and speed up sales processes — no strings attached! Grab yours today and start seeing the benefits by tomorrow.
Through best-in-class marketing strategy, leadership, and advisory support, I help B2B tech startups and consultancies achieve consistent and predictable revenue growth, all without the full-time CMO price tag.