- First steps to thinking about defining a machine learning problem
- Some thoughts on building useable data sets
- How to train and test ML
Advanced measurement opportunities have increased marketing visibility like never before. Innovative technologies including AI, Machine Learning, and algorithmic solutions have propelled the marketing world forward, whether in the form of custom packaging or out of the box platform offerings.
It’s an exciting time for Machine Learning (ML), in particular. Knowledge of ML has been growing over recent years, but now we finally have both the scale of data and the ability to put it into use. With developing technologies from companies like Amazon, Google and Microsoft it’s never been easier to leverage. The hard part now is determining what to dive into first! Here are a few fundamental guidelines to follow when considering a custom approach to ML:
Step 1. Define your problem
A few examples for marketers include ad copy generation, product recommendation, and personalized experiences. In addition to identifying what you’re looking to solve with ML, it’s also a good time to pick the right type of model such as GANs (generative adversarial networks) for generating imagery, RNNs (recurrent neural networks) for text analysis and generation and CNNs (convolutional neural networks) for analyzing imagery.
Step 2. Building your data set for ML
Traditionally, machine learning works best when you provide good features or inputs and attach relevant labels with the data. For example, if you are building a network for image recognition, your features/inputs might be the pixels of the image, and the label would be the name of the object.
Step 3. Training and Testing
If you don’t have specific data for training, validation, and testing, you can split your overall data set into 60% training, 20% testing, and 20% validation. This training data should solely be used for training the model. The testing set should be used to identify the effectiveness of your model – is it accurate? The validation set helps to determine that you’re not overfitting the data while training. Overfitting is when your model knows the immediate correct answer for a given input, but can’t handle other inputs outside of the training set. For example, imagine if you memorized the answer for a question on a math exam, but the professor increased the variable values by 1. Now your answer doesn’t work at all for those inputs.
Step 4. Use in production
At this point, you’ve tested and validated your model and feel comfortable with the results it’s outputting, so it’s time to put it to use. Even with proper testing, the real world may not work completely with your model. If you’re building something like a product recommendation system for your website, you may just have new products that you need to train your system to recognize. You may also have real user data that can be used to improve your system’s ability to recommend products that are most likely to lead to cross- and up-sell opportunities. Keep training and improving your model, with care to not overfit.
If you don’t have the resources to develop your own ML solutions, there are numerous platforms used within the marketing industry that are increasingly powered by advanced measurement technologies – think Oracle, Facebook, Factual, and Place IQ to name a few. These help to paint a full picture of the consumer journey and have increased the capabilities of reaching users where and when it matters most. Curating audiences based on previous customer data, in addition to location-based targeting has opened the door to a more accurate pool of users who are currently in-market. This also allows you to customize content to speak to the individual consumer, whether it be in efforts to gain recursion or a conquest a customer from a competing brand. In doing this, you are ensuring marketing dollars are working harder for the brand and influencing real-time decisions during the most pivotal points in the consumer journey.