As a discipline, machine learning has been around since the very beginning of computer science, but in the last 15 years, new algorithms such as deep learning have come into their own and enabled widespread success across a variety of applications, ranging from predictive forecasting to image processing.
At the highest level, machine learning offers a different way to create computer programs. Instead of having a person write each step of the program explicitly, machine learning techniques allow parts of the program’s behavior to be learned automatically from data.
In what’s called “supervised learning”, we start with a set of labeled training data. The analogy is a teacher providing a list of examples, each labeled with the correct answer. The assumption is that future data will be similar to the training data, and the job of the machine learning algorithm is to look for correlations between the labeled training data so that it can make predictions about data it has not seen yet.
The model is the end result of training a machine learning algorithm: a program that you can use to make predictions with.
The inputs to a model describe the input data in some way. As an example, when predicting impression to install rates, we provide input signals such as:
- Geographic signals
- Device signals
- Contextual supply signals
- User history from our
- Slot frequencies
- And many more…
In this case, the model output is a numeric prediction indicating the probability that this impression will result in an install.
ML is a tool that when used for programmatic advertising can adjust bidding and targeting decisions based on historical data.
LifeStreet’s Prediction Engine uses ML to continuously learn from campaign data by identifying the patterns and signals belonging to your most valuable users. This information is collected and processed to help our DSP make better bidding, targeting and campaign optimization decisions.
Whether its user engagement, in-app purchases, or user lifetime value, our machine learning models learn what makes a valuable user to you and bids dynamically for each impression opportunity for each user.
Every time a bid request comes in, we run many models to make a final decision on things like:
- If we show the ad, how likely are they to install?
- Which campaign should we show?
- If the user installs the app, how likely will they be to make an IAP?
- Which ad group will have the highest yield?
- How similar is this user to existing user cohorts?
The answers to all of these questions help us bid effectively (i.e., the right user at the right price) to drive advertiser ROAS.