The machine learning process (ML)
The chart below illustrates the use of an initial testing budget. At each stage, our machine learning algorithms learn and optimize the data to maximize the result at the start of your work with Hybe.
We recommend to start developing machine learning algorithms before campaign starting*, but we can also start the user engagement process without ML. Below is a general description of how the optimization process works.
Event 1
At the first stage, if we don't have enough data to develop ML algorithms, we launch our campaigns without using ML models. This process is required to collect sufficient sampling for ML models training.
If we have access for non-atributed data, it is possible to use this data for creating look-a-like audiences, app and user targeting for our user acquisition campaigns. This helps save the exploration budget until we have enough data to connect the first ML model to optimize installation probability.
ML optimization of creatives works from the beginning and helps to scale the most effective creatives by lowering the priority for the worst ones.
Examples of events:
- Installation
Event 2
Once we've gathered the initial events to run ML**, we begin using ML algorithms to optimize campaigns. During this process, our media buying experts also apply various manual optimizations to ensure maximum performance and quickly block unwanted results***.
Examples of events;
- Completion of training
- Registration
Event 3
Once the initial ML training steps are complete, we continue to optimize and incorporate ML models for additional funnel events.
Examples of events are:
- Reaching level 5.
- KYC data completed.
Event 4
The final step in ML optimization will be to connect a model trained on a target event, such as a paying user.
Examples of events:
- Depositor.
* Share non-attributable events from your MMP before launching your UA campaign.
** We need ± 800 initial events before we can activate the first ML algorithms
*** High eCPIs, creative optimization, WL/BL optimization, targeting by user characteristics, separating top and average targeting, etc.