Preparing user journey data
Before attribution modeling begins, Funnel processes user journey data to structure it for training. Each journey consists of a time-ordered sequence of events, such as impressions, clicks, and conversions, captured for both converting and non-converting users. These sequences include:
Timestamps to reflect the order and timing of events
Channel metadata including source, campaign, and device information
Conversion outcomes (positive or negative signals)
This structure allows the model to evaluate both effective and ineffective journeys, which is critical for distinguishing correlation from causation.
How the LSTM model is trained
Funnel uses Long Short-Term Memory (LSTM) networks to learn patterns from historical user behavior. The model evaluates thousands of journeys to estimate how likely each sequence is to result in a conversion. The training process includes the following steps:
Data preprocessing to clean and normalize journey inputs
Sequence encoding to feed events, metadata, and timing into the LSTM
Prediction of the conversion likelihood for each journey
Credit assignment to allocate attribution weights to each touchpoint based on its contribution
Funnel uses counterfactual modeling to understand the effect of removing or changing a touchpoint. This helps isolate the role each touchpoint plays in influencing the final outcome.
Attribution weighting logic
Attribution credit is assigned based on how much each touchpoint contributes to a successful conversion. The model does not assign credit based on pre-defined rules. Instead, it uses observed patterns from actual user behavior to determine influence.
For example:
A click that appears early in a journey may receive more or less credit depending on what follows.
Multiple touchpoints from the same channel may have diminishing returns.
Interactions with short or long delays between them are weighted differently depending on modeled outcomes.
Model refresh and maintenance
To ensure attribution results stay up to date with changes in marketing performance and user behavior, Funnel retrains the model on a regular schedule.
Daily updates refresh the attribution weights and predictions using the latest data.
Periodic full retraining occurs when there are major changes in tracking setup, channel mix, or feature engineering.
This approach balances modeling stability with adaptability to real-world campaign changes.