Start with the target task
A ranking model, a recommendation system, and a demand forecast need different views of the same behavior. The target determines which events matter, how far the sequence should extend, and what counts as a useful outcome.
BGraph scopes the dataset around that decision. The result can focus on retrieval relevance, next-category movement, repeat research, or another documented target.
- Ranking and retrieval context
- Recommendation transitions
- Demand and trend features
- Task-specific outcome definitions
Turn journeys into model inputs
Query-to-click paths connect an approved query class with the domain or category selected next. Ordered transitions add the surrounding journey. Graph edges show recurring relationships between entities, while time windows make recency and change visible.
Each representation serves a different model design. Teams can inspect the field definitions and transformation notes before adding the signal to a pipeline.
- Query classes and destination categories
- Ordered event or stage transitions
- Entity edges and normalized weights
- Windowed frequency, recency, and change
Keep evaluation separate from training
Random row splits can leak later behavior into an earlier prediction task. Time-based splits and market-specific slices make the test closer to the decision the model will face.
Coverage notes, missingness, taxonomy changes, and source shifts belong beside the features. Production validation remains the buyer's responsibility, including bias review and permitted-use checks.
- Time-based train and evaluation windows
- Market and category slices
- Distribution-shift checks
- Coverage and quality flags