From a batch of histories to predictionsCompare each history item with the candidate, turn each comparison into a relevance score, and use the scores to pool the history into one candidate-specific interest vector before predicting a click.
Build a target-aware embedding moduleTargetAwareEmbedding looks up IDs, scores each real history item against the candidate, masks padding, and pools. ClickPredictor adds a head so one loss trains the table, scorer, and head together.
Candidate relevanceTo rank C candidates, repeat each history C times, flatten the pairs into the batch axis, score them with the same model, and reshape the results back to [B, C].
Summarize within categories before combining themFirst average items within each category using their relevance, then weight each category summary by its own relevance to the candidate and add the summaries together.
Train one model from IDs to click logitsHierarchicalRecommender connects ID lookup, two-level pooling, and a prediction head, so one binary cross-entropy loss trains every part.
Multiply features to express interactionsA cross layer multiplies the original features by a learned weighted sum of the current ones and adds the previous result back, so stacking layers builds higher-degree feature interactions.
Use cross features in a complete predictorRun a cross branch and a deep branch on the same features, concatenate them, and map the result to one logit per example; one loss trains both branches and the head.
Build matrix feature crosses with DCN v2A matrix cross layer gives each feature its own learned mixture of all features, at a cost of D² + D parameters per layer, and multiplies it by the original features to build interactions.