How synerise/recsys2025 works
This repository is a behavioral-profiles challenge baseline plus its organizer-side evaluation pipeline. Participants generate per-user representations from interaction history; the pipeline validates those embeddings, prepares task-specific targets, trains a separate model per task, and aggregates scores. The baseline’s event-feature aggregation is only one representation approach. The README also documents open and hidden downstream tasks. Some modules are indexed but not excerpted, so their internal relationships are omitted unless the index or README establishes them.
Main entry points: train_runner.py Training runner · create_embeddings.py Profile generation · validate.py Validation entry · task_constructor.py Task setup
Components 24
Profile baseline
- Profile generation
baseline/aggregated_features_baseline/create_embeddings.py - Feature aggregation
baseline/aggregated_features_baseline/features_aggregator.py - Feature calculators
baseline/aggregated_features_baseline/calculators.py
Challenge data
- Data directory
data_utils/data_dir.py - Event loading
data_utils/utils.py
Validation
- Validation entry
validator/validate.py - Embedding loading
validator/loading_validator.py - Embedding checks
validator/embeddings_validator.py
Task training and scoring
- Training runner
training_pipeline/train_runner.py - Task definitions
training_pipeline/tasks.py - Task setup
training_pipeline/task_constructor.py - Target calculators
training_pipeline/target_calculators.py - Target data
training_pipeline/target_data.py - Training data module
training_pipeline/data_module.py - Task model
training_pipeline/model.py - Metric calculators
training_pipeline/metric_calculators.py - Metric results
training_pipeline/metrics_containers.py - Score aggregation
training_pipeline/metric_aggregator.py - Training logging
training_pipeline/logger_factory.py - Logging configuration
training_pipeline/train_logging_config.py
Other components
- Participants
- Interaction data
- Embedding files
- Evaluation scores
Connections 24
- Participants → Profile generation: generate profiles
README.md - Interaction data → Profile generation: provides events
README.md - Profile generation → Event loading: loads events
baseline/aggregated_features_baseline/create_embeddings.py - Profile generation → Feature aggregation: builds features
baseline/aggregated_features_baseline/create_embeddings.py - Feature aggregation → Feature calculators: computes features
baseline/aggregated_features_baseline/features_aggregator.py - Participants → Embedding files: submit representations
README.md - Embedding files → Validation entry: loads for evaluation
- Training runner → Validation entry: validates and loads
training_pipeline/train_runner.py - Validation entry → Embedding loading: loads arrays
validator/validate.py - Validation entry → Embedding checks: checks embeddings
validator/validate.py - Training runner → Target data: reads targets
training_pipeline/train_runner.py - Training runner → Task setup: constructs task
training_pipeline/train_runner.py - Task setup → Task definitions: dispatches task
training_pipeline/task_constructor.py - Task setup → Target calculators: configures targets
training_pipeline/task_constructor.py - Training runner → Training data module: sets up data
training_pipeline/train_runner.py - Training runner → Task model: runs training
training_pipeline/train_runner.py - Task model → Metric calculators: calculates metrics
training_pipeline/model.py - Metric calculators → Metric results: returns metrics
training_pipeline/metric_calculators.py - Training runner → Score aggregation: aggregates scores
training_pipeline/train_runner.py - Score aggregation → Metric results: weights results
training_pipeline/metric_aggregator.py - Training runner → Training logging: sets up logger
training_pipeline/train_runner.py - Training runner → Logging configuration: uses configuration
training_pipeline/train_runner.py - Training logging → Logging configuration: uses configuration
training_pipeline/logger_factory.py - Score aggregation → Evaluation scores: saves scores