Welcome to Compresso Recsys’s documentation!

Compresso Recsys is the recommender-system companion package for Compresso. It contains dataset loaders, checkpoint-building utilities, checkpoint read/write helpers, and retrieval metrics used to demonstrate sparse representation learning.

The project distribution is named compresso-recsys and the Python package is imported as compresso_recsys.

Getting Started

• Installation
  • Install from PyPI
  • Optional Extras
  • Local Development
  • Install from GitHub
  • Build the Documentation Locally
• Getting Started
  • Basic Python Usage
  • Checkpoint Workflow
  • What Goes Into a Checkpoint
  • Split Modes
  • Retrieval Metrics
• Quickstart
  • Choose a Split Mode

User Guide

• Examples
  • Dataset Loader
  • Building Checkpoints
  • Checkpoint Read/Write
  • Evaluate Embeddings From Python
  • Evaluate Model Predictions
  • Train TEASER for Cold Items
  • Train TEASER with Gradient Descent
  • Serve New TEASER Candidates
  • Train and Evaluate ELSA
  • Using item embeddings
• Datasets
  • Reading the default tables
  • MovieLens 1M
  • MovieLens 20M
  • Goodbooks-10k
  • Amazon Reviews 2023
  • Steam
  • Netflix Prize
  • MSD Taste Profile
  • Gowalla
  • DBbook
  • Last.fm-2K
  • Using item embeddings
• Bringing Your Own Dataset
  • What a dataset has to produce
  • The dataset
  • The dataset class
  • Two defaults that will empty your dataset
  • What is in it
  • Does it work?
  • Cold items, for free
  • What timestamps would have added
  • You do not need a dataset class at all
• Implementing a Recommender
  • What the base class owns
  • Data
  • Configuration
  • The complete model
  • Low-level prediction
  • Stable-ID recommendation
  • Standalone persistence
  • Persistence inside a data checkpoint
  • Evaluation
  • Executable contract checks
  • Extending the pattern
• Reproducing SASRec Results on ML1M
  • Configuration and run modes
  • Build the MovieLens 1M checkpoint
  • Inspect the split
  • Train SASRec
  • Score the catalog
  • Published protocol: 100 sampled negatives at 10
  • Package protocol: full catalog at 20
  • Check the epoch budget on validation data
  • Save the trained model
• Comparing Models Statistically
  • Why the aggregates are not enough
  • Do the comparison
  • A complete worked example
  • Reading the output
  • Five traps
  • Choosing the settings
  • When one user owns several rows
  • What the methods do, formally
  • Reporting the result
  • References
  • See also
• Citing Compresso Recsys
  • Compresso
  • EASE
  • Neighborhood Models
  • Mult-VAE, Mult-DAE, and AutoRec
  • TEASER
  • ELSA
  • ELSA with sampled output candidates
  • Compressed ELSA
  • SASRec
  • SimpleRNN
  • SimpleGPT and SimpleBidirectionalTransformer
  • Content, Popularity, and Random Baselines
  • Datasets
  • Evaluation Protocol
  • Statistical Comparison
  • SWAP multimodal ML-1M, DBbook, and Last.fm-2K
• Checkpoint CLI Reference
  • Temporary MovieLens download workaround
  • Steam
  • Amazon Reviews 2023
  • Checkpoint Split Schema
  • Builder Parameters
  • Supported Amazon Reviews 2023 Datasets

API Reference

• API Reference
  • Core API
  • Datasets API
  • Checkpoint API
  • Fitted Model Persistence
  • Sequences API
  • Models API
  • Retrieval API
  • Ranking Evaluation API
  • Statistics API

Indices and Tables

• Index

• Module Index

• Search Page