implicit

Project Overview

implicit is a collaborative filtering recommendation library focused on implicit feedback data such as clicks, purchases, and plays. It implements multi-threaded training with Cython and OpenMP, fitting models efficiently across all CPU cores. It provides four mainstream algorithms: implicit ALS, Bayesian Personalized Ranking (BPR), logistic matrix factorization, and item-item nearest neighbour models using Cosine, TFIDF, or BM25 distances. ALS and BPR ship custom CUDA kernels, and Annoy, NMSLIB, or Faiss can accelerate serving recommendations.

Train industrial recommendation models straight from behaviour logs, no ratings required.

Key Features

• Implicit ALS: alternating least squares tuned for implicit feedback with multi-threaded training
• BPR: Bayesian Personalized Ranking for ranking scenarios
• Logistic matrix factorization: a probabilistic factorization of interactions
• Item-item KNN: models using Cosine, TFIDF, or BM25 distances
• GPU training: custom CUDA kernels for ALS and BPR
• ANN acceleration: integrates Annoy, NMSLIB, and Faiss to speed up recommendations
• Similar lookups: one call returns similar items or similar users

Target Audience

• Backend engineers building recommender systems
• Algorithm engineers handling implicit behaviour data
• Researchers studying recommendation algorithms
• Competitors working on recommendation challenges

Quick Start

1. Install:

pip install implicit

2. Train an ALS model and recommend:

import implicit

model = implicit.als.AlternatingLeastSquares(factors=50)
model.fit(user_item_data)
recommendations = model.recommend(userid, user_item_data[userid])

3. Find related items:

related = model.similar_items(itemid)

Core Advantages

• Fast training: Cython plus OpenMP saturates all cores, far beyond pure NumPy
• Optional GPU: mainstream algorithms ship CUDA kernels, fitting large datasets in seconds
• Implicit-feedback focus: interfaces and algorithms designed around clicks and purchases

System Requirements

Notes

• Input format: models take sparse matrices; build the user-item confidence matrix first
• BLAS threads: on OpenBLAS set OPENBLAS_NUM_THREADS=1; disabling its internal threading actually speeds this library up
• GPU setup: GPU support requires CUDA 13 and the rmm dependency; the CPU build has no such requirement