Data scientists and machine learning engineers working in ecommerce and media industries use session-based recommendation algorithms to predict a user's next action within a short time period, particularly for anonymous users (i.e, to tackle the user cold-start problem) or when users' interests are very contextual and change within a session. Providing relevant recommendations to first-time or early-visit users helps engagement, retention, and signing up for subscription services. Transformers4Rec is an open-source library that streamlines building pipelines for session-based recommendations and is included with NVIDIA Merlin™, an open source framework that accelerates the entire pipeline, from ingesting and training to deploying a GPU-accelerated recommender system.
The NVIDIA Merlin team designed Transformers4Rec to help machine learning engineers and data scientists explore and apply Transformer architectures when building sequential and session-based recommendation pipelines.
Inspired by NLP Transformers, Designed for Recommenders
Transformers4Rec is designed for recommender workflows. It takes inspiration from Transformer architectures and training methods utilized within NLP (Natural Language Processing) to support language modeling and sequence-to-sequence tasks. Transformers are an efficient replacement for RNNs (Recurrent Neural Networks). Evaluations have indicated that RNN-based session-based recommender solutions can often be outperformed by simpler algorithms, including k-NN (k-Nearest Neighbor), for session-based recommendations. The NVIDIA Merlin team developed Transformers4Rec by embedding learnings from research projects, participating in industry competitions, and leveraging Hugging Face Transformers, a popular NLP library.LEARN MORE
Solving for Users Cold-Start Problem
Recommender methods popularized in mainstream media often rely upon long-term user profiles or lifetime user behavior. Yet, ecommerce and media companies acquiring new ongoing active users must provide relevant recommendations to first-time and early-visit users. Relevant recommendations enable increased user engagement, retention, and conversion to subscription services. Utilizing session-based recommenders with Transformers4Rec, data scientists and machine learning engineers are able to solve the cold-start problem by leveraging contextual and recent user interactions to predict a user's next action and provide relevant recommendations. Transformers4Rec can be used as a standalone solution or within an ensemble of recommendation models.LEARN MORE
Flexibility Supports Experimentation
Transformers4Rec supports multiple input features and provides many options for representing and combining them. It also provides flexible building blocks for creating architectures with multiple towers, heads, and loss functions for item recommendation or sequence classification (e.g., for predicting cart abandonment). It includes popular ranking metrics for offline evaluation, and the whole pipeline can be easily deployed in NVIDIA Triton™ Inference Server for GPU-accelerated inference.LEARN MORE
Merlin Team SIGIR eCom Data Challenge
Predict User Intent, a Winning Solution
Data scientists and machine learning engineers are able to leverage Transformers4Rec as a standalone library or within an ensemble model solution. The NVIDIA Merlin team leveraged Transformers to win the SIGIR eCom Data Challenge 2021, an industry competition with tasks predicting the next interacted item and the cart abandonment probability within a shopping session. The Merlin team used data augmentation and feature engineering techniques, as well as an ensemble of two Transformer models based on XLNet and Transformer-XL for the winning solution.Learn more about the SIGIR eCom Data Challenge
Review the code used
Streamlined Workflows with Merlin
Transformers4Rec is included in NVIDIA Merlin, an open source framework that accelerates the entire pipeline, from ingesting and training to deploying a GPU-accelerated recommender system. All Merlin components are available as open-source projects on GitHub or as containers in the NVIDIA NGC catalog. Containers package the software application, libraries, dependencies, and runtime compilers in a self-contained environment. This way, the application environment is both portable, consistent, reproducible, and agnostic to the underlying host system software configuration.
Merlin on NGC
Enables users to do preprocessing and feature engineering with NVTabular and then train a deep learning-based recommender system model with HugeCTR.PULL CONTAINER FROM NGC
Merlin TensorFlow Training
Utilize preprocessing and feature engineering with NVTabular and then train a deep learning-based recommender system model with TensorFlow.PULL CONTAINER FROM NGC
Merlin PyTorch Training
Leverage preprocessing and feature engineering with NVTabular and then train a deep learning-based recommender system model with PyTorch.PULL CONTAINER FROM NGC
Container allows users to deploy NVTabular workflows and HugeCTR or TensorFlow models to the NVIDIA Triton™ Inference Server for production.PULL CONTAINER FROM NGC
Merlin on GitHub
The GitHub repository provides documentation, tutorials, examples, and notebooks to help users get started with NVIDIA Merlin.
ACM RecSys Transformers4Rec Paper
Read more about NVIDIA Merlin team's ACM RecSys paper on bridging the gap between NLP and session-based Recommendations.
Explore the components of NVIDIA Merlin, which include Merlin Feature Engineering: NVTabular, Merlin Training: HugeCTR, Merlin Inference: NVIDIA® TensorRT™ and Triton, and Merlin Reference Applications.
Merlin Technical Resource Kit
Learn how to accelerate the entire pipeline, from ingesting and training to deploying GPU-accelerated recommender systems.
NVIDIA Deep Learning Institute (DLI)
Register for the Building Intelligent Recommender Systems DLI workshop to learn the fundamental tools and techniques for building highly effective recommender systems.
Merlin Transformers4Rec is available to download.