Recommender systems are an essential component of online platforms providing users with personalized experiences. Some recommendation scenarios such as social networks and news are extremely dynamic in nature with user interests changing over time and new items being continuously added due to breaking news and trending events.

Newcomers to recommender systems often face challenges related to their lack of understanding of how these systems operate in real life. In most online content related to this topic, the focus is on models and algorithms that score items based on the user’s preferences. However, the recommender model alone does not comprise everything needed for serving optimized recommender systems that meet the company’s business objectives. An industry-standard recommender system involves a number of steps, including data preprocessing, defining and training recommender models, as well as filtering and business logic for serving. In this work, we propose the four-stage recommender system, an industry-wide design pattern we have identified for production recommender systems. The four-stage pipeline includes an item retrieval step that prepares a small subset of relevant items for scoring. The filtering stage then cleans up the subset of items based on business logic such as removing out-of-stock or previously seen items. As for the ranking component, it uses a recommender model to score each item in the presented list based on the preferences of the user. In the final step, the scores are re-ordered to provide a final recommendation list aligned with other business needs or constraints such as diversity. In particular, the presented demo demonstrates how easy it is to build and deploy a four-stage recommender system pipeline using the NVIDIA Merlin open-source framework.

Much of the recent progress in sequential and session-based recommendation has been driven by improvements in model architecture and pretraining techniques originating in the field of Natural Language Processing. Transformer architectures in particular have facilitated building higher-capacity models and provided data augmentation and training techniques which demonstrably improve the effectiveness of sequential recommendation. But with a thousandfold more research going on in NLP, the application of transformers for recommendation understandably lags behind. To remedy this we introduce Transformers4Rec, an open-source library built upon HuggingFace’s Transformers library with a similar goal of opening up the advances of NLP based Transformers to the recommender system community and making these advancements immediately accessible for the tasks of sequential and session-based recommendation. Like its core dependency, Transformers4Rec is designed to be extensible by researchers, simple for practitioners, and fast and robust in industrial deployments.

The lines between data science (DS), machine learning (ML), deep learning (DL), and data mining continue to be blurred and removed. This is great as it ushers in vast amounts of capabilities, but it brings increased complexity and a vast number of tools/techniques. It's not uncommon for DL engineers to use one set of tools for data extraction/cleaning and then pivot to another library for training their models. After training and inference, it's common to then move data yet again by another set of tools for post-processing. The RAPIDS suite of open source libraries not only provides a method to execute and accelerate these tasks using GPUs with familiar APIs, but it also provides interoperability with the broader open source community and DL tools while removing unnecessary serializations that slow down workflows. GPUs provide massive parallelization that DL has leveraged for some time, and RAPIDS provides the missing pieces that extend this computing power to more traditional yet important DS and ML tasks (e.g., ETL, modeling). Complete pipelines can be built that encompass everything, including ETL, feature engineering, ML/DL modeling, inference, and visualization, all while removing typical serialization costs and affording seamless interoperability between libraries. All experiments using RAPIDS can effortlessly be scheduled, logged and reviewed using existing public cloud options. Join our engineers and data scientists as they walk through a collection of DS and ML/DL engineering problems that show how RAPIDS running on Azure ML can be used for end-to-end, entirely GPU pipelines. This tutorial includes specifics on how to use RAPIDS for feature engineering, interoperability with common ML/DL packages, and creating GPU native visualizations using cuxfilter. The use cases presented here give attendees a hands-on approach to using RAPIDS components as part of a larger workflow, seamlessly integrating with other libraries (e.g., TensorFlow) and visualization packages.

The selection of features and proper preparation of data for deep learning or machine learning models plays a significant role in the performance of recommender systems. To address this we propose a tutorial highlighting best practices and optimization techniques for feature engineering and preprocessing of recommender system datasets. The tutorial will explore feature engineering using pandas and Dask, and will also cover acceleration on the GPU using open source libraries like RAPIDS and NVTabular. Proposed length is 180min. We’ve designed the tutorial as a combination of a lecture covering the mathematical and theoretical background and an interactive session based on jupyter notebooks. Participants will practice the discussed features by writing their own implementation in Python. NVIDIA will host the tutorial on their infrastructure, providing dataset, jupyter notebooks and GPUs. Participants will be able to easily attend the tutorial via their web browsers, avoiding any complicated setup. Beginner to intermediate users are the target audience, which should have prior knowledge in python programming using libraries, such as pandas and NumPy. In addition, they should have a basic understanding of recommender systems, decision trees and feed forward neural networks.

This industry talk covers the deep learning architecture developed at Realtor.com to recommend real estate listings to our userbase. The recommendation of homes is a different problem than most other domains both in the sense that listings are unique and that there are additional geographic and time constraints that increase the sparsity of interactions and make recommendation of individual listings more challenging. In particular time on market in a hot area can be limited to weeks or even days, and listing cold-start is critical to providing up to date market information. Thankfully the structured feature data for listings is incredibly rich and provides a framework from which to map listings into a meaningful vector space. User first impressions are also incredibly important in this highly competitive field, and offline recommendation or models that don't adapt during the users session are less desirable.