Notable Papers:

A Systematic Review and Replicability Study of BERT4Rec for Sequential Recommendation

Note: fancy models do not necessary provide best results across the board.

Paper:

Talk:

Denoising Self-Attentive Sequential Recommendation

Note: differentiable masking for regularization. Not sure if this is borrowed from NLP space.

Paper:

Talk:

Modeling Two-Way Selection Preference for Person-Job Fit

Note: useful for two-way marketplaces.

Paper:

Talk:

Augmenting Netflix Search with In-Session Adapted Recommendations

Note: Useful for QUIQSearch in the future.

Paper:

Talk:

Exploring the longitudinal effects of nudging on users’ music genre exploration behavior and listening preferences

Note: longer term user modeling instead of session based user modeling.

Paper:

Talk:

RADio – Rank-Aware Divergence Metrics to Measure Normative Diversity in News Recommendations

Note: Useful for QUIQSearch in the future.

Paper:

Talk:

Industry Chatters:

Industry vs Academic Skew

• Industry prefer fast models for real time inference. Academics favor larger/complex models.
• Industry paper architectures are highly bespoke to their software application. Academics are limited to MovieLens, Amazon Reviews, or very low sample survey data
• Transformers (BERT4Rec) still dominates session based topics, but it’s results does not apply across different datasets.
• No one knows how to evaluate embeddings. Dimensionality reduction does not work.

RecSys/Search/Data Science Team Organizations

• Applied Scientists/Data Scientists to Software Engineers ratio ranges from 1:1 at the lowest to around 8:1 for most companies.
• RecSys/Search & Discovery/Data Science teams are usually built on top and supported by data engineering and platform engineering teams.
• Even split between data scientists embedded in product teams lead by PM vs RecSys/Search & Discovery/Data Science teams lead by tech lead.

ML Engineering/Ops/Infrastructure

• Big skew between infrastructure maturity between smaller and bigger companies. Few builds in-house tools. Most buys and stitch together in-house platform.
• Low adoption of NVIdia’s in-house frameworks.
• It is still common to run batch jobs over Spark/Ray on a hourly/daily/weekly basis for candidate generation and persist results in Redis.
• Lot’s of interest for streaming/real time recommendation, but the challenge is support in applications and data engineering. AdTech companies don’t face such issues.

Tutorials:

Tutorial: Hands on Explainable Recommender Systems with Knowledge Graphs

Video:

Code:

Tutorial: Training and Deploying Multi-Stage Recommendation System

Video: ](https://vimeo.com/749421236)

Code:

Event: DLI Event – NVIDIA. Event code: DLI_XLAB_SR22