Summary of Efficient User History Modeling with Amortized Inference For Deep Learning Recommendation Models, by Lars Hertel et al.
Efficient user history modeling with amortized inference for deep learning recommendation models
by Lars Hertel, Neil Daftary, Fedor Borisyuk, Aman Gupta, Rahul Mazumder
First submitted to arxiv on: 9 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Information Retrieval (cs.IR)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper explores user history modeling in deep learning recommendation models (DLRM) using Transformer encoders, which can improve recommendation quality but require significant infrastructure upgrades or small model sizes. The authors revisit early fusion methods and compare concatenating candidate items with appending them to the end of the list as separate items. They reformulate M-FALCON, an amortized history inference algorithm, for DLRM models. Experimental results show that appending with cross-attention performs similarly to concatenation, while amortization reduces inference costs. The authors conclude by deploying this model on LinkedIn’s Feed and Ads surfaces, achieving a 30% latency reduction via amortization. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how we can improve online recommendations using special computer models called Transformers. These models are good at giving us suggestions based on what someone has looked at before, but they need lots of power to work quickly. The authors want to make these models better and faster by combining different pieces of information together in a new way. They test this idea and show that it works just as well as some other methods, while also being more efficient. Finally, they use this new method on LinkedIn’s website, where it makes the page load 30% faster. |
Keywords
» Artificial intelligence » Cross attention » Deep learning » Inference » Transformer
