Skip to content

BaseModelAI/EMDE

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 

Repository files navigation

Achievements

1️⃣st place at SIGIR eCom Challenge 2020

2️⃣nd place and Best Paper Award at WSDM Booking.com Challenge 2021

2️⃣nd place at Twitter Recsys Challenge 2021

3️⃣rd place at KDD Cup 2021

Applications

We are currently using EMDE for generating candidates to facilitate downstream recommendation systems. It generates recommendations using density-based rich customer representation. It allows us to trace customer look-alikes (‘People Like You’) to find similar users with similar cuisine/taste preferences as well as price affinity. We used Cleora for customer-restaurants graph data […] And to our delight, the embedding generation was superfast (i.e <5 minutes). For context, do remember that GraphSAGE took ~20hours for the same data in the NCR region. Cleora + EMDE gives us a generalised framework for recommendations […] We are exploring ways to use it in other applications such as search ranking, dish recommendations, etc.
~ Zomato.com Data Science team

Dailymotion has applied EMDE to personalize video recommendations in native applications, leading to improved relevance and catalog coverage.

Synerise is using EMDE working with clients from many industries such as:
retail, ecommerce, banking, telco, travel, health, insurance, automotive, fashion

for tasks including:
recommendations, propensity prediction, churn prediction, anaomaly detection, customer scoring, customer matching, behavioral super-segmentation

EMDE

Efficient Manifold Density Estimator is a framework utilizing arbitrary vector representations with the property of local similarity to succinctly represent smooth probability densities on Riemannian manifolds.

EMDE can be used to:

  • create a fixed-size aggregated representation
  • serve as an input and output to the neural network
  • work with multi-modal data
  • work with various embedding types

EMDE competitive edge:

  • significantly increases model performance
  • shifts heavy lifting part from neural network to data representation
  • results in radically smaller models and shorter training time

In most of our use cases, we use a sparse feed-forward network with around 5 layers and trained for 2 to 5 epochs on a single GPU.

Blog posts

Papers

Cite

Please cite our paper:

@inproceedings{dkabrowski2021efficient,
  title={An efficient manifold density estimator for all recommendation systems},
  author={D{\k{a}}browski, Jacek and Rychalska, Barbara and Daniluk, Micha{\l} and Basaj, Dominika and Go{\l}uchowski, Konrad and B{\k{a}}bel, Piotr and Micha{\l}owski, Andrzej and Jakubowski, Adam},
  booktitle={Neural Information Processing: 28th International Conference, ICONIP 2021, Sanur, Bali, Indonesia, December 8--12, 2021, Proceedings, Part IV 28},
  pages={323--337},
  year={2021},
  organization={Springer}
}