Factored, an AI and data powerhouse, demonstrates its technical expertise through groundbreaking research published in Leveraging User Embeddings and Text to Improve CTR Predictions With Deep Recommender Systems. Authored by Factored’s engineers, this work unveils innovative methodologies that ranked 6th globally in the prestigious 2020 RecSys Challenge, sponsored by Twitter.
Redefining CTR Predictions for Social Media Engagement
Click-through rate (CTR) prediction is a cornerstone of modern marketing systems. Factored’s team, Los Trinadores, tackled the challenge of predicting user engagement on Twitter, leveraging a dataset of over 160 million tweets. The project introduced unique strategies to process text data and optimize user embeddings, ensuring both computational efficiency and predictive accuracy.
Key innovations include:
- Efficient Text Processing: Using BERT encodings as non-trainable features, reducing computational overhead without sacrificing information richness.
- User Embedding Optimization: Introducing "user buckets" to cluster similar users, enabling quality embeddings even for sparse data.
- Tailored Neural Architecture: Incorporating custom modules such as a modified Factorization Machine (FM) and attention mechanisms, enhancing the model's ability to capture high-order interactions.
Factored’s Full Spectrum Leadership
Factored’s engineers led every stage of this project, from conceptual design to deployment on Amazon Web Services (AWS), optimizing large-scale data processing pipelines. Under the leadership of Carlos Miguel Patiño, Camilo Velásquez, Juan Manuel Muñoz, Juan Manuel Gutiérrez, David Ricardo Valencia, and Cristian Bartolome Aramburu, the team exemplified technical excellence and collaborative synergy.
Recognized for Excellence
We achieved 6th place among global contenders—cementing Factored’s reputation as a leader in data-driven AI. By making the model's code available as open source, Factored also contributes to the broader AI community, fostering transparency and innovation.
To see the full paper click here.
