Content Recommender Engines

Objective

To develop and deploy an advanced content recommender engine that enhances the personalization of daily content delivery for users by leveraging their historical activities, bio, and demographic features.

Methodology

Employed TensorFlow Recommenders (TFRS) for the construction of the recommendation model. The system was fine-tuned using Kubeflow and orchestrated via an Apache Airflow pipeline on Google Cloud Platform (GCP), ensuring a scalable and robust deployment.

Technical Stack

TensorFlow Recommenders (TFRS), Kubeflow, Apache Airflow, Google Cloud Platform (GCP).

Execution

Analyzed and understood the existing system's codebase, identifying key areas for enhancement. Successfully expanded the model's capacity to process bios in 105 languages, up from its previous single-language (English) capability. Additionally, I optimized the post-processing step to exclude content that was phased out by the company, ensuring that recommendations remain relevant and up-to-date.

Challenges

The primary challenge was to extend the model's language processing capabilities to include a diverse set of languages while maintaining the accuracy and relevance of content recommendations. Another challenge was refining the post-processing algorithm to dynamically adjust to the changing content offerings of the company.

Impact

The enhancements led to a more inclusive and refined user experience, potentially increasing user engagement and satisfaction with the personalized content provided.

Future Direction

Plans are in place to conduct A/B testing to assess the impact of different versions of the recommender system on user engagement and further refine the recommendation algorithm based on the insights gained.