Looking to elevate your user experience to unprecedented heights?
Our customer recommendation engine could be the indispensable asset you need to incorporate into your business workflow.
Explore the sophisticated TV catch-up recommendations, tailored to your audience using an advanced software algorithm developed by the Smartivus team. Intrigued to learn more? Read on to understand the mechanics behind our proprietary model.
The Core: Crafting a Customer Recommendation Engine
This article unveils the intricacies involved in developing a dynamic customer recommendation engine specifically designed for catch-up events on linear TV channels. Below is a glimpse into the preparation phase of our machine learning (ML) models.
Software Environment:
Tools: Python (Pytorch, FastAI, Sklearn) + GPU
Automation: Airflow
Data Source: Redis
AI Model Storage: Redis + Redis AI Module
The Algorithm: Optimizing the ML Model
Our customer recommendation engine empowers your platform by elevating viewer satisfaction through relevant and current content, fine-tuned via a cutting-edge algorithm:
- Data Storage: All statistical usage data is retained in Redis for a 14-day period, coinciding with the depth of the catch-up archive.
- Initial Cleanup: Removal of erroneous or corrupted entries, if any.
- Event Aggregation: Aggregating and grouping of the events.
- Secondary Cleanup: Removal of non-essential events such as “end of the program.”
- Weighting: Addition of weights to each event, based on comprehensive research indicating that events from high-traffic TV channels generally outperform those from thematic packages.
- Normalization: Data normalization for consistency.
- Segmentation: Customer profiles are segmented into eight distinct groups based on similar viewing patterns.
- Recommendation Generation: Calculating tailored recommendations for each segment, resulting in eight distinct models.
All the models are subsequently integrated back into the Redis system, facilitating real-time recommendation delivery via the Redis AI module.
Data Specifics:
Approximately 17 million statistical entries
Over 20,000 TV events
120 TV channels
Around 100,000 customer profiles
Real-World Implementation
The algorithm is set to run automatically through the Airflow scheduler, twice daily. On a dedicated server equipped with a GPU card, full model regeneration is achieved in roughly 25 minutes.
Server Specifications:
2×8-core Intel Xeon E5–2620 v4
16GB DDR4 ECC
Nvidia TU104 GPU
Want to Dive Deeper?
Eager to learn more? Let’s connect with us on LinkedIn.
