Project Synopsis
A large stock-audio marketplace needed a recommendation system for stock audio clips. Users of the system could be anonymous or old users. The system was expected to learn individual taste profiles and provide recommendations based on their past purchases and interactions with the system. It was also required to provide recommendations for new users who had no interactions with the system before (cold start). The logic behind recommending a certain item to the user was also required to be flexible enough to take into account new changes in business requirements, and also new information gathered about the products.
The developed platform was expected to have the following features:
- Recommendations to users listening to an audio clip
- Recommendations to anonymous/new users to the system
- Capable of ingesting and processing large amounts of audio and user data
- Flexibility to tune the business logic any time
Rare Mile Solution
We designed and developed the complete solution using leading technologies and proven algorithms in the industry. The delivered solution had the following features:
- Support for analyzing large scale data
- Tunable recommendation system providing precise control for business logic
- User taste profiles built using past interactions with the system
- Real-time feedback mechanism allowing the system to keep the recommendations fresh
- Fallback logic allowing the system to recommend audio to users without taste profiles
Project Highlights
- Analyzed audio and extracted features using Deep Learning
- Backed by Spark to allow distributed computations
- Developed APIs which allow the system to be deployed on large clusters and used as an external service
- Created algorithms to tune recommendations and provide flexibility for the business logic
