SPARK
A recommendation engine that learns from what shoppers actually do — and keeps adapting as their tastes shift, instead of relying on fixed rules that go stale.
Overview
SPARK — the Smart Product Analysis and Recommendation Kernel — is a deep-reinforcement-learning agent that personalises product recommendations in response to real customer interactions.
Rather than scoring items with static rules, it treats recommendation as a decision it learns to make better over time: it observes behaviour, acts, sees the response, and updates. Because tastes change, the agent changes with them.
How it works
- 1
Behaviour as signal
An e-commerce interface gathers real interactions — clicks, likes, reviews and purchases — as the raw signal of intent.
- 2
Feature engineering into latent taste
Past and real-time interactions are engineered into latent features that capture user preference, feeding the reward function that guides the agent.
- 3
A reinforcement-learning agent
The agent recommends, observes how the user responds, and updates — learning a policy that adapts to shifting behaviour over time.
- 4
A closed feedback loop
Recommendations return to the interface as products the user can interact with; that interaction becomes the next round of training data — a continuous online loop.
Interface
