On the value of diversified recommendations

What this post added

This post introduces the concept of diversified recommendations as a crucial evolution beyond simple predictive accuracy in recommendation systems. It uses toy models (Polya and balancing processes) to illustrate the trade-offs between positive reinforcement (leading to boredom) and negative reinforcement (leading to lack of engagement). The post then outlines practical methods for achieving recommendation diversity, including author, media type, and semantic diversity, exploring similar semantic nodes, maintaining long and short-term preferences, utilizing explore-exploit algorithms (epsilon-greedy, UCB), qualitative trade-off computation, and negative guardrails. It emphasizes optimizing for long-term user satisfaction over purely metric-driven approaches.