Construyendo un recomendador de emparejamiento de expertos

Creating an expert recommendation engine presents unique challenges. Unlike content recommendation, the item being recommended is a person with finite capacity. Recommending an expert to many people simultaneously can be costly on both ends - the requestor wasting a request and the expert wasting an hour of their time. Both parties learn to ignore the surface quickly. The statement must be checkable for credibility.

The challenge lies in the justification process. Three independent evaluators assess the eligible pool of experts, combined via Reciprocal Rank Fusion. RRF is the correct primitive here, as it normalizes the varying outputs of individual evaluators. It disregards the magnitude of scores and focuses solely on their ranking order. This order is the only consistent information across samples of limited size.

The standard constant 'k' used in the RRF formulation is 60. Scoring combines multiple components with specific weights. Expert quality saturates, meaning an expert with no rating history scores 0.5, not 0. The exponential saturation on experience prevents overvaluing veteran experts. Fairness incorporates a logarithmic penalty for overexposure. The overall approach aims to create a balanced global allocation of requests to experts, preventing a few top experts from monopolizing all the resources.