Online Absolute Ranking with Partial Information: A Bipartite Graph Matching Approach
share
Ever since the introduction of the secretary problem, the notion of selecting from a pool of candidates before encountering every candidate has been of significant research interest. Motivated by the real-world problem of hiring in a supply-constrained environment, e.g., the tech industry, we extend the secretary problem setting to the case where the complete score of a candidate is revealed only after a certain number of subsequent candidates (characterizing the delay in the partial score) have been encountered. A novel aspect of our study is that of predicting each candidate's absolute rank before all of the candidates have been revealed. A key contribution of our study involves the innovative use of weighted bipartite matching to find the maximum likelihood assignment of candidates to order-statistic distributions. We run extensive experiments on synthetically generated data as well as real-world data obtained from a university course. On synthetic data, we observe an average deviation of at most 1.6 ranks out of 25 from the true ranking, and on real-world partially observed data we observe a median deviation of 5.5 ranks out of 50. On fully observed real-world data, the algorithm successfully identifies the ten best students immediately upon their arrivals into the system, demonstrating that the algorithm has wide applicability for online hiring problems. Furthermore, we generalize the algorithm to match candidates to an arbitrary set of template distributions. This is particularly useful when we are interested in a certain fraction of the candidates only (e.g. the top 10 the order statistics. For this case, we introduce a novel sparsified randomized matching algorithm with a better complexity ( O(N^2 N), where N is the number of candidates) than the naive algorithm ( O(N^3)).
READ FULL TEXT