For >70 years, a 4-fold or greater rise in antibody titer has been used to confirm influenza virus infections in paired sera, despite recognition that this heuristic can lack sensitivity. Here we analyze with a novel Bayesian model a large cohort of 2,353 individuals followed for up to 5 years in Hong Kong to characterize influenza antibody dynamics and develop an algorithm to improve the identification of influenza virus infections. After infection, we estimate that hemagglutination-inhibiting (HAI) titers were boosted by 16-fold on average and subsequently decrease by 14% per year. Greater boosting in HAI titer is observed in epidemics with a circulating strain that is different from the previous epidemic. In six epidemics, the infection risks for adults were 3%-19% while the infection risks for children were 1.6-4.4 times higher than that of younger adults. Every two-fold increase in pre-epidemic HAI titer was associated with 19%-58% protection against infection. Among the 1731 infections inferred by our model, around half were missed by the 4-fold rise criteria, suggesting that this criteria underestimates infection risks by 23-70%. The sensitivity and specificity of identifying infections for our approach are 87% (95% CrI: 85%, 89%) and 98% (95% CrI: 97%, 98%) respectively, which are higher than 82% (95% CrI: 80%, 84%) and 96% (95% CrI: 96%, 97%) for using 4-fold rise criteria. Our inferential framework clarifies the contributions of age and pre-epidemic HAI titers to characterize individual infection risk and offers an improved algorithm to identify influenza virus infections.