Accurate prediction of carbon content and temperature is the crucial to the endpoint control of converter steelmaking. For the large sample fluctuation, it is difficult to measure the similarity of samples in just-in time learning(JITL), which leads to the problem of low prediction accuracy, therefore, this paper proposes a similarity measurement strategy based on an improved spectral clustering algorithm. Firstly, according to the coupling relationship between process variables and dominant variables, a spectral clustering algorithm with global weighted KL measurement standards is constructed, thus, the clustering subsets with large between clusters variance and small intra-cluster variance are obtained to eliminate the fluctuation among the furnace samples. Secondly, according to the difference information between class clusters, the local weighted KL metric criterion is integrated to calculate the posterior probability of the predicted samples belonging to various clusters, then, a similarity measurement strategy suitable for describing the complex characteristics of converter steelmaking process is constructed. Finally, this measurement strategy is used to calculate a subset of samples that are more similar to the properties of the new furnace, and the RVM model is established to predict the end point carbon content and temperature. The simulation results of actual converter steelmaking process show that the prediction accuracy of carbon content within pm0.02% error range reaches 89%, temperature within pm$10\textdegree$C error range reaches 92%.