Abstract: In the study of mineral resources evaluation model, the problem that the number of positive samples must be specified in the TSVM in the unlabeled samples. Then this article applied the improved algorithm to the evaluation of mineral resources, and the algorithm is TSVM based on BP neural network and BP-TSVM for short. The SVM is applied to studying the discernment and classification of labeling samples. We will annotate the plus or minus of unlabeled data within the boundary of already classified by utilizing BP neural network. The labeling samples as learned by SVM in the next iteration. The new algorithm will avoid decreased classified ability. TSVM can have more accurate estimates of number of plus of labeled data. Also, the number of training samples can reduce training times. The experimental result proves that compared with conventional SVM,TSVM and BP neural network the BP-TSVM is more accurate in evaluation of mineral resources.