Abstract: With the increase in scientific literature, classifying scientific literature has become an important focus. Effectively selecting representative features from scientific literature has become a key step in scientific literature classification and influences the performance of scientific literature classification. According to the structural characteristics of scientific literature, we combine an optimized K-medoids algorithm, which firstly adopts information entropy to empower clustering objects to correct the distance function and then employs the corrected distance function to select the optimal initial clustering centres, with the Apriori algorithm to propose a four-layer feature selection method. The proposed feature selection method firstly divides scientific literature into four layers according to their structural characteristics, selects features layer by layer from the former three layers by means of the optimized K-medoids algorithm, subsequently mines the maximum frequent item sets from the fourth layer by the Apriori algorithm to act as the features of the fourth layer, and finally merges selected features of every layer and eliminates duplicate features to obtain the final feature set. Experimental results show that the proposed four-layer feature selection method achieves higher performance in scientific literature classification.

Key words: feature selection, k-medoids algorithm, information entropy, apriori algorithm