Insect Voice Recognition And Quantitatively Structure-Activity Relationship Of Mosquito Repellent Based On SVM

The traditional taxonomic identification of insect morphology is time or labor consuming. However, automatic identification method based on insect sound is more straightforward, convenient and prompt. Insect repellent can prevent mosquito harassing people's daily life, and effectively control the dissemination of a variety of infectious diseases. So, it is of great significance for developing new effective insect repellent. Support Vector Machine (SVM) epitomized the field of machine learning, including Support Vector Classification (SVC) and Support Vector Regression (SVR). It has many advantages such as suitable for modeling on small sample and nonlinear data, effectively avoid over fitting and curse of dimensionality, meanwhile the generalization ability is excellent. This paper applied SVM to insect voice recognition and Quantitatively Structure-Activity Relationship (QSAR) of mosquito repellent. The results are showed as follows:Automatic recognition of insect voice based on MFCC_GS_SVC. Mel Frequency Cepstrum Coefficients (MFCC) does not consider the context correlation of sound signals. This article introduced the conception of Geo-statistics (GS) with one dimensional time series, and then characterized the context correlation of sound signals using the semivariograms between time-domain signals in different time interval. The final features contain 8832 MFCC and 150 GS features, which were screened a large number of irrelevant and redundant features by Binary Matrix Shuffling Filter (BMSF). We obtained 1958 reserved features, containing 1900 MFCC and 58 GS features. For a dataset of insect sound which consists of 100 samples locating in 10 classes, first transform the multi-class to binary-class problems and then, combine SVC and the simple voting correction strategy to perform prediction. Finally, we achieved a 100% independent prediction accuracy on 30 test samples selected randomly..Nonlinear QSAR modeling of mosquito repellent based on SVR. The repellent activities are determined by the effective protection time of 40 kinds of amides compounds to Aedes aegypti. We obtained 1773 initial molecular descriptors for each compound by using the stoichiometry software PCLIENT. Then the Binary Matrix Shuffling Filter and Worst Descriptor Elimination Multi-round methods were successively used to conduct the nonlinear selection on the initial descriptors. We finally obtained 8 well-defined molecular descriptors in physical and chemical characteristics. The SVR was used to construct the nonlinear QSAR model, which is highly significant with the F= 8465 and R2= 0.9996. The analyses of single factor significance and effect showed that:the relative importance of 8 reserved descriptors is TPSA (Tot)> RDF035e> BLTF96> Eigle> RDF055u> IC4> RDF075e> G (N..O), in which the descriptors IC4, RDF055u and RDF075e show positive correlation with the compound activities, but the G(N..O) and TPSA(Tot) reveal negative correlation with the repellent activities. In addition, the RDF035e, Eigle and BLTF96 present parabolic relation with the bioactivities.