Prediction of uniaxial compressive strength of rock samples using density weighted least squares twin support vector regression

In this study, the uniaxial compressive strength (UCS) of rock samples has been predicted using a novel machine learning (ML) algorithm. The efficacy of the algorithm was evaluated by testing the same on a tiny dataset with only 47 samples as well as a large dataset with 170 samples. The UCS of rock samples has some outlier points in the dataset. It is well known that the samples are equally responsible for the end regressor in the case of random forest (RF), extreme learning machine (ELM), least squares support vector regression (LSSVR) primal least squares twin SVR (PLSTSVR), and even few of them act as outliers. Due to this, the prediction performance may degrade. In this study, a new density weighted approach for PLSTSVR is proposed as density weighted least squares twin support vector regression (PDWLSTSVR) in primal space, to deal with input samples in the presence of outliers. Hence, it boosts the performance of PDWLSTSVR in terms of efficiency. Here, the weights are determined with the help of k-Nearest Neighbour (k-NN) distance. Further, the proposed PDWLSTSVR is applied to real-world application like the prediction of the UCS of rock samples. To assess the competence of the proposed PDWLSTSVR, the performance of the models is tested based on different evaluation measures like RMSE, MAE, SMAPE, MASE, SSE/SST, SSR/SST and R-2. The result shows that PDWLSTSVR outperforms the RF, ELM, LSSVR and PLSTSVR in terms of all the evaluation measures.

Automated summary

The study introduces a novel density weighted approach for PLSTSVR to handle input samples in the presence of outliers, boosting the performance of PDWLSTSVR in terms of efficiency. The weights are determined with the help of k-Nearest Neighbour (k-NN) distance. Further, the proposed PDWLSTSVR outperforms the RF, ELM, LSSVR and PLSTSVR in terms of all evaluation measures when applied to the real-world application of predicting the UCS of rock samples.