Multi-disciplinary Spatio-temporal prediction is the task of predicting sequence values over different time series. Localized Support Vector Machine (LSVM) with an efficient algorithm called Profile SVM efficiently classified data by extracting smaller clusters which achieved comparable accuracy on both temporal and spatial data. However, the learning task was concerned with only the classification problem whereas regression and time series prediction was not concentrated. Digital elevation model (DEM) with SVM classification predicted the distribution of rare species in island forest ecosystems resulting in higher classification accuracy. But the application of DEM was confined to specific species was not addressed on predicting diverse elevation (i. e.,) at different user conditions. To improve the linear data regression performance level, Classification-based Quadratic Mean Loss function (CQML) is proposed in this paper. CQML function aims to produce the linear regression model by predicting the Web Spatio-temporal data by reducing the average error rate. CQML function segments each space-time series data to easily predict the regression. A method, Distribution Time Transform is carried out for performing segmentation for different space-time series data. The segmented data are grouped concurrently in CQML function by significantly optimizing the average loss on each group. The grouping operation is performed using the quadratic mean computation to minimize the loss and variance of loss while developing the linear regression model. CQML function is experimented on factors such as linear regression rate, average error rate and segmentation time. © Research India Publications.