POLIBITS, Vol. 57, pp. 19-37, 2018.
Abstract: least square regression, singular value decomposition, propensity, anomaly, accuracy, precision, learning management systems
Keywords: The classical Ordinary linear Least Square approximation (OLS) model has been used as the best fit regression for linear trend data. In data analysis, the accuracy of analysis depends on the model as well as the metric used to measure the error. Singular Value Decomposition (SVD) is also applied for Normal linear Least Square (NLS) approximation along the perpendicular to the approximating line. The OLS line is not sensitive to temporal variation in time variables whereas SVD is sensitive, it renders OLS less suitable for time sensitive data. Both OLS and SVD use quantitative metric for regression analysis, and SVD has inherent constraints. Propensity score analysis is an innovative class new technique for qualitative error analysis. Propensity score method is easier to communicate to non-expert audience. Moreover, propensity score estimates are often more robust than the percent error estimates of predicted values over the true values. Herein we present a hybrid algorithm that achieves a balance between quantitative and qualitative approximation accuracy of both OLS and NLS (SVD). This metric has also proved useful for evaluating the effects of treatments in real patient data. This technique is also suitable for anomaly removal. Visualization is a preferred way to ascertain the quality of a new algorithm and is used to demonstrate the hybrid algorithm. We have applied this criteria for comparison with other existing methods. We have found that this technique is reliable and preferable to explain to the expert as well as non- expert. The empirical tests show the accuracy improvements over conventional methods.
PDF: Exploration, Exploitation Phenomena and Regression Analysis: Propensity Metric, Anomaly Reduction, Dimensionality Reduction
PDF: Exploration, Exploitation Phenomena and Regression Analysis: Propensity Metric, Anomaly Reduction, Dimensionality Reduction
https://doi.org/10.17562/PB-57-2
Table of contents of POLIBITS 57