ADAPTIVE RIDGE REGRESSION-FS WITH WEIGHTED NORMALIZATION OF INDONESIA'S GDP GROWTH PREDICTION USING GOOGLE TRENDS

• Main Author: Suwardiman
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55677
• Institution: Institut Teknologi Bandung

Adaptive Ridge Regression Selection Operator (ARSO) is an elaboration of the Ridge Regression by adjusting the variable selection. In ARSO the weighting is calculated using the Ordinary Least Squares (OLS) and the weighting formulation is still using conventional methods. Therefore, in this research, changes were made on the weighting technique with expectations the accuracy and ability of ARSO in selecting variables are increased. In this research, the weighting technique was done by conducting to use several feature selection methods such as the lasso, RReliefF and Correlation Feature Selection (CFS) and the weighting formulation was was done by normalizing the value of the selection feature’s results. Experiments were performed by indicating the case studies on the prediction of Indonesia's Gross Domestic Product (GDP) by utilizing Google Trends. This case research was chosen because the need for GDP information is indispensable for better policy planning. The current condition in estimating Indonesia's GDP using economic variables, the data is incomplete and the release frequency of economic variables varies are monthly and quarterly. Therefore, convenient data were really needed,, and also abundant and available in real-time, one of the solutions is to use Google Trends as a predictor variable of Indonesia's GDP predictions. The use of selection variables on Google trends is intended to extract the most significant variables from the large number of Google Trends data. One of the selections variable used is the lasso in which lasso penalizes variables that do not significantly take effect in Indonesia's GDP. In penalizing the variable lambda value (?), it greatly affects the results of the selection variable. The determination of the lambda value on the lasso usually uses cross-validation where the value of each ? in a certain range will be evaluated entirely and then the ? value which has the lowest error is taken. Determination of ? in this way is less efficient so that an algorithm is developed that determines the lambda value based on the number of variables selected from each iteration as well as changes in the lasso convergence criterion which is when there are variables selected iteration for the ? value is terminated. So it is expected that a faster search for ? could increase the execution time and its accuracy value. The result of this experiment was showed 14 scenarios which were a combination of the methods and features of the selection to be tested. The test results showed that from the 14 scenarios performed, the results showed that the ARSO method using normalized weighting is the method with the best predictive value with a RMSE value of 0.74 which when compared to ARSO using vi conventional weighting reaches 1.19. In addition, ARSO with normalized weighting succeeded in capturing the pattern of economic crises that occurred during pandemic in 2020. The results of normalized ARSO were obtained using the lasso selection feature as weighting where the ? value was obtained from the algorithm proposed in this study. When it’s compared with the cross-validation algorithm proposed conducting the search ? managed to outperform both the execution time and the final accuracy of the method, namely for the execution time reaching 0.047 seconds and the RMSE value at 0.74 when compared to cross-validation is 3.9 seconds and the value RMSE 1.16. In addition, from this research it can be seen that data from Google Trends is able to reflect events that are happening in Indonesia, in this study the search keywords that indirectly reflect that the state of the Indonesian economy is declining based on the best prediction methods are dana_blt and pandemic.