PREDICTING FOLLOWER GROWTH FOR SOCIAL MEDIA BLOGGERS AND MODELING USER FOLLOWING BEHAVIOR USING XGBOOST AND RANDOM FOREST ENSEMBLE LEARNING

This study addresses user interaction dynamics on social media platforms by constructing separate models for predicting new follower counts for bloggers and classifying users' targeted following behavior. Quantitative analysis of new follower counts reveals a high correlation between interaction behavior and follower growth, with correlation coefficients exceeding 0.89. By constructing an XGBoost model that utilizes second-order Taylor series expansion to optimize the objective function and incorporates 1- to 3-day lagged features to capture time-series patterns, the model demonstrated robust performance on the test set with an average absolute error of 20.71. For the user targeted following behavior classification prediction, a Random Forest model was selected to address the low-dimensionality and strong nonlinearity of the features. The study employed target encoding strategies to smooth user and blogger IDs, preventing overfitting, and defined interaction intensity formulas based on behavior weights. Experimental results show the Random Forest model achieved an AUC of 0.83 and an F1 score of 74.25% in user follow prediction tasks, effectively enhancing behavioral prediction accuracy. This study provides quantitative insights into user growth dynamics and micro-interaction patterns on social platforms through ensemble learning algorithms.