@phdthesis{digilib77432,
           month = {June},
           title = {OPTIMISASI REGRESI BINOMIAL NEGATIF LASSO
BERBASIS AIC PADA FITUR TEKSTUAL PROMOSI DI
PLATFORM X (STUDI KASUS: @SHOPEEID)},
          school = {UIN SUNAN KALIJAGA YOGYAKARTA},
          author = {NIM.:  22106010078 Dea Iswari},
            year = {2026},
            note = {Sri Utami Zuliana, S.Si., M.Sc., Ph.D.},
        keywords = {binomial negatif; overdispersi; regularisasi LASSO; AIC; TF-IDF},
             url = {https://digilib.uin-suka.ac.id/id/eprint/77432/},
        abstract = {Overdispersion in count data leads to a violation of the assumption of equidispersion, meaning that the Poisson distribution is no longer capable of adequately modeling the variability of the data. Negative Binomial regression is used as an alternative approach by adding a dispersion parameter to accommodate variance that exceeds the mean. However, modeling becomes complex when it involves predictor variables in the form of textual features extracted via TF-IDF, which may lead to redundancy and multicollinearity among features. To produce a parsimonious model, LASSO regularization with an L1 penalty is applied, which shrinks the coefficients of variables with small contributions to exactly zero. The penalty parameter {\ensuremath{\lambda}} is determined by minimizing the AIC criterion to achieve a balance between model complexity and performance. The implementation of this method on textual features from the @ShopeeID account?s promotional posts on the X platform demonstrates that AIC-based Negative Binomial LASSO regression effectively reduces model dimensionality by eliminating 30 redundant variables. Although the model becomes more compact, its explanatory power remains relatively consistent, as indicated by a very small decrease in McFadden?s Pseudo R2, specifically, 0.0013 compared to the full model. The research results show that the integration of Negative Binomial regression and LASSO regularization is effective in improving model interpretability without significantly sacrificing statistical quality.}
}