Debiasing delle predizioni fotometriche del redshift nei quasar

This work addresses the significant challenges posed by selection biases in obtaining accurate photometric redshift estimations for quasars, which serve as essential probes of the early universe and cosmic struc- ture. We introduce a comprehensive debiasing methodology that improves these predictions in low redshift ranges, z = [0,4], by incorporating a selection bias correction term into the loss function of a neural net- work, effectively assigning higher weights to underrepresented quasar samples. Both homoscedastic and heteroscedastic regression models were evaluated, with the heteroscedastic model showing enhanced uncer- tainty estimation and achieving 91.2% accuracy within a 15% relative error threshold, alongside a reduction in outlier percentage to 16.87%. We introduce a debiased prediction framework first using an approxi- mated Gaussian selection function and then with an unapproximated empirical selection function to correct for selection biases. To prevent overcompensation in well-sampled regions, a flat turnover was applied to the unapproximated selection function in the low-redshift regime (0 ≤z ≤1.5). Additionally, sym- bolic regression was utilized to derive an interpretable model that maps photometric features to redshift: ZSR = 0.24 ×(bp−g) + sin[(MW1 )−0.6 ×(Mrp)] + 1.64. We then proceeded to gather visualizations of model predictions to reveal key trends influencing photometric redshift estimations. Higher astrometric excess noise, RUWE, and interstellar extinction (E(B-V)) correlate with larger deviations from true red- shifts, indicating the significant impact of observational uncertainties like poor astrometric fits and dust extinction on model performance. Sources with unreliable astrometric solutions or high extinction exhibit greater errors, highlighting the need for improved handling of these uncertainties. Furthermore, redshift versus position residuals from DECaLS images, considering standard deviation, show that higher redshift sources tend to have greater discrepancies due to morphological complexity and observational uncertainties. The relationship between positive residual counts and variability underscores the challenges of modeling high-variance sources. Finally, quasar classification using XGBoost on a multi-feature set, including stellar parameters and color indices, achieved a classification accuracy of 100%, demonstrating its effectiveness for high-performance quasar identification. By addressing selection biases and enhancing photometric redshift estimations, this research contributes to a more complete understanding of quasar demographics, enabling more accurate cosmological studies and advancing quasar evolution research.