In ground-based EAS (Extensive Air Shower) array experiments for gamma-ray observations, hadron-induced background cosmic-ray showers constitute over 99.9$\%$ of the shower particles detected at ground level compared to those originated by gamma rays. Therefore, a crucial step in analyzing data from these ground arrays involves the essential removal of these hadron-induced background showers. This work utilizes data recorded by both the surface array and underground muon detector array of the Tibet AS$\gamma$ experiment to discriminate between gamma-ray and hadron-induced showers within the energy range of 3 TeV to 1 PeV. The traditional gamma/hadron separation relies on two parameters: the sum of the particle density ($\sum{\rho}$) measured by the AS array and the number of muons ($\sum{N_{\mu}}$) measured by the MD array. This work introduces machine learning (ML) methods (MLP, XGBoost, LightGBM) that incorporate additional characteristic parameters—shower core position ($c_{x}$, $c_{y}$) and zenith angle ($\theta$) to improve the discrimination ability of gamma/hadron. Our study reveals that ML methods demonstrate significantly improved performance in gamma/hadron separation compared to conventional approaches in the high-energy (HE, E > $\sim$10 TeV) range. ML-based selections achieve over 10 times higher cosmic-ray rejection above 60 TeV, making these methods highly effective for studying PeVatrons. Moreover, ML methods provide more uniform discrimination across the entire detector area, unlike the traditional method which performs poorly outside the muon detector coverage. In contrast, in the low-energy (LE) mode (E < $\sim$10 TeV), where events are restricted to the central dense array, the performance of ML methods is comparable to the traditional approach, as both rely on selecting events with $\sum{N_{\mu}} = 0$.