Structured light is attracting significant attention for its diverse applications in both classical and quantum optics. The so-called vector vortex beams display peculiar properties in both contexts due to the non-trivial correlations between optical polarization and orbital angular momentum. Here we demonstrate a new, flexible experimental approach to the classification of vortex vector beams. We first describe a platform for generating arbitrary complex vector vortex beams inspired to photonic quantum walks. We then exploit recent machine learning methods -namely convolutional neural networks and principal component analysis -to recognize and classify specific polarization patterns. Our study demonstrates the significant advantages resulting from the use of machine learning-based protocols for the construction and characterization of high-dimensional resources for quantum protocols.Introduction-Light is endowed with Orbital Angular Momentum (OAM) [1, 2], a degree of freedom associated with structured, non-plane wavefronts, and characterized by an azimuthal phase dependence. When a nontrivial phase dependence is coupled with a helicoidal transverse polarization pattern, one talks of a Vector Vortex Beam (VVB) [2,3]. The interest in such states is motivated by the applications in multiple fields of classical and quantum optics [4,5]: from particle trapping to metrological applications in microscopy [6,7], and for OAM-based communications schemes in free-space and in-fibre [8,9]. VVBs are also often employed in quantum information protocols due to the hyperentanglement between their polarization and spatial degrees of freedom. Photonic platforms for quantum sensing and metrology leveraging such encoding have also been reported [10,11]. OAM-based schemes for investigating quantum causal structures [12], quantum communication and cryptography [13][14][15][16][17][18], quantum walks [19][20][21], quantum simulation [22,23], and quantum state engineering [24,25], have been previously demonstrated.Despite the potential of VVBs, many questions regarding the decoding of information stored in OAM and polarization remain unanswered. Various techniques of OAMdemultiplexing envisage the need of additional instruments -such as interferometry [26][27][28] or spatial filtering [29][30][31] -to be efficiently implemented. These introduce detrimental effects of loss and noise [32]. Moreover, the challenge of performing state tomography in such a high-dimensional framework, a fundamental task in quantum information processing [33,34], can hardly be overestimated. The design and demonstration of reliable techniques for the generation and classification of VVBs is thus highly desirable. Indeed, substantive efforts on finding novel platforms are subject of intense research activities [6,7,35,36], including in integrated photonics [37][38][39] and generation by plasmonic metasurfaces [40,41].Recently, Machine Learning (ML) has emerged as a versatile toolbox to tackle a variety of tasks arising in experi-