AI-driven insights into the impact of tourism on local cultures: a machine learning approach

Tourism is increasingly recognized as a powerful driver of economic growth, spatial restructuring, and cultural change, yet empirical evidence on how tourism intensity and short-term rentals reshape local cultures remains fragmented. Existing research is often split between qualitative case studies of touristification and gentrification, and quantitative models that forecast visitor flows but provide limited insight into downstream cultural externalities such as authenticity loss, commodification, noise, and crowding. Moreover, many high-performing machine learning forecasters operate as opaque black boxes, hindering their adoption in policy-facing tourism analytics. This study proposes an AI-driven framework that integrates multi-source tourism signals with culture-sensitive outcomes and an interpretable, long-horizon forecaster. Our primary novelty is methodological: rather than introducing new tourism concepts, we contribute a forecasting and interpretability model that operationalizes established tourism-theory constructs at a monthly scale and links them to upstream tourism/STR signals through transparent temporal attribution. Grounded in debates in tourism studies on overtourism, touristification, and authenticity, we treat "cultural perception" not as a generic sentiment proxy but as a contested evaluative construct through which visitors and residents negotiate the meaning of place amid intensifying platform-mediated tourism. In this framing, short-term rentals (STRs) are conceptualized as socio-technical infrastructures that reconfigure the conditions of "local experience" (everyday life, encounter, and consumption), shaping how authenticity, commodification, noise, and crowding are perceived and narrated. We assemble three complementary panels: regional tourism pressure (Eurostat nights per resident), city-scale short-term rental (STR) micro-intensity (Inside Airbnb), and monthly cultural-perception indices derived from Yelp reviews and photos (authenticity, commodification, noise, crowding). We introduce the Patch-Agnostic Temporal Transformer (PATT), an encoder-only architecture that combines reversible instance normalization, a Local Perception Unit for short-range denoising, and offset-guided sparse attention that learns to attend to a compact set of temporally salient events without relying on large token patches. Cross-modal analysis exploits attention maps and lagged correlations to surface lead-lag linkages between tourism/STR shocks and subsequent shifts in cultural perceptions. Because the cultural-perception indices are constructed from platform reviews and photos, we interpret the observed lead-lag structure as a temporally structured association in discursive perceptions rather than a direct measurement of cultural welfare or causal impact. On the Eurostat dataset, PATT attains MASE [Formula: see text] and CRPS [Formula: see text]; on Inside Airbnb STR micro-intensity, MASE [Formula: see text] and CRPS [Formula: see text]; and on Yelp cultural-perception indices, MASE [Formula: see text] and CRPS [Formula: see text], while yielding interpretable temporal attributions and cross-modal alignment diagnostics that support monitoring and hypothesis generation about tourism-related shifts in review-derived perception proxies. Code is archived at https://doi.org/10.5281/zenodo.18980783 .