Research on Construction of Rural Tourism Management Platform Based on XGBoost-CNN

To address issues such as data fragmentation, inaccurate demand forecasting, and inefficient resource allocation in rural tourism management, this paper proposes an intelligent management platform model that integrates eXtreme Gradient Boosting (XGBoost) with a Convolutional Neural Network (CNN). Regarding data acquisition, the model aggregates multi-source, heterogeneous data—including tourist profiles, scenic area environmental monitoring data, and social media content—encompassing both structured data and unstructured image data. In terms of data processing, structured data undergoes cleaning, feature engineering, and normalization before being fed into the XGBoost module for high-precision forecasting of tourist volume and consumption preferences; unstructured image data undergoes denoising, enhancement, and size normalization before entering the CNN module for tasks such as scenic area popularity recognition, safety hazard detection, and landscape quality assessment. The core innovation of our model lies in its architectural design and fusion strategy. This model adopts a parallel dual‑branch feature extraction architecture and performs a feature‑level fusion strategy based on this architecture. Specifically, the feature vectors output by the two branches are concatenated after dimensional alignment, and the concatenated vector is fed into a fully connected decision network. This fusion strategy differs from result‑level ensemble methods in the structure of the information merging stage. By integrating deep features from both branches at an early‑to‑mid stage rather than merely averaging or voting on final predictions, our approach preserves cross‑modal correlations and enables richer interaction between structured and visual cues, which is a key distinction from conventional hybrid architectures that rely on late fusion or output‑level integration. Experimental results based on approximately 870,000 data samples from a rural tourism demonstration zone in a certain province (covering the period from 2023 to 2025) indicate that the hybrid model achieved an Root Mean Square Error (RMSE) of 0.32—a reduction of 21.3% compared to the standalone XGBoost model and 25.6% compared to the standalone CNN model. Additionally, the F1-score for scenic area popularity recognition reached 0.92, representing an improvement of 0.09 over XGBoost and 0.14 over CNN.