ACLM Model: A CNN-LSTM and Machine Learning Approach for Analyzing Tourist Satisfaction to Improve Priority Tourism Services

Abstract

Tourist satisfaction is a key proxy for destination service quality, yet automatic sentiment analysis of online reviews still faces class imbalance, overfitting, and limited deployability. This study proposes ACLM, a hybrid sentiment classification pipeline that learns semantic and temporal features with a CNN-LSTM backbone and evaluates three classifier heads (Softmax, Logistic Regression, XGBoost) on a three-class corpus (neutral, satisfied, dissatisfied). The objective is to deliver an accurate and operational model for decision support in tourism services. The idea combines Word2Vec embeddings, a compact CNN for local patterns, an LSTM for sequence dependencies, and a training workflow with text cleaning, SMOTE based balancing, and regularization to curb overfitting; outputs are exposed through a simple Streamlit interface. Results show that CNN-LSTM with a Softmax head attains accuracy 0.89, macro precision 0.89, macro recall 0.84, and macro F1 0.86, outperforming Logistic Regression (accuracy 0.87, macro precision 0.84, macro recall 0.82, macro F1 0.82) and XGBoost (accuracy 0.85, macro precision 0.80, macro recall 0.82, macro F1 0.80). The findings indicate that deep sequence features paired with a simple Softmax head provide the best tradeoff between accuracy and stability for three-way sentiment classification. The contribution is a reusable, end to end blueprint from preprocessing and balanced training to quantitative evaluation and an inference GUI, and the novelty lies in testing interchangeable classifier heads on a single CNN-LSTM feature extractor while explicitly addressing data imbalance and deployment constraints. The GUI is implemented using the highest accuracy model, namely CNN-LSTM with Softmax.