AI for Historical Structure Classification & Tourism Recommendation
Image classification for heritage structures (TensorFlow) + exploratory analytics and recommendation engine for tourism.
Problem Statement
Historical structures preserve cultural heritage and attract tourism. A government agency wants to use machine learning to:
- Monitor conditions of historical structures automatically (image-based classification).
- Understand tourists and recommend places to improve marketing and engagement.
Part 1 — Image Classification (TensorFlow / ResNet50)
Objective
Predict the category (one of 10 classes) of a structure from an image to support automated monitoring.
Dataset
- Training images + separate test set across 10 categories.
- Training split further into training / validation.
Model & Training
- Base model: ResNet50 (pretrained), with custom dense + dropout layers.
- Loss:
sparse_categorical_crossentropy. - Regularization: early stopping + dropout.
- Training epochs: 50 planned, early stopped at 26 (non-augmented) and 15 (augmented).
- Also trained on augmented images for improved robustness.
Model Snippet
# CNN architecture
base_model = tf.keras.applications.ResNet50(
input_shape=(224,224,3),
include_top=False,
weights='imagenet' #Initialize the weights (parameters) using the model that was already trained on ImageNet.
)
#Add the layers
model = models.Sequential([
base_model,
layers.GlobalAveragePooling2D(), #Flatten- for transfer we use this layer rather than flat layer
layers.Dense(128, activation='relu'),
layers.Dropout(0.2), # Dropout for regularization
layers.Dense(10, activation='softmax') # 10 classes
])
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
Training Summary
Conclusion (Part 1)
Validation accuracy > Training accuracy and Validation loss < Training loss, indicating strong generalization. Augmentation increased robustness but made training accuracy lower due to higher data variability. Additional data or longer training could further improve performance.
Part 2 — Tourism Analytics & Recommendation Engine
Objective
Perform EDA and build recommenders to help tourists discover places of interest and guide tourism marketing.
Data Preparation
- Merged three datasets into a single DataFrame.
- Cleaned data and removed irrelevant columns.
- Translated Indonesian text to English (Google Translate) for analysis.
Key Analytics Findings
| Insight | Summary |
|---|---|
| Top rating age group | Users aged 25–35 provided most ratings |
| Top tourist origin | Bekasi, Jawa Barat |
| Popular cities | Bandung, Jakarta, Yogyakarta City |
| Top category by visits | Amusement Parks |
| Highest-rated category | Nature preserves |
Recommendations Built
- Cosine similarity — collaborative approach on user-place rating matrix (low similarity scores; mostly < 0.4).
- GenAI recommender — generative/contextual recommendations using combined dataset; produced category-aware suggestions and often diverged from cosine-based results.
Takeaways
- Cosine similarity struggled due to sparse overlaps in user ratings.
- Generative AI gave more thematic recommendations.
Conclusion (Part 2)
The analytics uncovered clear demographic and location patterns (top cities and categories). The hybrid approach (collaborative + generative) can be used together: collaborative filtering where data overlap exists and GenAI for contextual, category-based suggestions.
Charts & Diagrams
Tech Stack & Future Improvements
Tech Stack
- Deep Learning: TensorFlow, Keras, ResNet50
- Data Analysis: Pandas, NumPy
- Visualization: Matplotlib, Seaborn
- Recommendation: Cosine Similarity, GenAI
- Translation: Google Translator API
Future Improvements
- Increase and diversify the image dataset; fine-tune deeper layers of ResNet50.
- Combine collaborative and content-based recommenders into a hybrid system.
- Deploy a web dashboard for real-time monitoring and recommendations.
- Collect more explicit user feedback to improve collaborative filtering.