CultureConnect Machine Learning

Focusing on the machine learning aspects of our project CultureConnect, an app designed to help users understand their emotions and discover suitable cafe recommendations based on their mood.

📊 Tasks

1. Data Collection:

• Collect mood-related text data from various sources.
• Gather cafe information tailored to different emotional states.
• Compile a comprehensive dataset of emotional expressions and cafe environments.

2. Data Labeling:

• Label the dataset with the following mood categories:
  • Sadness
  • Joy
  • Love
  • Anger
  • Fear
  • Surprise

3. Modeling:

Build three distinct models:

• Mood Detection Model: Utilize the Bidirectional LSTM with Embedding Neural Network architecture for emotion classification.
• Cafe Recommendation Model: Implement content-based filtering using TensorFlow and cosine similarity.
• Text Summarization: Fine-tune the architecture to generate mood-based insights.

🔧 Model Architecture

• Text Preprocessing:

  • Text data is tokenized and padded for uniform input sequences.
  • Employed one-hot encoding for mood labels.

• Mood Detection Model:

  • Built using a sequential neural network architecture:

    • Embedding Layer: Maps words to dense vector representations.
    • Bidirectional LSTM: Captures both past and future dependencies in text.
    • Global Max Pooling Layer: Aggregates sequence information.
    • Dense Layers: Fully connected layers for classification.

  • Optimized using cross-entropy loss and Adam optimizer.

    

• Cafe Recommendation System:

  • Employs cosine similarity to match user moods with cafe data embeddings.

  • Determines recommendations based on user reviews and lighting levels of cafes.

    

📃 Dataset

• Mood Detection Training Data: Labeled and weighted text data.
• Cafe Recommendation Database: Information on cafes, including names, locations, ratings, and emotional suitability.

🎨 Model Demo

Explore our model demonstrations:

• Mood Detection Model and Cafe Recommendation System

---

🔢 Model Evaluation

Comprehensive analysis of model performance, including:

• Accuracy metrics.
• Confusion matrix.
• Emotional classification confidence scores.

---

📖 References

1. Sanh, V., et al. (2020). DistilBERT, a distilled version of BERT: smaller, faster, cheaper, and lighter.
2. Houlsby, N., et al. (2019). Parameter-Efficient Transfer Learning for NLP.
3. Agrippina Fleta (2021). Analisis Pencahayaan Alami dan Buatan pada Ruang Kantor terhadap Kenyamanan Visual Pengguna. JURNAL PATRA Vol. 3 No. 1. Available Online
4. Selfiyani Lestari, Bagus Takwin, Dianti Endang Kusumawardhani (2021). Pencahayaan Baik untuk Emosi yang Positif: Analisis Emosi Saat Malam Hari Berdasarkan Penilaian Terhadap Pencahayaan Lokasi. INQUIRY Jurnal Ilmiah Psikologi Vol. 12 No. 1, hlm 38-52.
5. Rr. Puruwita Wardani (2018). Pengaruh Mood Konstruktif dan Tidak Konstruktif terhadap Pengambilan Keputusan dalam Audit. JURNAL ONLINE INSAN AKUNTAN Vol. 3 No. 1, Juni 2018.
6. Peilu Wang, Yao Qian, Frank K. Soong (2015). A Unified Tagging Solution: Bidirectional LSTM Recurrent Neural Network with Word Embedding. arXiv. DOI:10.48550/arXiv.1511.00215.
7. Shi, J., Ye, M., Chen, H. et al. (2023). Enhancing Efficiency and Capacity of Telehealth Services with Intelligent Triage: A Bidirectional LSTM Neural Network Model Employing Character Embedding. BMC Med Inform Decis Mak 23, 269. DOI:10.1186/s12911-023-02367-1.