Personal note: You should know that I had very little knowledge of machine learning and deep learning before starting this project. I had to go through a lot of content and courses (Thank you Andrew Ng and folks at Google BI specialization!) while implementing this project, and I hope it can help, and inspire others. As I mentioned, I am relatively new to the tech industry and data science, so constructive criticism would be greatly appreciated.
Dashboard Preview:
Recording.2024-07-29.134052.mp4
The Spending Tracker project leverages AI to categorize financial transactions, aiming to inspire the tech community to integrate AI into daily tasks. By fine-tuning the DistilBERT model, I achieved a multiclass classification system that sorts transactions into 20 distinct categories. The open-source model ensures data privacy as it operates locally, not requiring internet connectivity for processing personal bank data. The beauty of fine-tuning is that you can get personized result. I tested the trained model with my data and then compared its performance against an online service provider (Wallet by budget bakers). My model performed 15% better than the OSP.
The goal of this project is to:
- Demonstrate the effectiveness of fine-tuning open-source models on personal data.
- Create a system to track financial spending
- Encourage non-data scientists to experiment with AI models.
- Fine-tuned DistilBERT Model: Classifies transactions into 20 categories based on transaction notes.
- Privacy-Focused: Processes data locally, ensuring no personal data is shared online.
- Performance: Outperforms a paid service provider by 15% in categorizing transactions.
Performance test:
- Model Training: Notebooks,scripts, for training the DistilBERT model along with the saved model. Perhaps the most important.
- Dashboarding: Processing data for further visualization.
- Data Extraction: Scripts for extracting data from PDF bank statements. This might not for everyone. This was coded specifically for myself, but I included it in the git hub respository.
- Dataset: Approximately 300-400 rows of text data with corresponding labels for transaction categories.
- Combined Text: Merges transaction notes with amount information (e.g., "Customer Transfer Cr. INVESTMENT PROCEEDS: $678.64").
- Class Imbalance Handling: Techniques like oversampling used to address class imbalances.
- Training: Utilizes the Trainer API from Hugging Face Transformers for efficiency.
- Setup: Small model with 110M parameters trained on a GPU with 6GB VRAM (Can work with CPU if you have enough RAM).
- Clone the repository:
git clone https://github.com/prayag-purohit/Financial_tracker.git
- Navigate to the project directory:
cd Financial_tracker - Follow the instructions in the Jupyter notebooks within subdirectories to set up and use the model.
- Python 3.x
- Jupyter Notebook
- Required Python libraries: Transformers, sklearn, torch, camelot, imblearn, subprocesses, os, pandas, numpy, Matplotlib
-
Install the required libraries:
pip install -r requirements.txt
-
Launch Jupyter Notebook and edit the notebooks to begin categorizing your transactions.
- Experiment with text-generation models for comparison.
- Remove classification heads from pre-trained models for better personalization.
- Implement relational database structures instead of using CSVs.
- Data augmentation is very hard for transaction data with conventional methods. Suggestions are highly appreciated.
- Small Dataset: Limited to 300 rows due to the sensitive nature of bank transaction data.
- Overfitting Risk: High risk due to the small dataset size, mitigated by careful monitoring and regularization techniques.
- Hardcoded paths
This project welcomes contributions and constructive criticism. If you are from a different background or have suggestions for improvement, feel free to contribute or provide feedback.
Detailed documentation and usage instructions can be found within the subdirectories in the Jupyter notebooks.