The MARFIN project revolutionizes financial risk assessment by employing Multimodal Analogical Reasoning over knowledge graphs. This innovative approach integrates diverse financial data sources, including market trends, asset performance, and economic indicators, to uncover hidden relationships and patterns. By fusing visual and textual cues, researchers aim to enhance genre classification accuracy, while financial institutions seek to mitigate potential losses and optimize investment strategies using MARFIN's advanced risk identification techniques.
- Multimodal Knowledge Graph Integration: Combines various financial data types such as textual reports, numerical data, and multimedia elements like graphs or charts.
- Financial Data Relationships: Uncovers and represents complex relationships between entities such as companies, financial instruments, economic indicators, and market trends.
- Advanced Risk Identification: Utilizes multimodal data to enhance risk assessment and identification techniques.
- IMDB Multimodal Dataset: Leverages the IMDB Multimodal Vision & NLP Genre Classification dataset, offering a curated collection of movie posters and detailed plot summaries for genre classification.
In MARFIN, a multimodal knowledge graph integrates various types of financial data. For instance:
- Nodes: Represent entities such as companies, financial instruments, economic indicators, or market trends.
- Edges: Denote relationships between these entities (e.g., ownership, correlation, causality).
The FinancialKnowledgeGraph class constructs the knowledge graph. This example includes:
- Entities: "StockMarket," "MarketCrash," "CreditDefault."
- Relationships: "Affects" and "LeadsTo."
You can extend and modify this knowledge graph to represent more complex relationships in financial data.
The MultimodalEmbeddingGenerator class generates embeddings for nodes in the knowledge graph based on multiple modalities (text, image, and numerical attributes). While the example uses a basic linear layer, real-world scenarios would likely use sophisticated models such as multimodal transformers or graph convolutional networks (GCNs) to capture complex relationships within and between modalities.
The goal of this project is to develop a predictive model for bankruptcy using a dataset with various financial ratios and features. The project includes:
- Data exploration
- Data preprocessing
- Model training using Random Forest with hyperparameter tuning
- Model evaluation
- Feature importance visualization
- Knowledge graph construction
- Probability distribution visualization for bankruptcy prediction
The following Python packages are required to run the code:
- pandas
- numpy
- matplotlib
- seaborn
- scikit-learn
- imbalanced-learn
- pickle
- networkx
The data were collected from the Economic Journal for the years 1999 to 2009.
Install the required packages using pip:
pip install pandas numpy matplotlib seaborn scikit-learn imbalanced-learn networkx