Skip to content

Knowledge Graph

The MzeeChakula Knowledge Graph (KG) is the semantic brain of our system. Unlike traditional databases that store data in rows and columns, our KG stores data as entities (nodes) and relationships (edges), allowing the AI to reason about connections between food, health, and culture.

Ontology & Schema

The graph follows a strict ontology designed to capture the complexities of elderly nutrition in Uganda.

Core Entities (Nodes)

Entity Description Examples Count
Food Local food items available in Uganda. Matoke, Millet Porridge, Nile Perch 5,005
Nutrient Micro and macronutrients. Vitamin A, Iron, Protein 30
Condition Health conditions affecting the elderly. Hypertension, Type 2 Diabetes 4,852
Region Geographic regions of Uganda. Central, Northern, Western 5
Season Time periods affecting availability. Dry Season, Wet Season 12
Culture Cultural groups and practices. Baganda, Acholi 8

Core Relationships (Edges)

The power of the KG lies in how these entities are connected:

  • Food -[CONTAINS]-> Nutrient

    • Properties: amount (per 100g), unit (mg, g, ug).
    • Usage: Allows the system to calculate the nutritional value of a meal.

  • Nutrient -[AFFECTS]-> Condition

    • Properties: impact (Positive/Negative), mechanism (e.g., "Raises blood pressure").
    • Usage: The core reasoning path. If Salt AFFECTS Hypertension negatively, and Chips CONTAINS Salt, the system infers Chips are bad for Hypertension.

  • Food -[AVAILABLE_IN]-> Season

    • Usage: Ensures recommendations are practical. We won't recommend mangoes when they are out of season.

  • Food -[CULTURALLY_RELEVANT]-> Culture

    • Usage: Personalization. A Muganda elder might prefer Matoke, while an Acholi elder might prefer Millet.

Reasoning Capabilities

The Knowledge Graph enables several types of advanced reasoning:

1. Multi-Hop Reasoning

The system can traverse multiple edges to find indirect connections.

Query: "What should I eat for Anemia?" Path: Anemia <---(treats)--- Iron <---(contains)--- Liver Result: Recommend Liver.

2. Negative Inference

The system understands what not to recommend.

Query: "I have High Blood Pressure." Path: Hypertension <---(worsens)--- Sodium <---(contains)--- Salted Fish Result: Avoid Salted Fish.

3. Contextual Filtering

Combining health needs with environmental constraints.

Query: "Recommend a fruit for Vitamin C in January." Path: Vitamin C <---(contains)--- Orange ---(available_in)---> January Result: Recommend Orange (if available).

Graph Construction

The graph was built using a semi-automated pipeline:

  1. Data Ingestion: Structured data from food composition tables (FAO) and medical guidelines (WHO).
  2. Entity Linking: Mapping local food names (e.g., "Posho") to standardized ontology terms (e.g., "Maize Meal").
  3. Graph Projection: Stored in Neo4j for querying and exported as PyTorch Geometric tensors for GNN training.

Visualization

Below is a simplified visualization of a subgraph centered around "Matoke":

graph LR
    Matoke(Food: Matoke) -->|CONTAINS| K(Nutrient: Potassium)
    Matoke -->|GROWN_IN| Central(Region: Central)
    Matoke -->|AVAILABLE_IN| AllYear(Season: All Year)
    K -->|HELPS| Hypertension(Condition: Hypertension)

    style Matoke fill:#d90000,color:white
    style K fill:#fcdc04,color:black
    style Hypertension fill:#078930,color:white