StripeSankey 0.2.0

Interactive Sankey diagram widget for topic modeling analysis with sample flow tracing

Stripe Sankey Widget

An interactive Sankey diagram widget for understanding how samples are represented by topics across different LDA models. This tool can also be applied to similar algorithms like NMF and LSA.

LDA is a powerful tool for achieving soft clustering and finding latent groups beyond initial hypotheses. However, the challenge of applying LDA to 16S rRNA data is that sequence reads are less informative than human language when users want to evaluate a model's results. Even when metrics are applied, making decisions remains difficult.

We designed this StripeSankey diagram to make LDA results more accessible and integrated two widely used metrics—perplexity and coherence score—into novel visual encodings.

Overview of StripeSankey diagram:

Legend of metrics coour:

After click one flow:

Installation

From PyPI

pip install StripeSankey

Data Preprocessing

Quick Start

import stripe_sankey
from stripe_sankey import StripeSankeyInline

# Load your processed topic modeling data
sankey_data = {
    "nodes": {
        "K2_MC1": {
            "high_count": 45,
            "medium_count": 23,
            "model_metrics": {"perplexity": 1.2},
            "mallet_diagnostics": {"coherence": -0.15}
        },
        # ... more nodes
    },
    "flows": [
        {
            "source_segment": "K2_MC1_high",
            "target_segment": "K3_MC2_medium", 
            "source_k": 2,
            "target_k": 3,
            "sample_count": 15,
            "samples": [{"sample": "doc_1", "source_prob": 0.8, "target_prob": 0.6}]
        },
        # ... more flows
    ],
    "k_range": [2, 3, 4, 5]
}

# Create and display widget
widget = StripeSankeyInline(sankey_data=sankey_data)
widget

Visualization Modes

Default Mode

Shows topic representations with high/medium probability segments:

widget = StripeSankeyInline(sankey_data=data, mode="default")

Metric Mode

Color-codes topics by quality metrics (perplexity + coherence):

widget = StripeSankeyInline(sankey_data=data, mode="metric")

# Customize metric weights
widget.update_metric_config(red_weight=0.9, blue_weight=0.7)

Interactive Features

Sample Flow Tracing

• Click any flow to highlight sample trajectories across K values
• Orange trajectories show where samples move between topics
• Count badges display number of traced samples in each segment
• Line thickness represents sample flow volume
• Click background to clear selection

Visual Elements

• Stacked bars: High (dark) and medium (light) probability representations
• Curved flows: Proportional thickness based on sample counts
• Barycenter layout: Optimized positioning to reduce visual complexity
• Hover tooltips: Detailed information on flows and segments

Data Format

Your data should follow this structure:

{
    "nodes": {
        "K{k}_MC{mc}": {
            "high_count": int,           # Samples with prob ≥ 0.67
            "medium_count": int,         # Samples with prob 0.33-0.66
            "total_probability": float,
            "model_metrics": {
                "perplexity": float      # Lower is better
            },
            "mallet_diagnostics": {
                "coherence": float       # Higher (less negative) is better
            }
        }
    },
    "flows": [
        {
            "source_segment": "K{k}_MC{mc}_{level}",
            "target_segment": "K{k+1}_MC{mc}_{level}",
            "source_k": int,
            "target_k": int,
            "sample_count": int,
            "average_probability": float,
            "samples": [
                {
                    "sample": str,           # Sample identifier
                    "source_prob": float,    # Probability in source topic
                    "target_prob": float     # Probability in target topic
                }
            ]
        }
    ],
    "k_range": [2, 3, 4, 5]  # Topic numbers analyzed
}

Configuration Options

Widget Parameters

widget = StripeSankeyInline(
    sankey_data=data,
    width=1200,           # Canvas width
    height=800,           # Canvas height  
    mode="default"        # "default" or "metric"
)

Metric Mode Configuration

widget.update_metric_config(
    red_weight=0.8,       # Perplexity influence (0-1)
    blue_weight=0.8,      # Coherence influence (0-1) 
    min_saturation=0.3    # Minimum color brightness
)

Color Schemes

widget.color_schemes = {
    2: "#1f77b4",  # Blue for K=2
    3: "#ff7f0e",  # Orange for K=3
    4: "#2ca02c",  # Green for K=4
    5: "#d62728"   # Red for K=5
}

Use Cases

• Topic Model Analysis: Understand how topics evolve across different K values
• Sample Trajectory Tracking: Follow samples through topic assignments
• Model Quality Assessment: Visual comparison of perplexity and coherence metrics
• Flow Bottleneck Detection: Identify where samples cluster or disperse
• Research Presentation: Interactive demonstrations of topic modeling results

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Built with anywidget - modern Jupyter widget framework
• Visualization powered by D3.js

Support

• 📖 Documentation