4/ AI Transparency and Interpretability

Dash integrates cutting-edge Explainable AI (XAI) methodologies to ensure its predictions, alerts, and insights are fully traceable and understandable. This commitment to transparency empowers users by offering detailed explanations for every decision, enabling them to validate the AI’s conclusions and build trust in the system.

Core Principles of Explainable AI

1. Feature Importance Analysis:

   • Dash identifies the most influential variables (features) that contribute to each prediction or alert. These include:

     • Transaction volume (TvT_vTv​).

     • Liquidity metrics (LLL).

     • Sentiment scores (SSS).

     • Historical price trends (PtP_tPt​).

   • SHAP Values (SHapley Additive ExPlanations):

     • Quantify the contribution of each feature to the AI’s output:

       SHAP(fi)=E[f(X)∣Xi=xi]−E[f(X)]SHAP(f_i) = E[f(X) | X_i = x_i] - E[f(X)]SHAP(fi​)=E[f(X)∣Xi​=xi​]−E[f(X)]

       Where:

       • f(X)f(X)f(X): Model output.

       • E[f(X)]E[f(X)]E[f(X)]: Average model output over all feature values.

       • Xi=xiX_i = x_iXi​=xi​: The specific feature value being evaluated.

     • Example:

       • If a token price prediction outputs +5%+5\%+5%, and TvT_vTv​ contributes +2%+2\%+2%, LLL contributes +1.5%+1.5\%+1.5%, and SSS contributes +1.5%+1.5\%+1.5%, SHAP values highlight these proportions.

2. Traceable Pipelines:

   • Every AI-generated alert or prediction includes metadata outlining the underlying data sources, model parameters, and decision logic. This ensures full traceability.

   • Example:

     • For a high-priority alert on liquidity shifts:

       • Source: Raydium API.

       • Metric: Liquidity drop of 40%.

       • Model Used: Bayesian Inference.

       • Confidence Score: 92%.

Interpretable Visualizations

Dash provides dynamic visualizations to help users intuitively understand the AI’s conclusions:

1. Feature Contribution Graphs:

   • Display the relative importance of each feature in a prediction.

   • Example: A bar graph showing:

     • Sentiment Score: 40%.

     • Liquidity Shift: 30%.

     • Price Momentum: 30%.

2. Decision Trees:

   • Simplify complex predictions by showing step-by-step decision paths.

   • Example:

     • A decision tree for predicting a token’s price surge:

       • If Tv>10,000T_v > 10,000Tv​>10,000, move to liquidity conditions.

       • If L>500,000L > 500,000L>500,000, sentiment is evaluated.

       • If S>0.3S > 0.3S>0.3, output = High Confidence of Surge.

3. Real-Time Alerts Dashboard:

   • Interactive UI elements allow users to view detailed explanations for each alert with one click.

Confidence Scoring

Dash attaches a confidence score to every output, reflecting the AI’s certainty in its predictions. This score is calculated as:

C=Weighted Sum of Model AgreementsTotal ModelsC = \frac{\text{Weighted Sum of Model Agreements}}{\text{Total Models}}C=Total ModelsWeighted Sum of Model Agreements​

• Example:

  • If three models predict a price increase with probabilities of 0.8, 0.7, and 0.9: C=0.8+0.7+0.93=0.8 (80%)C = \frac{0.8 + 0.7 + 0.9}{3} = 0.8 \, (80\%)C=30.8+0.7+0.9​=0.8(80%)

Audit Logs

1. Actionable Insights:

   • Every decision is logged with timestamps, source data, and model parameters, ensuring that all outputs are auditable.

2. User Access:

   • Logs are accessible through the dashboard, allowing users to review:

     • Input data (e.g., transaction volumes, sentiment scores).

     • Model outputs (e.g., price predictions, anomaly detections).

     • Decision logic (e.g., thresholds and weights).

Example: Explaining a Liquidity Alert

Scenario:

• An alert indicates a 40% liquidity drop for a token.

1. Feature Importance:

   • SHAP values show:

     • Liquidity metric: 50%.

     • Transaction spike: 30%.

     • Sentiment score: 20%.

2. Traceability:

   • Data Source: Raydium API.

   • Time of Event: 14:32 UTC.

   • Model: Anomaly Detection via GNN.

3. Confidence:

   • Confidence Score: 88%.

4. Visualization:

   • A bubble map highlights wallets responsible for the liquidity withdrawal.

Future Enhancements

1. Natural Language Explanations:

   • The AI will generate textual summaries for every prediction:

     • Example: “This token is predicted to increase in price by 8% due to high trading volume and a positive sentiment score of 0.6.”

2. Explainable Reinforcement Learning (XRL):

   • Provides insights into the decision-making process of RL agents:

     • Example: Why a buy action was recommended in a specific liquidity condition.