Please implement an experiment to test whether integrating dynamic graphlets with temporal motifs improves the accuracy of dynamic structural embeddings in capturing the temporal evolution of structural roles over weekly intervals, as measured by cosine similarity.

Experiment Overview

This experiment will compare three approaches for creating dynamic structural embeddings:
1. Baseline 1: Dynamic graphlets only
2. Baseline 2: Temporal motifs only
3. Experimental: Integrated approach combining dynamic graphlets and temporal motifs

The hypothesis is that the integrated approach will outperform both baselines in terms of cosine similarity between node embeddings across weekly intervals, as well as in downstream tasks like node classification and community detection.

Data Requirements

Please use a temporal network dataset that has:
- Timestamped interactions between nodes
- Weekly intervals (or data that can be aggregated into weekly snapshots)
- Sufficient temporal depth (at least 4-8 weeks of data)
- Node labels for classification tasks (if available)

Suggested datasets include:
- DBLP co-authorship network
- Email communication networks
- Social network interactions

Implementation Steps

1. Data Preparation

• Load the temporal network data
• Preprocess and clean the data
• Divide the data into weekly snapshots
• Create training and testing splits (e.g., use earlier weeks for training, later weeks for testing)

2. Baseline 1: Dynamic Graphlets Only

• Implement dynamic graphlet construction for each weekly snapshot
• Count the occurrences of different dynamic graphlet types for each node
• Create a feature vector for each node based on these counts
• Generate node embeddings using these feature vectors

3. Baseline 2: Temporal Motifs Only

• Implement temporal motif identification for the network
• Define a suitable null model to evaluate the significance of temporal motifs
• Count the occurrences of different temporal motif types for each node
• Create a feature vector for each node based on these counts
• Generate node embeddings using these feature vectors

4. Experimental: Integrated Approach

• Implement both dynamic graphlet construction and temporal motif identification
• Create a unified embedding framework that combines both approaches
• This could involve:
  a) Concatenating the feature vectors from both approaches
  b) Creating a weighted combination of the feature vectors
  c) Using one approach to guide or constrain the other
• Generate node embeddings using this integrated approach

5. Evaluation

• Primary Metric: Calculate cosine similarity between node embeddings across weekly intervals for all three approaches
• Secondary Metrics:
  a) Node classification accuracy (if labels are available)
  b) Community detection performance (modularity, NMI, etc.)
• Statistical Analysis: Compare the performance of the three approaches using appropriate statistical tests

Pilot Mode Implementation

Implement a global variable PILOT_MODE with three possible settings: 'MINI_PILOT', 'PILOT', or 'FULL_EXPERIMENT'.

MINI_PILOT Configuration

• Use only 2-3 weeks of data
• Select a small subset of nodes (e.g., 50-100 nodes)
• Run with minimal iterations/epochs for any optimization procedures
• Purpose: Quick debugging and verification of code functionality
• Expected runtime: 5-10 minutes

PILOT Configuration

• Use 4-6 weeks of data
• Select a moderate subset of nodes (e.g., 500-1000 nodes)
• Run with moderate iterations/epochs for optimization procedures
• Purpose: Verify if the approach shows promising results and if differences between methods are emerging
• Expected runtime: 30-60 minutes
• Use training data for training and validation data for evaluation

FULL_EXPERIMENT Configuration

• Use all available weeks of data
• Include all nodes in the network
• Run with full iterations/epochs for optimization procedures
• Purpose: Complete evaluation of the hypothesis
• Use training data for training, validation data for hyperparameter tuning, and test data for final evaluation

Start by running the MINI_PILOT first, then if everything looks good, proceed to the PILOT. After the PILOT completes, stop and do not run the FULL_EXPERIMENT (a human will manually verify the results and make the change to FULL_EXPERIMENT if appropriate).

Output Requirements

1. Log files containing:
2. Parameters used for each approach
3. Training progress (if applicable)
4. Intermediate results

1. Results file containing:
2. Cosine similarity scores for each approach across weekly intervals
3. Node classification accuracy (if applicable)
4. Community detection performance metrics
5. Statistical comparison between approaches
6. Visualizations of the embeddings (e.g., t-SNE or PCA projections)

1. Saved models and embeddings for each approach

1. A summary report with:
2. Description of the implemented methods
3. Key findings
4. Statistical analysis results
5. Conclusions regarding the hypothesis

Please implement this experiment with clear, well-documented code that follows best practices for reproducibility.