exploratory-data-analysis（探索性数据分析）

Exploratory Data Analysis

Discover patterns, anomalies, and relationships in tabular data through statistical analysis and visualization.

Supported formats: CSV, Excel (.xlsx,.xls), JSON, Parquet, TSV, Feather, HDF5, Pickle

Standard Workflow

1. Run statistical analysis:

python scripts/eda_analyzer.py <data_file> -o <output_dir>

1. Generate visualizations:

python scripts/visualizer.py <data_file> -o <output_dir>

1. Read analysis results from <output_dir>/eda_analysis.json
2. Create report using assets/report_template.md structure
3. Present findings with key insights and visualizations

Analysis Capabilities

Statistical Analysis

Run scripts/eda_analyzer.py to generate comprehensive analysis:

python scripts/eda_analyzer.py sales_data.csv -o ./output

Produces output/eda_analysis.json containing:

• Dataset shape, types, memory usage
• Missing data patterns and percentages
• Summary statistics (numeric and categorical)
• Outlier detection (IQR and Z-score methods)
• Distribution analysis with normality tests
• Correlation matrices (Pearson and Spearman)
• Data quality metrics (completeness, duplicates)
• Automated insights

Visualizations

Run scripts/visualizer.py to generate plots:

python scripts/visualizer.py sales_data.csv -o ./output

Creates high-resolution (300 DPI) PNG files in output/eda_visualizations/:

• Missing data heatmaps and bar charts
• Distribution plots (histograms with KDE)
• Box plots and violin plots for outliers
• Correlation heatmaps
• Scatter matrices for numeric relationships
• Categorical bar charts
• Time series plots (if datetime columns detected)

Automated Insights

Access generated insights from the "insights" key in the analysis JSON:

• Dataset size considerations
• Missing data warnings (when exceeding thresholds)
• Strong correlations for feature engineering
• High outlier rate flags
• Skewness requiring transformations
• Duplicate detection
• Categorical imbalance warnings

Reference Materials

Statistical Interpretation

See references/statistical_tests_guide.md for detailed guidance on:

• Normality tests (Shapiro-Wilk, Anderson-Darling, Kolmogorov-Smirnov)
• Distribution characteristics (skewness, kurtosis)
• Correlation methods (Pearson, Spearman)
• Outlier detection (IQR, Z-score)
• Hypothesis testing and data transformations

Use when interpreting statistical results or explaining findings.

Methodology

See references/eda_best_practices.md for comprehensive guidance on:

• 6-step EDA process framework
• Univariate, bivariate, multivariate analysis approaches
• Visualization and statistical analysis guidelines
• Common pitfalls and domain-specific considerations
• Communication strategies for different audiences

Use when planning analysis or handling specific scenarios.

Report Template

Use assets/report_template.md to structure findings. Template includes:

• Executive summary
• Dataset overview
• Data quality assessment
• Univariate, bivariate, and multivariate analysis
• Outlier analysis
• Key insights and recommendations
• Limitations and appendices

Fill sections with analysis JSON results and embed visualizations using markdown image syntax.

Example: Complete Analysis

User request: "Explore this sales_data.csv file"

# 1. Run analysis
python scripts/eda_analyzer.py sales_data.csv -o ./output

# 2. Generate visualizations
python scripts/visualizer.py sales_data.csv -o ./output

# 3. Read results
import json
with open('./output/eda_analysis.json') as f:
    results = json.load(f)

# 4. Build report from assets/report_template.md
# - Fill sections with results
# - Embed images: ![Missing Data](./output/eda_visualizations/missing_data.png)
# - Include insights from results['insights']
# - Add recommendations

Special Cases

Dataset Size Strategy

If < 100 rows: Note sample size limitations, use non-parametric methods

If 100-1M rows: Standard workflow applies

If > 1M rows: Sample first for quick exploration, note sample size in report, recommend distributed computing for full analysis

Data Characteristics

High-dimensional (>50 columns): Focus on key variables first, use correlation analysis to identify groups, consider PCA or feature selection. See references/eda_best_practices.md for guidance.

Time series: Datetime columns auto-detected, temporal visualizations generated automatically. Consider trends, seasonality, patterns.

Imbalanced: Categorical analysis flags imbalances automatically. Report distributions prominently, recommend stratified sampling if needed.

Output Guidelines

Format findings as markdown:

• Use headers, tables, and lists for structure
• Embed visualizations: ![Description](path/to/image.png)
• Include code blocks for suggested transformations
• Highlight key insights

Make reports actionable:

• Provide clear recommendations
• Flag data quality issues requiring attention
• Suggest next steps (modeling, feature engineering, further analysis)
• Tailor communication to user's technical level

Error Handling

Unsupported formats: Request conversion to supported format (CSV, Excel, JSON, Parquet)

Files too large: Recommend sampling or chunked processing

Corrupted data: Report specific errors, suggest cleaning steps, attempt partial analysis

Empty columns: Flag in data quality section, recommend removal or investigation

Resources

Scripts (handle all formats automatically):

• scripts/eda_analyzer.py - Statistical analysis engine
• scripts/visualizer.py - Visualization generator

References (load as needed):

• references/statistical_tests_guide.md - Test interpretation and methodology
• references/eda_best_practices.md - EDA process and best practices

Template:

• assets/report_template.md - Professional report structure

Key Points

• Run both scripts for complete analysis
• Structure reports using the template
• Provide actionable insights, not just statistics
• Use reference guides for detailed interpretations
• Document data quality issues and limitations
• Make clear recommendations for next steps