results-analysis

Load your result files (CSV, JSON, or TensorBoard logs), validate data completeness, then compute means with standard deviations across runs. Run pre-tests for normality and variance homogeneity, apply the appropriate significance test, and report p-values with effect sizes before drafting the Results section.

For two paired models with normal data use a paired t-test; for non-normal data use Wilcoxon signed-rank. For three or more models use ANOVA with Tukey HSD post-hoc, or Kruskal-Wallis with Dunn's test when normality fails. Always verify assumptions first with Shapiro-Wilk and Levene tests.

The workflow supports CSV files for tabular metrics, JSON files for structured results, TensorBoard logs for training curves, and Python pickle files for complex objects. Data validation checks completeness, consistency, and reproducibility information such as random seeds.

Run experiments at least 3-5 times with different fixed random seeds; use 10 or more runs for high-variance tasks. Report all runs as mean plus or minus standard deviation, and never cherry-pick the best single run.

Raster formats like PNG and JPG lose quality when scaled and often fail publisher requirements. Use vector formats (PDF or EPS) instead, with colorblind-friendly palettes such as Okabe-Ito, error bars, and fonts of at least 8 points so figures remain readable in black-and-white print.

Standard deviation describes the spread of your data across runs, while standard error describes uncertainty in the estimated mean and shrinks with sample size. Always state explicitly which one you report, since standard error appears smaller and can mislead readers if unlabeled.