NEWS

detectXOR 0.2.0

The chi-square comparison of observed and expected tile counts and the fixed absolute-difference threshold (abs_diff_threshold) were replaced. A pair is now tested with the XOR purity: the balanced accuracy of the XOR rule within the four critical tiles, maximized over both label assignments and both XOR manifestations (default, rotated). A pair is called XOR-positive if
- the purity exceeds chance in a class-label permutation test, adjusted for the number of variable pairs (Westfall-Young max-T by default, p_adjust_method, n_perm);
- the one-sided lower confidence bound of the purity reaches purity_threshold (default 0.8);
- each class populates both of its expected tiles (min_tile_share, default 0.05);
- the class-wise Kendall tau criterion is met (check_tau, as before);
- optionally, all four group-wise Wilcoxon tests are significant (compute_axes_parallel_significance, now FALSE by default). Previously these tests were computed but did not affect the decision.
abs_diff_threshold is deprecated and ignored.
The results table has new columns (xor_purity, purity_lower, coverage_ok, p_value, p_adjusted, candidate_pattern, axis_tests_passed); chi_sq and chi_sq_p_value were removed.

Detection no longer depends on which class label is coded first.
Class columns not named "class" no longer cause an error.
extreme_handling = "remove" is accepted as a synonym of "trim", and trimming no longer considers the class column.
Tile assignment for the Wilcoxon tests now uses the same grid as the main analysis; pairs of binary variables are exempt from these tests.
The sequential fallback of the parallel processing wrapper was repaired.
Winsorization handles missing values when use_complete = FALSE.

New argument split_quantiles to change the tile boundaries (default: tertiles).
New argument balance_classes to down-sample the larger class before screening, recommended for marked class imbalance.
Unit tests were added (testthat).
Examples use XOR_data instead of the three-class iris data.