wenmin-wu/tabular-anonymized-target-reverse-engineering

Overview

When a competition host hides the target ("loyalty score, normalized") your first move is not to build a model — it's to figure out what the number actually means. Histogram it; try 2**y, np.exp(y), y * std + mean for various candidate scales. If a transform makes the distribution snap onto integer values, dollar amounts, or a known KPI range, you have just unlocked feature engineering the host did not intend to share. On Elo, applying 2**target revealed an integer-like structure consistent with a churn-month count; the discovery motivated month-recency features that pushed solutions up the LB. This is dual-use: also use it to detect leakage (if the transform recovers a column you already have).

Quick Start

import numpy as np, matplotlib.pyplot as plt

y = train.target.values

# 1) raw histogram
plt.hist(y, bins=200); plt.title('raw'); plt.show()

# 2) try common inverse transforms
for name, t in [('2**y', 2**y), ('exp(y)', np.exp(y)),
                ('y**2', y**2), ('1/(1-y)', 1/(1-y))]:
    vals = t[np.isfinite(t)]
    plt.hist(vals, bins=200); plt.title(name); plt.show()
    print(name, 'min=', vals.min(), 'max=', vals.max(),
          'frac_int=', np.mean(np.isclose(vals, np.round(vals), atol=1e-3)))

Workflow

1. Plot the raw target — note any spikes, gaps, hard floors/ceilings (sentinel values are clues)
2. Try the standard transform menu: exp, log, 2**, **2, affine rescales using train.describe() percentiles as anchors
3. For each transform compute: range, fraction-near-integer, mode, top-10 most-frequent values
4. If a transform yields integer-like or round-number structure, search the host's data for a column whose values match
5. Cross-check on a small sample by hand — if 2**y == months_since_last_purchase for 5 random cards, you've recovered the target
6. Use the discovered semantics to design features (e.g., monthly aggregates if the target is monthly)

Key Decisions

• Spikes are tells, not noise: a single huge spike usually points to a sentinel value or a clipped boundary; both reveal target semantics.
• Try inverses of the obvious normalizations: hosts most often use z-score, min-max, or log — invert each.
• Stop when you have a hypothesis, not a proof: even a plausible meaning is enough to seed feature engineering; certainty isn't required.
• Don't post the discovery: if it's exploitable you want to keep it, if it's leakage report it to the host.
• Combine with adversarial validation: if your transform recovers a feature train and test disagree on, that's a leak hint.
• Beware of overfitting to the discovery: validate that features motivated by the recovered semantics actually generalize.

References

• Target — true meaning revealed (Elo Merchant)