jackstev/hw1-tabular-worn-items

Worn Items: Waist-Up vs Waist-Down

jackstev/hw1-tabular-worn-items

Thirty hand-measured worn items with five features and one binary target, plus explicitly marked synthetic training variants. Built for CMU 24-679 (Designing and Prototyping AI Systems), Homework 1. The classification task predicts whether an item is worn at or above the waist from its flat dimensions, opening and fastener counts, and colour.

Composition

Column	Role	Description
source_id	identifier	gNN measured, suffixed for synthetic rows.
flat_width_in	continuous	Max horizontal extent laid flat, inches.
flat_height_in	continuous	Max vertical extent laid flat, inches.
openings	count	Through-holes plus pockets.
fasteners	count	Buttons, zips, clasps. Drawstrings excluded.
colour	nominal category	Dominant colour, eight buckets assigned by eye.
waist_up	target	1 = worn at or above the waist, 0 = below.
parent_id	provenance	Source row. Equals source_id when measured.
second_parent_id	provenance	Mixup partner. Equals parent_id otherwise.
mix_weight	provenance	Mixup interpolation weight. 1.0 otherwise.
augmentation	provenance	Method that produced the row, none if measured.
is_augmented	provenance	False for measured rows, True for synthetic.

Provenance fields are not predictors.

Collection

Thirty items were measured by the author from a personal wardrobe with a tape measure, laid flat on a hard surface. Conventions were fixed before collection and applied uniformly:

• Width is the maximum horizontal extent including sleeves. A sleeveless top measures body width; a long-sleeve top measures sleeve-tip to sleeve-tip.
• Openings counts through-holes plus pockets. Belt loops are excluded.
• Fasteners counts buttons, zips and clasps. Drawstrings are excluded.
• Colour was bucketed into a fixed vocabulary rather than recorded freely.

An item column naming each garment ("shorts", "shirt", "sock") was recorded during collection and removed before publication, since it restates the target and would leak the label.

Labels

waist_up is assigned by where an item is worn, not by garment category. Head, torso, arm and hand items are 1; leg and foot items are 0. Gloves and a wrist sweatband are 1 under the arms-and-hands rule, a judgement call worth naming since hands rest near hip height.

The population is worn items, not strictly garments. Goggles, a balaclava, gloves and a shin guard sleeve sit alongside conventional clothing. This widens the feature distributions and makes the task harder than a clothing-only set.

Splits

Split	Measured rows	Synthetic rows	Total rows
train	21	389	410
validation	4	0	4
test	5	0	5

Split	Class	Rows
train	waist-down	197
train	waist-up	213
validation	waist-down	2
validation	waist-up	2
test	waist-down	3
test	waist-up	2

The thirty measured items were partitioned with a two-stage stratified split: 30% reserved as holdout, then 50% of that holdout assigned to test. Both holdouts are entirely unaugmented. The seed was selected so that the dataset's one near-duplicate pair falls in a single partition; no model was fit while choosing it.

Splitting precedes augmentation. Augmenting first would place perturbed copies of validation and test items into training and inflate held-out scores. The notebook asserts that both parents of every synthetic row belong to the training partition.

Augmentation

Synthetic rows were generated from the training partition only, by four labelled methods run over 5 copies each. Rows whose features were unchanged after rounding, and repeated feature combinations from the same parent, were removed before packaging.

Training method	Stored rows
additive_numeric_jitter	105
categorical_perturbation	75
multiplicative_numeric_scale	105
none	21
within_class_mixup	104

• additive_numeric_jitter: Gaussian noise on each numeric feature, scaled to a fraction of the training interquartile range, then returned to its domain.
• multiplicative_numeric_scale: each numeric feature multiplied by a factor drawn uniformly from 0.90 to 1.10, so proportions stay recognisable.
• within_class_mixup: features interpolated between two training rows of the same class, primary parent weighted 0.60 to 0.90.
• categorical_perturbation: colour replaced with another colour observed in the training partition.

Noise scales and the colour vocabulary are estimated from training rows only, so no holdout value influences any synthetic row.

Label preservation is structural rather than argued. One-parent methods copy the parent's label and never recompute it. Mixup draws both parents from the same class, so the blended row inherits that class unchanged; interpolating a binary target across classes would produce a fractional label and break the contract. The notebook asserts that every synthetic label matches both of its parents.

Intended use and limitations

Suitable for teaching, feature-engineering practice and small-model prototyping. Not suitable for any real application.

• Validation and test hold a handful of items each. Accuracy on them moves in very large steps and indicates little. They demonstrate the workflow rather than support model selection.
• All items come from one person's wardrobe, so sizing, style and colour reflect a single adult and do not generalise.
• Synthetic rows carry no information beyond the training originals. They add sampling density, not diversity, and a model evaluated on them would score optimistically.
• Numeric perturbations are clipped at their contracted floors, so rows sitting at a floor drift slightly upward relative to their parents. Labels are unaffected.
• colour is near-uninformative for the target and its rarest buckets have very few measured samples.
• Two measured rows are identical on all four numeric features and differ only in colour. The random seed was chosen so both fall in the same partition, since splitting them would let a model score one for free from the other. The notebook asserts this rather than assuming it.

Ethical considerations

No people, faces or personally identifying information are included. The items are the author's own everyday clothing and accessories. The dataset carries no sensitive attributes and poses no foreseeable privacy risk.

License

CC BY 4.0.

AI usage disclosure

All thirty measurements were taken manually by the author. Measurement conventions, feature selection, label rules and the choice of augmentation strategy were the author's decisions.

An LLM assistant (Claude) was used to write and refine the notebook code, the plots and this card, and to review the measured data for label leakage and class balance. No feature value or label in the measured rows was generated by AI. Synthetic rows are produced algorithmically by the code in the accompanying notebook, not by a language model.