Expedia Hotel Search — Learning to Rank

Ranking hotel listings for a search query using XGBoost's LambdaMART implementation, on the Expedia personalised sort dataset. Built as coursework in data mining.

Ranking is not classification with extra steps. A classifier asks "will this hotel be booked?" one row at a time; a ranker asks "given these 30 hotels returned for search #48291, what order should they be in?" The unit of prediction is the whole result list, and a model that scores every hotel plausibly but orders them badly is useless.

How it works

Group structure is the whole thing. The dataset is grouped by srch_id — one search, many candidate properties. The ranker has to be told those boundaries explicitly:

groups_size_train = OrderedDict(...)   # rows per srch_id, in order

model = XGBRanker(
    objective='rank:ndcg',
    learning_rate=.1,
    max_depth=7,
    n_estimators=1500)

model.fit(X_train, y_train, group=list(groups_size_train.values()))

The group argument tells XGBoost which rows compete with each other. Get it wrong — or leave the frame unsorted so the group sizes no longer line up with the rows — and training silently optimises comparisons between hotels from unrelated searches. It won't error; it will just produce a bad model.

rank:ndcg makes the objective position-aware. Normalised Discounted Cumulative Gain discounts relevance logarithmically by rank, which encodes something true about search: users read from the top, so an error at position 2 costs far more than the same error at position 25.

Prediction is a sort, not a label. The model emits a relevance score per row, and the ranking comes from ordering within each search:

df_pred.sort_values(['srch_id', 'score'], ascending=[True, False])

Supporting work, in the other two notebooks:

• PCA.ipynb — dimensionality reduction across the feature set, which spans property attributes, price, location desirability scores and competitor pricing signals.
• Resampling.ipynb — the dataset is severely imbalanced, since almost every hotel shown is not booked. Undersampling the majority class with RandomUnderSampler keeps the bookings from being drowned out entirely.

Results

Not published. The notebooks train the ranker and produce ordered predictions, but no NDCG evaluation was recorded against a held-out split, and the dataset is not committed, so there is nothing here to reproduce a score from. The rank:ndcg above is the training objective, not a measured result — worth saying plainly, because the two are easy to confuse.

Run it

pip install -r requirements.txt
jupyter notebook Ranking.ipynb

Download the training data from the competition page and update the path in the loading cell. Note the ordering requirement above: the frame must be sorted by srch_id before group sizes are computed.

Stack

Python, XGBoost (XGBRanker), scikit-learn, imbalanced-learn, pandas, NumPy, matplotlib, seaborn.