Feature Engineering for Recommendation and Search Advertising

Feature engineering transforms raw data into new vector spaces, enabling models to learn patterns more effectively, improve performance, simplify model complexity, and reduce maintenance costs. High‑quality features can make even simple linear models perform well, emphasizing the "Garbage In, Garbage Out" principle.

Common misconceptions:

Deep learning eliminates the need for feature engineering – while end‑to‑end models can learn some row‑based transformations, column‑based statistical features remain essential, especially for relational data in search, advertising, and recommendation.

AutoFE tools replace manual work – current AutoFE solutions are immature, lack business insight, and cannot fully substitute the creativity and domain knowledge of data scientists.

Feature engineering has no technical depth – it involves statistical reasoning, scaling, binning, encoding, and robust handling of outliers, all of which require solid technical expertise.

What makes good features? They should be discriminative, independent, interpretable, scalable to high‑cardinality, efficient for online inference, flexible across models, and robust to data distribution shifts (e.g., promotional events).

Common transformation operations:

Numerical scaling (Min‑Max, Z‑score, log, L2‑norm, Gauss‑Rank) – essential for distance‑based algorithms and gradient stability.

Handling outliers – robust scaling using median and IQR preserves discriminative power.

Bin (discretization) – converts continuous features into categorical bins, improving non‑linear modeling, interpretability, and outlier resistance.

Category feature techniques – cross‑feature combinations, binning high‑cardinality fields, and statistical encodings (Count, Target, Odds Ratio, Weight of Evidence).

Temporal features – aggregates over recent days/weeks, differences, and sequence statistics.

Examples include quantifying video popularity with log‑based scaling, price “expensiveness” using category‑wise Z‑score, and user preference scoring via binning and statistical encoding.

Feature engineering in search advertising: The workflow follows a "Bin & Counting" paradigm – first bin entities (users, items, contexts), then compute statistical counts (positive/negative samples) per bin, optionally perform cross‑counting, and finally apply scaling/encoding. All statistics must be computed using data prior to the prediction event to avoid leakage.

Typical pipeline steps:

Identify column‑store entities (users, items, contexts).

Perform entity binning based on natural attributes or business logic.

Generate feature crosses (second‑order, third‑order, etc.) on binned features.

Interactive Q&A:

Q1: Manual feature engineering vs. automatic binning in tree models – manual work offers better control for high‑dimensional data, while tree‑based binning can be computationally heavy.

Q2: Acquiring feature‑engineering knowledge – deep business understanding, predictive proxy features (e.g., "shopping age"), and collaboration with experienced teams.

In summary, effective feature engineering bridges data and algorithms, significantly impacting model performance in recommendation and search advertising scenarios.