02 — Feature Types in Recommender Systems

“Garbage in, garbage out — but in RecSys, the right features are 80% of the lift.”

Introduction

Features are the language your model uses to describe users, items, and context. Get them right and even a logistic regression can outperform a transformer. Get them wrong and no architecture will save you. This post is a taxonomy of the feature families used in modern recommenders, the engineering tricks each requires, and the gotchas they hide.

1. The Feature Taxonomy

flowchart TD
    F[Recommender Features] --> U[User Features]
    F --> I[Item Features]
    F --> C[Context Features]
    F --> X[Cross / Interaction Features]
    F --> S[Sequence Features]

    U --> U1[Demographic]
    U --> U2[Behavioral / Historical]
    U --> U3[Profile / Preference]
    U --> U4[Social / Graph]

    I --> I1[Content / Metadata]
    I --> I2[Visual / Audio Embeddings]
    I --> I3[Text / NLP Embeddings]
    I --> I4[Statistical / Popularity]

    C --> C1[Temporal]
    C --> C2[Location / Device]
    C --> C3[Session]

    X --> X1[User x Item history]
    X --> X2[User x Category affinity]

    S --> S1[Last-N clicks]
    S --> S2[Time-decayed actions]

2. User Features

2.1 Demographic

Age, gender, country, language. Pros: stable, cheap. Cons: privacy-sensitive (GDPR/CCPA), sparse, often missing or self-reported.

Production note: demographic features often add little lift once behavioral features are present — but they’re crucial for cold start (Blog 18).

2.2 Behavioral / Historical

Counts and aggregates over user actions: clicks, purchases, watch time, ratings, skips.

These are typically computed in a feature store (Blog 22) with online/offline parity guaranteed by point-in-time joins.

2.3 Profile / Preference

Explicit signals — followed creators, saved searches, “thumbs up”, muted categories.

2.4 Social / Graph

Friends, follows, co-views. Powers graph models (Blog 10). Features: number of friends who liked item, friend-of-friend signals, embedding from node2vec on the social graph.

3. Item Features

3.1 Content / Metadata

Title, brand, category, price, tags. Categorical features are typically hashed + embedded.

3.2 Text Embeddings

Title and description → embeddings via BERT, sentence-transformers, or fastText.

3.3 Visual / Audio

Image embeddings (ResNet, CLIP) for fashion/visual products. Audio embeddings (VGGish, MERT) for music.

3.4 Statistical / Popularity

CTR, CVR, average rating, click count over various time windows. Beware of leakage — must be computed as of the prediction time.

4. Context Features

import pandas as pd
df["hour"] = pd.to_datetime(df["ts"], unit="s").dt.hour
df["is_weekend"] = pd.to_datetime(df["ts"], unit="s").dt.dayofweek >= 5

5. Cross & Interaction Features

The deep learning revolution started by automating cross features (Wide & Deep, DCN — Blog 06/07). But hand-crafted crosses still beat neural ones in many shops:

• user_country x item_category
• device x ad_creative_size
• query_intent x item_brand

In TensorFlow:

crossed = tf.feature_column.crossed_column(
    ["user_country", "item_category"], hash_bucket_size=10000)

6. Sequence Features

Users are sequences, not bag-of-actions. Modern systems feed last-N item IDs + dwell times into transformer encoders (Blog 08).

# 50 most recent item ids, padded with 0
seq = user_history[-50:]
seq = [0] * (50 - len(seq)) + seq

7. Engineering Tricks

7.1 Categorical Encoding

• One-hot: only for low-cardinality
• Hashing trick: fast, no vocab maintenance, allows collisions on purpose
• Embeddings: $d \approx \min(50, \lceil n^{0.25} \rceil)$
• Target encoding: replace category with smoothed mean target (careful with leakage — use out-of-fold)

7.2 Numerical

• Log-transform heavy-tailed features (revenue, view-count)
• Bucketize then embed for non-monotone effects
• Standardize for neural nets

7.3 Missing Values

• Use explicit “missing” embedding, not zero — zero collides with informative inputs.

7.4 Freshness

Cache hot features in Redis; tier the rest in offline parquet + DynamoDB.

8. Feature Importance & Selection

# Permutation importance with LightGBM
import lightgbm as lgb
from sklearn.inspection import permutation_importance

model = lgb.LGBMClassifier(n_estimators=200)
model.fit(X_train, y_train)
r = permutation_importance(model, X_val, y_val, n_repeats=5, random_state=0)
imp = pd.Series(r.importances_mean, index=X_val.columns).sort_values(ascending=False)
print(imp.head(20))

For deep models, use integrated gradients or SHAP sparingly — they’re expensive at scale.

9. End-to-End Code: Build a Feature Table on RetailRocket

# pip install pandas pyarrow
import pandas as pd

# RetailRocket events: timestamp, visitorid, event, itemid, transactionid
events = pd.read_csv("events.csv")
events["ts"] = pd.to_datetime(events["timestamp"], unit="ms")

def user_features(events, as_of):
    past = events[events["ts"] < as_of]
    g = past.groupby("visitorid")
    return pd.DataFrame({
        "n_views_7d": g["event"].apply(lambda x: (x == "view").sum()),
        "n_carts_7d": g["event"].apply(lambda x: (x == "addtocart").sum()),
        "n_purch_7d": g["event"].apply(lambda x: (x == "transaction").sum()),
        "last_seen_hr": (as_of - g["ts"].max()).dt.total_seconds() / 3600,
    }).reset_index()

def item_features(events, as_of):
    past = events[events["ts"] < as_of]
    g = past.groupby("itemid")
    return pd.DataFrame({
        "item_ctr_30d": g["event"].apply(lambda x: (x == "addtocart").sum() / max(1, len(x))),
        "item_popularity": g.size(),
    }).reset_index()

snapshot = pd.Timestamp("2015-09-01")
u_feat = user_features(events, snapshot)
i_feat = item_features(events, snapshot)
print(u_feat.head(), i_feat.head())

This is essentially what a feature store does at scale, with point-in-time correctness guarantees.

10. Feature Engineering Mistakes (and Fixes)

11. Public Datasets

• MovieLens — has tag genome (rich item features) — https://grouplens.org/datasets/movielens/
• Amazon Reviews 2018 — title, brand, image, price — https://nijianmo.github.io/amazon/
• H&M Personalized Fashion — demographic + images — https://www.kaggle.com/c/h-and-m-personalized-fashion-recommendations
• Criteo Display Advertising — 26 categorical + 13 numerical anonymized features — https://www.kaggle.com/c/criteo-display-ad-challenge/data
• Avazu CTR — 24 features for mobile ads — https://www.kaggle.com/c/avazu-ctr-prediction

12. Further Reading

• Cheng et al., Wide & Deep Learning for Recommender Systems (DLRS 2016)
• He & Chua, NFM: Neural Factorization Machines (SIGIR 2017)
• Covington et al., Deep Neural Networks for YouTube Recommendations (RecSys 2016) — masterclass in features
• Tecton & Feast docs on feature stores (Blog 22)