Feature Vector

The number of elements in the vector, representing the total count of distinct features. Dimensionality is a critical architectural constraint.

• High-dimensional vectors (hundreds to thousands of features) capture complex patterns but increase computational cost and the risk of overfitting.
• Low-dimensional vectors are computationally efficient but may lack the predictive signal needed for nuanced decisions.
• Dimensionality reduction techniques like Principal Component Analysis (PCA) are often applied to project high-dimensional data into a lower-dimensional space while preserving variance.

All values within a feature vector must be numerical for mathematical operations. Raw data must be transformed through feature engineering.

• Continuous features: Real-valued numbers like price = 49.99 or temperature = 72.5.
• Categorical features: Converted via encoding schemes. One-hot encoding creates binary columns for each category, while label encoding assigns an integer.
• Text and image data: Transformed into dense numerical embeddings using models like BERT or ResNet, which capture semantic meaning in a vector space.

A feature vector is always keyed to a specific entity, such as a user ID, product SKU, or device ID. This association is what links a prediction to a real-world subject.

• The entity serves as the primary join key when assembling features from multiple feature groups in a feature store.
• A complete prediction input often combines features from multiple entities, such as a user vector and an item vector, concatenated for a recommendation model.
• Without a strict entity association, a vector is just a meaningless array of numbers.

A feature vector is a snapshot of an entity's state at a precise moment. Using future data to construct a historical vector causes data leakage and invalidates the model.

• Point-in-time joins in a feature store ensure that when training on a historical event, only feature values that existed before that event are used.
• For example, a vector predicting a purchase on Monday must not include a user_total_spend feature that was updated on Tuesday.
• This temporal discipline is the single most critical factor in building models that perform consistently when deployed.

A vector is sparse when the vast majority of its elements are zero. This is common after one-hot encoding high-cardinality categorical features.

• A vector with 100,000 dimensions might have only 50 non-zero values, representing a user's active product categories.
• Sparse vectors are stored and computed using specialized formats like CSR (Compressed Sparse Row) to avoid wasting memory on zeros.
• Many algorithms, such as linear models and factorization machines, are optimized to handle sparse input efficiently.

The feature vector used for online inference must be an exact logical match to the vector used during offline training. Any discrepancy is a training-serving skew.

• A feature store's feature view defines the transformation logic once and applies it identically in both environments.
• If a training pipeline applies a z-score normalization, the online serving pipeline must use the identical mean and standard deviation calculated from the training set.
• Consistency extends to data types, encoding schemes, and handling of missing values, ensuring the model receives data in the format it learned from.

This vector captures a shopper's immediate intent during a live session. It is served from the Online Store with sub-millisecond latency.

• Inputs: Last 5 SKUs viewed, current cart value, session duration, device type.
• Encoding: Categorical actions are one-hot encoded; numerical values are normalized.
• Use Case: A Next-Best-Action Model uses this to decide whether to show a discount or a product recommendation.

A pre-computed Batch Feature vector summarizing historical loyalty and preferences, materialized daily in the Offline Store.

• Inputs: 90-day purchase frequency, lifetime value tier, average order value, preferred category affinity scores.
• Mechanism: Generated via Point-in-Time Correctness joins to avoid data leakage from future returns.
• Use Case: Injected into a Deep Learning Recommender System to seed the candidate generation layer.

A dense vector representation of a product, often generated by a neural network, stored in a Vector Database for similarity search.

• Inputs: Product title tokens, description semantics, brand, price bucket, image pixels.
• Technique: A multimodal model processes text and images to output a fixed-length embedding.
• Use Case: Finding visually and contextually similar items for a 'Complete the Look' carousel.

A dynamic feature vector representing the decision context for a reinforcement learning agent balancing exploration and exploitation.

• Inputs: Time of day, traffic source (email vs. social), geo-location, inventory level.
• Mechanism: Contextual features are concatenated with user and item vectors to estimate the reward probability.
• Use Case: A Dynamic Pricing Algorithm adjusts a discount multiplier based on this context to maximize margin.

A unified vector that resolves anonymous and logged-in sessions across devices into a single profile using Feature Engineering logic.

• Inputs: Probabilistic matches on IP prefix, deterministic matches on login ID, device graph cluster ID.
• Challenge: Mitigates the Cold Start Problem by merging mobile browsing features with desktop purchase history.
• Use Case: Ensures a user who browsed on a phone sees consistent Dynamic Assortment Optimization on their laptop.

A real-time feature vector computed from Change Data Capture feeds, reflecting current stock levels and velocity.

• Inputs: Current stock count, sell-through rate (last hour), trending velocity score, warehouse proximity.
• Freshness: Requires Feature Freshness of under 1 second to prevent overselling.
• Use Case: A Demand Forecasting Model consumes this to trigger a 'Low Stock' urgency badge on the product display page.