Item2Vec

The core learning algorithm that treats each item as a 'word' and each user session or basket as a 'sentence'. The model is trained to predict which items are likely to co-occur within the same context window. Negative Sampling is used to efficiently approximate the full softmax by updating only a small random subset of negative items, avoiding the prohibitive cost of normalizing over the entire item catalog during training.

Item2Vec produces embeddings where cosine similarity directly corresponds to item co-occurrence probability. Items frequently consumed together—like a phone and its compatible case—are mapped to nearby vectors. This spatial property enables powerful analogical reasoning: the vector arithmetic king - man + woman ≈ queen in NLP translates to item_A - co_purchased_item + substitute_item ≈ alternative_item in retail, capturing complex substitution and complementarity relationships.

Unlike traditional collaborative filtering that relies on a static user-item matrix, Item2Vec operates on ordered sequences of item interactions within a defined time window. This makes it inherently robust to:

• Cold-start users: No user identity is needed; only the current session basket
• Anonymous shoppers: Works without login or persistent cookies
• Shifting intent: Captures short-term, context-specific preferences rather than long-term static profiles

Item sequences are constructed by grouping user interactions into sessions. A sliding context window (typically 5-10 items) slides across each sequence, generating (target, context) pairs. Key design decisions include:

• Session boundary definition: Timeout-based (e.g., 30 min inactivity) vs. fixed-length
• Item deduplication: Whether repeated interactions within a window are treated as distinct
• Temporal weighting: Optionally decaying the influence of items further apart in the sequence

Once trained, item embeddings are indexed using Approximate Nearest Neighbor (ANN) structures like Hierarchical Navigable Small World (HNSW) graphs. At inference time, a user's recent items are averaged or pooled to form a dynamic session embedding, and the ANN index retrieves the top-k most similar items in logarithmic time. This two-stage pipeline—embedding lookup followed by ANN search—enables sub-millisecond retrieval even across catalogs of hundreds of millions of items.

Item2Vec is a direct conceptual descendant of Mikolov et al.'s word2vec, specifically the Skip-Gram architecture. The key insight is that the distributional hypothesis—'words that occur in similar contexts tend to have similar meanings'—applies equally to items in consumption sequences. The primary adaptation is the removal of natural language-specific preprocessing like subsampling of frequent words, as item frequency distributions in retail follow different power-law dynamics than word frequencies in text corpora.