"algo-rec-session"

Session-Based Recommendation

Overview

Session-based recommendation predicts the next item a user will interact with based on their current session's click/view sequence, without relying on long-term user profiles. Uses Markov chains, association rules, or neural approaches (GRU4Rec). Operates in real-time with O(sequence_length) inference.

When to Use

Trigger conditions:

• Anonymous users (no login, no long-term profile)
• Short browsing sessions where recency matters most
• Real-time "next item" prediction during active sessions

When NOT to use:

• When rich user history is available (use CF or content-based for better personalization)
• When sessions are extremely short (1-2 clicks) — insufficient signal

Algorithm

IRON LAW: First Few Clicks Are Disproportionately Important
Session-based methods operate WITHOUT long-term profiles. Intent must
be inferred from SHORT sequences. The first 2-3 clicks establish the
session's intent — misreading early signals derails the entire session.

Phase 1: Input Validation

Parse clickstream into sessions (by session ID or timeout-based splitting, typically 30min inactivity). Filter sessions below minimum length (3+ events). Gate: Sessions parsed, minimum length threshold applied.

Phase 2: Core Algorithm

Markov Chain approach:

1. Build transition matrix from item-to-item sequences across all sessions
2. For current session [A, B, C], predict next item from P(next | C) or higher-order P(next | B, C)

Association Rules approach:

1. Mine frequent item sequences (sequential pattern mining)
2. Match current session suffix against known patterns
3. Recommend items that frequently follow the matched pattern

Phase 3: Verification

Evaluate with leave-one-out: hide last item in each session, predict, check hit rate and MRR (Mean Reciprocal Rank). Gate: Hit@20 significantly above random baseline.

Phase 4: Output

Return ranked next-item predictions with confidence scores.

Output Format