Sequential recommendation considers the order and timing of user interactions — modeling how preferences evolve over time and predicting next items based on interaction sequences, capturing temporal dynamics that static models miss.

What Is Sequential Recommendation?

• Definition: Recommend based on ordered sequence of past interactions.
• Input: Time-ordered user history (item1 → item2 → item3 → ...).
• Output: Next items user likely to interact with.
• Goal: Capture temporal patterns, evolving preferences.

Why Sequence Matters?

• Temporal Patterns: Preferences change over time.
• Context: Recent items more relevant than old ones.
• Intent: Sequence reveals user goals (browsing → comparing → buying).
• Seasonality: Holiday shopping, back-to-school, summer trends.

Techniques

Markov Chains: Model item-to-item transitions. RNNs/LSTMs: Learn long-term sequential dependencies. Transformers: Self-attention over interaction history (BERT4Rec, SASRec). Temporal CNNs: Convolutional networks over time. Memory Networks: Explicitly model short and long-term memory.

Applications: E-commerce (next purchase), streaming (next video/song), news (next article), social media (next post).

Challenges: Long sequences, concept drift, computational cost, cold start.

Tools: RecBole, TensorFlow Recommenders, SASRec, BERT4Rec, GRU4Rec.