Evolution and Future Trends of Recommendation Systems: From Deep Learning to Large Language Models and AI Agents

The presentation introduces the development history of recommendation systems over the past ten years and highlights emerging trends in the era of large language models (LLMs).

1. Recommendation System Problems and Background

Recommendation systems personalize content by analyzing users' historical behaviors (clicks, views, purchases) to predict interests with minimal interaction cost. Two main product forms are discussed: listwise recommendation (static ranked lists) and conversational recommendation (multi‑turn natural‑language interaction).

2. Development Trend Diagram

A timeline visualizes three technical milestones—deep learning, LLMs, and LLM‑based agents—across the two product forms, showing how research has progressed toward a personalized AI assistant.

3. Deep Learning Era

Key listwise models (DeepFM, DCN, DIN) share a common architecture: sparse categorical features → embedding → feature‑interaction layer → output. Feature‑interaction learning is categorized into multiplication (e.g., FM, DCN), convolution (CNN/GNN), and attention (AutoInt, FiBiNet). User‑behavior modeling evolved from single‑sequence attention (DIN, CAN) to handling ultra‑long sequences (SIM, ETA), multi‑behavior sequences, and multi‑attribute tensors (SC‑CNN).

Limitations of deep models include noisy implicit feedback, lack of semantic understanding, and high interaction cost for conversational systems.

4. Conversational Recommendation

Conversational systems obtain explicit feedback via natural‑language dialogue, reducing ambiguity. Two research streams are identified: item/attribute‑driven dialogue (modeled as an MDP, e.g., Microsoft PMF) and generative dialogue (e.g., Redial’s encoder‑decoder architecture).

5. Large Language Model Era

LLMs bring strong world knowledge and reasoning to recommendation. They can be used as feature enhancers (e.g., KAR) or as core components. KAR combines LLM‑generated textual knowledge with a multi‑expert adapter to produce low‑dimensional embeddings that augment traditional CTR models, achieving a 1.5 % AUC lift and comparable inference latency.

LLM‑based recommendation can be categorized by (a) whether the LLM is fine‑tuned and (b) whether traditional recommenders are still used at inference, forming four quadrants of research.

6. AI‑Agent Era

The ultimate goal is a personalized AI assistant that infers user intent with minimal input. Five essential capabilities are identified: multimodal understanding, planning, personalized memory, tool‑calling, and execution. Both single‑agent and multi‑agent designs (e.g., RAH) are described.

7. Future Development Trends

From browsing to experience: multimodal, immersive interfaces (voice, video, AR/VR) and personal assistants.

From utility to responsibility: fairness, reliability, and explainability.

From understanding to cognition: modeling the underlying user decision logic.

From product to ecosystem: integrating recommendation services across platforms.

The Q&A section addresses practical integration of LLMs with CTR models, performance gains, evaluation metrics, and the current role of LLMs as complementary tools rather than a complete replacement for traditional pipelines.