New Generation Rank Technology: Search‑Based Interest Model (SIM) and Dynamic Computation Allocation Framework (DCAF) for Alibaba Directed Advertising

The talk introduces the background of Alibaba's directed advertising ecosystem, distinguishing between banner‑style non‑product ads and product‑centric ads, and explains why accurate ranking of top‑k items is crucial for revenue maximization.

It then outlines the evolution of ranking algorithms from early MLR models to DIN, DIEN, MIMN, and finally the Search‑Based Interest Model (SIM), which expands user behavior modeling from hundreds to tens of thousands of actions, enabling lifelong interest capture.

SIM adopts a two‑stage retrieval process: a lightweight General Search (soft or hard) reduces the massive behavior history to a few hundred relevant actions, followed by an Exact Search that applies fine‑grained CTR models such as DIN/DIEN on this subset, achieving better long‑term interest modeling without prohibitive online cost.

The article also discusses the limitations of memory‑based approaches like MIMN, including scalability, update latency, and memory size constraints, motivating the shift to SIM.

Beyond modeling, the presentation introduces the Dynamic Computation Allocation Framework (DCAF), a combinatorial optimization that assigns personalized compute budgets to each request based on its expected revenue contribution, allowing the system to allocate more resources to high‑value traffic while conserving compute on low‑value traffic.

Experimental results show that DCAF can increase eCPM by up to 3.7% under the same compute budget or reduce compute usage by nearly 50% without sacrificing revenue, demonstrating the practical impact of joint algorithm‑architecture‑compute co‑design.

Finally, the speaker shares insights on the importance of algorithm‑engineering co‑design, the need for lifelong user modeling, and encourages collaboration between research and production teams to drive continuous innovation in large‑scale advertising systems.