Tencent Advertising Algorithm Competition Live Stream FAQ: Using Angel Platform, Resource Configuration, and Embedding Evaluation

To help participants prepare for the 2021 Tencent Advertising Algorithm Competition, Tencent held a series of livestreams from May 10 to May 12. On May 12, senior big‑data researcher Sun Ruihong presented an introduction to Angel usage and answered audience questions.

Q1: How do I run a baseline with Angel, and does using Angel eliminate the need for programming?

The answer depends on the algorithm. If Angel already provides the algorithm, you can run it directly without writing code. If the algorithm is available on the TI‑ONE platform, you must check whether it meets your application requirements; if it does, you can use it without programming, configuring the job on the canvas and retrieving results from the specified output path.

Q2: How should Angel resources be configured when using TI‑ONE?

Angel involves two resource components: Spark and Angel Parameter Server (PS). For typical machine‑learning algorithms with modest model size, allocate 2–3 times the estimated data size. For deep‑learning or graph algorithms, consider both data size and model size, estimating parameters and then allocating roughly 2–3 times that combined size. PS and Spark have slightly different calculation methods because PS stores models and data, especially for graph algorithms that involve nodes, edges, and features. Detailed calculations are documented in the official guide.

Q3: How can I evaluate the quality of the embedding vectors obtained from training?

For supervised tasks, use a validation set to verify model performance. For unsupervised models, apply auxiliary measures such as clustering and visualisation to compare results with expectations, often requiring some manual inspection. For recommendation, ranking, or recall scenarios, rely on expert experience and domain‑specific criteria to judge embedding quality.