Search Term Recommendation: Scenarios, Algorithm Design, and Future Directions

The talk introduces the search term recommendation problem, describing five main scenarios in QQ Browser: personalized term recommendation, query auto‑completion, short‑video‑related queries, article‑related queries, and result‑page related queries, highlighting the three‑layer matching relationship between context, query, and result.

Three key challenges are identified: lack of evaluability in traditional recommendation, the absence of a stable content ecosystem for query items, and the rapidly changing attributes of query results, which increase technical difficulty.

The query library is organized into four categories—active search terms, generative queries (template‑based and extraction‑based), knowledge‑graph generated queries, and manually curated hot‑search terms—supported by machine‑learning operators, safety checks, and human review.

Result‑page satisfaction is evaluated on relevance, richness, timeliness, and content quality, forming the basis for downstream ranking.

Algorithm design follows a classic recommendation stack: query library → index → recall → coarse ranking → fine ranking → mixed ranking → business integration. Specific issues such as data sparsity in coarse ranking are addressed with dual‑tower models, teacher‑student training, and embedding compression via hash‑based partitioning.

Fine‑ranking employs multi‑task learning (ESMM, MMoE) to jointly predict click‑through rate, result‑page click‑through, consumption, and relevance, using both pointwise and pairwise losses, as well as lambda‑loss for position‑aware ranking.

In the query auto‑completion scenario, top‑N satisfaction is improved by combining pointwise estimation with pairwise loss and lambda‑loss, and evaluated using position‑wise click share, click‑through rate, and NDCG.

Future work focuses on multi‑task and session‑aware models to capture diverse user behaviors across QQ Browser’s integrated services, and on enhancing query understanding to filter high‑quality terms from billions of candidates.

The Q&A section addresses embedding training, timeliness updates, and the proportion of different query generation methods, emphasizing the dominance of massive active search queries.