Turning Real‑Time Hotspot Detection into AI‑Powered E‑Commerce Recommendations

Problem & Background

Traditional recommendation pipelines rely on historical in‑platform behavior and batch processing, which creates a latency gap for rapidly emerging external hotspots (e.g., viral social events, celebrity‑driven trends). The system cannot interpret newly coined terms or memes, leading to missed commercial opportunities and a stale user experience.

Technical Solution

Hotspot Perception Network

An hour‑level data collection service polls top‑ranked hot‑search lists from multiple external platforms. The raw items are normalized and stored in a dynamic “hotspot knowledge base”, providing a real‑time foundation for downstream processing.

Hotspot Understanding Agent

A large‑language‑model (LLM) driven agent performs multi‑round verification to turn short hotspot phrases (e.g., 雷军同款皮衣) into reliable, fact‑checked reports.

Step 1 – Platform‑targeted source: Use the site: operator to restrict the first search to the original platform, capturing the earliest context.

Step 2 – Open‑world verification: Expand site: to the whole web to collect corroborating media.

Step 3‑5 – Expert‑level deep dive: Refine the query with quoted phrases, exclusion operator ( -) and domain‑specific keywords (e.g., 通报, 成交额) to obtain high‑precision evidence.

Demand Inference & Material Recall

The agent extracts up to five core entities (IP, brand, product, location, etc.) from the verified report, infers the most likely purchase motive, and generates executable search queries using a “strong‑anchor + precise term” strategy. Example:

BanG Dream联名T恤
Ave Mujica角色扮演

Three parallel recall streams are then triggered:

Textual recall: Queries are sent to the product text index (title, description) to retrieve matching items.

Visual recall: Seed images from the hotspot are fed to a visual‑search service (e.g., vector‑based image retrieval) to find visually similar products.

Content seeding: Multimodal similarity between newly published content (reviews, tutorials) and the hotspot semantics is computed to surface relevant editorial pieces.

Self‑Evolution of Demand Inference

Each generated query is logged with online performance metrics (CTR, CVR). Queries are labeled “high‑quality” or “low‑quality”. A higher‑order LLM consumes these labeled cases, discovers patterns (e.g., queries containing model numbers convert better) and proposes prompt refinements. The refined prompt is validated offline before being deployed, forming a continuous evaluate → sink → reflect → optimize loop.

Relevance Machine Review

A cascade of three LLM‑based classifiers aligns machine decisions with expert standards:

R1 – Hotspot commercial intent: Filters out politically sensitive, disaster‑related or otherwise non‑commercial hotspots.

R2 – Hotspot‑query relevance: Checks whether the generated query realistically reflects user search behavior for the hotspot.

R3 – Hotspot‑item fit: Verifies that a concrete item can satisfy the inferred user need, ensuring core‑entity alignment.

All classifiers are prompted with strict decision trees and periodically fine‑tuned on human‑reviewed cases. A Retrieval‑Augmented Generation (RAG) store of expert‑validated hotspot‑decision pairs supplies few‑shot examples for each inference.

Automatic Topic Aggregation

To avoid redundant processing of multiple phrases describing the same event, a three‑step clustering pipeline groups related hotspots into a single Topic ID.

Keyword‑based candidate pairing: Tokenize each hotspot, merge extracted core entities into a keyword list, and quickly pair items sharing keywords (e.g., all containing “iPhone 17”).

Event‑level deduplication: An LLM compares each candidate pair, extracts structured elements (time, location, participants, phase) and labels the relationship as same_event, same_macro_event or different based on causal consistency.

Graph‑based topic generation: High‑confidence pairs become edges in a graph; maximal connected components are identified and each component receives a unique Topic ID. Cross‑time partition merging and out‑degree caps are applied to reduce false merges.

Results

The end‑to‑end pipeline reduces manual review workload, delivers hour‑level hotspot response, and produces a curated set of high‑conversion product displays across categories such as fashion, electronics, food and entertainment. Offline experiments show that the cascade classifiers improve precision for R1 by ≈ 50 % and for R2 by ≈ 15 % compared with baseline rules.

Future Work

The roadmap aims to evolve the rule‑driven automation chain into a fully autonomous AI agent that can set business objectives (e.g., maximize GMV), perform multi‑step reasoning across perception, understanding, demand inference and delivery, and continuously self‑optimize via a data‑flywheel of human feedback.