How Suning Leverages Query Logs to Auto‑Discover Product Synonyms

In the e‑commerce domain, users express the same product with varied terms, leading to recall gaps in search. Suning’s search R&D team mined massive product and query‑click logs to automatically discover domain‑specific synonyms, improving retrieval and data management.

Data Source and Candidate Extraction

The system relies on query‑click logs, which capture diverse user‑entered keywords. After cleaning noise, candidate terms of length 2‑4 characters are extracted from queries and filtered by frequency thresholds.

Computing Synonym Relations

Three similarity features are used:

Similarity of clicked product vectors.

Similarity of surrounding character (context) vectors.

Similarity of occurrence patterns.

Each feature yields a cosine‑based similarity score; the final score is a weighted sum.

Click‑Vector Similarity

Product click vectors are built from user clicks and compared using cosine distance.

Context‑Vector Similarity

For each candidate term, a character count vector is created from the five characters before and after each occurrence in queries; cosine similarity measures contextual closeness.

Pattern‑Based Similarity

Query terms are tokenized, sorted, and grouped by product ID and term count. Apriori frequent itemset mining extracts K‑1 itemsets; patterns are matched to find interchangeable terms, incrementing similarity counts.

Post‑Processing

After filtering by similarity thresholds, further cleaning removes incomplete terms and overlapping pairs to reduce noise.

Through this multi‑feature approach combining cosine similarity and Apriori mining on large‑scale query and click data, Suning generated tens of thousands of high‑quality synonym pairs, far exceeding manual knowledge and significantly enhancing search experience.