Amazon Product Search & Review Summarization
I co-developed an end-to-end search and review-summarization system for Amazon Office Products. The project connected data preparation, information retrieval, evaluation, and a lightweight Flask interface in one reproducible workflow.
At a glance
| Question | Can product metadata and customer reviews support fast, relevant search and concise review summaries? |
| Data | 300,000 customer reviews; 104,259 indexed products |
| Methods | TF-IDF, cosine similarity, extractive summarization, offline retrieval evaluation |
| Tools | Python, pandas, scikit-learn, SciPy, Flask |
| Role | Team project — co-developed the analysis and application workflow |
Approach
- Prepared the corpus. Product metadata and a controlled review sample were parsed, cleaned, and joined into a searchable product table.
- Built the retrieval layer. Product titles and descriptions were represented with TF-IDF. The final sparse search matrix covered 104,259 products across a 50,000-term vocabulary, with cosine similarity used for ranking.
- Added review summaries. The application selected informative review sentences and included safeguards for products with limited review coverage.
- Created a search interface. A Flask dashboard returned ranked products, metadata, and review summaries for a user query.
- Evaluated relevance. Product-title self-retrieval was tested on 200 queries using Hit@10, MRR@10, and nDCG@10 rather than relying only on anecdotal examples.
Results
| Metric | Result |
|---|---|
| Hit@10 | 98.0% |
| MRR@10 | 92.45% |
| nDCG@10 | 93.85% |
| Queries ranked first | 177 of 200 |
The strongest result was coverage: 196 of 200 title queries retrieved the expected product within the top ten. MRR@10 and nDCG@10 also show that the relevant item usually appeared near the top rather than merely somewhere in the result set.
Practical implications and limitations
This project demonstrates how a simple, interpretable retrieval baseline can perform strongly when paired with careful corpus construction and explicit evaluation. The title-based evaluation is intentionally narrow, however; a production system would also require human relevance judgments for natural-language queries, latency testing, and monitoring for catalog drift.
Review the work
Attribution: Team project. Results and methods are described as shared project work; no claim of sole ownership is made.