<![CDATA[The growth of digital platforms has generated large volumes of Web-derived interaction data, but these data are often noisy, duplicated, incomplete, and temporally unstable. Recommendation quality therefore depends not only on the ranking model, but also on how extraction, validation, and temporal control are integrated upstream. This paper presents an end-to-end architecture in which Web extraction, schema normalization, cleaning, deduplication, anomaly quarantine, recency-aware processing, and recommendation generation are treated as a single operational pipeline. The contribution is not the use of hybrid recommendation alone, which is already common, but the explicit integration of these quality-control stages with temporally valid offline evaluation and system-level monitoring. Four recommendation strategies are studied within the same pipeline: global popularity, recency-weighted popularity, implicit matrix factorization, and a hybrid method that combines collaborative filtering with a recency-based fallback for sparse-user cold-start situations. Experiments are conducted on a realistic e-commerce dataset comprising approximately 50,000 users, 18,000 items, and 1.2 million interactions under a strict chronological 80/20 split. Evaluation includes Precision@K, Recall@K, NDCG@K, Coverage@K, sparse-user cold-start analysis, and system indicators. Results indicate that the hybrid approach achieves the best observed aggregate ranking performance under the present protocol, improves sparse-user robustness (Recall@10 = 0.158), maintains broad catalog coverage (38.9%), and remains operationally stable under the tested evaluation conditions (p95 latency = 48 ms; uptime = 99.7%). These findings support assessing recommendation quality as a property of the full data-to-recommendation pipeline rather than of the ranking algorithm alone.]]>
