<![CDATA[Purpose – The user cold-start problem cripples recommenders when a new user has almost no interaction history. This study asks two questions prior work conflates: does semantic review embedding similarity outperform conventional rating-overlap collaborative filtering (CF) in cold-start, and does a magnitude-sensitive kernel over the embedding add anything beyond ordinary cosine similarity? Design/methods/approach – Under one identical protocol on two Amazon review domains of differing sparsity, RBF-SBERT (an RBF kernel over SBERT embeddings) is compared with cosine, Pearson, an XGBoost baseline, rating-overlap CF, and two popularity baselines. Cold-start users (one to three interactions) use a per-user sampled protocol; significance uses Wilcoxon and Cohen's d. Findings – In the sparse domain the embedding beats even the strongest simple baseline, item-count popularity, on precision (P@5 0.567 vs 0.252; d=0.65) and ranking (AUC 0.869 vs 0.821); in the denser domain that edge vanishes as popularity matches AUC and exceeds precision, so the advantage is governed by sparsity. Rating-overlap CF is weakest, collapsing to its mean-rating fallback. The RBF kernel is inert (≈ cosine, d=0.007–0.107) because the pooled magnitude it uses is nearly constant (CV 0.07–0.09) and tracks review length, not preference. Research implications/limitations – Cold-start gains arise from the representation, not the kernel; methods must be benchmarked against popularity under sparsity-aware protocols. Evaluation covers two Amazon categories under one sampled protocol. Originality/value – The study separates representation from kernel and precision from ranking quality, giving a reproducible account of when review-based embedding helps cold-start and why.]]>
