01 — Signals
History becomes a readable representation of preferences.
Views, add-to-cart events and purchases are weighted by strength and recency. They are then connected with product attributes to create affinities across explored categories, brands, ranges or other available catalogue characteristics.
This representation supports score explanation and a fallback when history is insufficient. It also separates repeated preferences from isolated events without exposing identity or individual-level data in the public presentation.
02 — Learning to rank
The model learns to reorder a candidate list.
LambdaMART directly optimizes relative product order over grouped results. Its XGBoost implementation combines affinity features, available context and product signals to assign each candidate a score.
The search engine remains responsible for producing a relevant initial set. The reranker then operates on a bounded list, clarifying each component’s role and allowing personalization to be turned off without losing baseline results.
03 — Protocol
A temporal split moves evaluation closer to the real scenario.
Training uses earlier interactions, then the model must rank a future positive product among negative candidates. The protocol covered tens of thousands of events and several thousand products, using NDCG, Hit Rate and Mean Reciprocal Rank (MRR) as complementary measures.
The comparison isolated two hypotheses. User-product affinities clearly improved all three metrics, while a model based only on broad context remained weak. Integration experiments also exposed excessive latency for some reranking strategies, making it an architecture criterion.
04 — Scope
The next proof belongs to an online experiment.
The evaluation relies on sampled negatives and one observed future outcome. It measures ranking ability within that protocol, but not the causal effect on visitor behavior.
The next step defines eligible segments, fallback behavior, diversity guardrails and an A/B test measuring clicks, add-to-cart events, conversion and latency. The offline model then becomes an instrumented product hypothesis rather than a claim of commercial impact.