The developer utilized the Coveo SIGIR eCOM 2021 dataset, consisting of three CSV files containing user actions, search queries, and product attributes. The dataset is characterized by anonymized identifiers and potential gaps in event logging, requiring session reconstruction based on timestamps and session IDs.
For the final evaluation, 60,000 sessions were prepared. The Hit@20 metric was calculated using 748 sessions that contained a recorded purchase following a search query, resulting in 441 hits, or 58.96%. Sessions without a recorded purchase were retained in the sample to maintain the integrity of historical statistics and graph connections.
The model training process involves a 24-hour window for identifying target purchases following a search query. If a purchase occurs within this timeframe, the SKU is added to the target set. The training data for the ranker was derived from 50,000 sessions, where 625 sessions contained a registered target.
The methodology employs implicit feedback, where candidates with registered purchases are prioritized over others. Sessions lacking a registered target are still included in the denominator for the Hit@20 metric to account for potential system misses. Statistical significance remains unconfirmed due to insufficient data for validating confidence intervals.
