Corruption-Aware Multi-View Metric Learning for Unseen-User Session Retrieval

Source abstract

Retrieving related visits from users unseen during training requires representations that preserve behavior across sessions. We propose a corruption-aware sequence--graph metric-learning framework that combines a gated recurrent unit with a directed, weighted session-graph encoder. Cross-session contrastive learning uses independently recorded visits of a training user as positives, while clean and perturbed inputs train complementary views in a shared normalized embedding space. An in-depth Coffee Shop study uses five user-disjoint splits and three initializations, with matched supervision, independent-encoder and training-regime controls. A matched positive-pair experiment increases clean mean reciprocal rank (MRR) from 0.1285 to 0.5948 at 1,000 candidates. Under 50\% combined view corruption, original fusion achieves MRR 0.1846, compared with 0.1638 for independent late fusion and 0.0953 for validation-tuned SASRec. Paired gains are 0.0208 (95\% user-bootstrap interval {[}0.0190, 0.0227{]}) and 0.0894 {[}0.0866, 0.0921{]}, respectively. Joint training therefore adds value under combined view damage, complementing alternatives that perform better on clean retrieval. Count-based and raw-event controls further characterize the operating conditions of this benefit. Frozen same-application repartitioning further supports the gated extension's graph-damage benefit against SR-GNN. The contribution is an integrated retrieval-oriented learning design and controlled evidence explaining how cross-session supervision and the availability of complementary views shape its benefits.