EVENT RECOMMENDATION UNDER EXTREME COLD-START CONDITIONS: THE VALUE OF SCHEDULE, WEATHER AND CALENDAR CONTEXT (BASED ON DATA FROM A KAZAKHSTANI TICKETING PLATFORM)

Abstract

This article addresses the problem of personalized recommendation of cultural and entertainment events under the extreme cold-start conditions typical of online ticketing platforms. The study is based on real anonymized data from a national ticketing platform of Kazakhstan, covering 12,715 interactions of 10,657 users with 2,074 events in four cities of the country from March 2023 to February 2024. The dataset is characterized by a median of one interaction per user, an 83 % share of cold-start users in the holdout month, and about 50 % of purchases falling on premiere events with no sales history. Within a unified protocol of strict temporal validation on six monthly folds, 13 algorithms were compared: from basic popularity models to a hybrid gradient-boosted ranker LightGBM (LambdaRank) enriched with weather, calendar and schedule features and rank blending via reciprocal rank fusion (RRF). Statistical significance was assessed using a paired bootstrap procedure (5,000 replicates, 95 % confidence intervals). It was found that knowledge of the next month’s event schedule provides the largest and statistically significant quality gain (ΔMAP@10 = +0.0084; p < 0.0001), exceeding the effect of any refinement of the ranking algorithm itself; the best configuration achieves MAP@10 = 0.0578, which is 17 % above the strongest non-personalized baseline. The weather context demonstrates a gain at the boundary of statistical significance (p = 0.053), while calendar features alone yield no significant effect. The results substantiate the priority of integrating schedule data into production recommendation pipelines of event platforms.