Sounds like gambling: detection of gambling venue visitation from sounds in gamblers' environments using a transformer.

Objective digital measurement of gamblers visiting gambling venues is conducted using cashless cards and facial recognition systems, but these methods are confined within a single gambling venue. Hence, we propose an objective digital measurement method using a transformer, a state-of-the-art machine learning approach, to detect total gambling venue visitations for gamblers who visit multiple gambling venues using sounds in gamblers' environments. We sampled gambling and nongambling event datasets from websites to create a gambling play classifier. We also sampled gambling and nongambling location datasets for a gambling location detector. Further, we sampled practical dataset with four different recording conditions and two different recording devices. Our Swin transformer model with 54 classes (4 gambling play classes and 50 nongambling event classes) achieved highest accuracy (0.801). The gambling location detector of the Swin transformer also achieved high performance; the areas under the receiver operating characteristic curves (AUCs) for bingo, mahjong, pachinko, and electronic gambling machine plays were 0.845, 0.780, 0.826, and 0.833, respectively. Moreover, gambling visitation detector of the Swin transformer showed high performance especially in Pachinko (AUCs 0.972-0.715) regardless of their recording conditions and devices. These preliminary findings highlight the potential of environmental sounds to detect visits to gambling venues.