TY - GEN
T1 - RLIFE
T2 - 32nd ACM International Conference on Information and Knowledge Management, CIKM 2023
AU - Zhong, Shuxin
AU - Yubeaton, William
AU - Lyu, Wenjun
AU - Wang, Guang
AU - Zhang, Desheng
AU - Yang, Yu
N1 - Publisher Copyright:
© 2023 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-0124-5/23/10.
PY - 2023/10/21
Y1 - 2023/10/21
N2 - Shared electric scooters (e-scooters) have been increasingly popular because of their characteristics of convenience and eco-friendliness. Due to their shared nature and widespread usage, e-scooters usually have a short lifespan (e.g., two to five months [2]), which makes it important to predict the remaining lifespan accurately, ensuring timely replacements. While several studies have focused on the lifespan prediction of various systems, such as batteries and bridges, they present a two-fold drawback. Firstly, they require significant manual labor or additional sensor resources to ascertain the explicit status of the object, rendering them cost-ineffective. Secondly, these studies assume that future usage is similar to historical usage. To solve these limitations, we aim at accurately predicting the remaining lifespan of e-scooters without extra cost, and its essence is to accurately represent its current status and anticipate its future usage. However, it is challenging because: i) lack of explicit rules for the e-scooters' status representation; and ii) e-scooters' future usage may significantly differ from their historical usage. In this paper, we design a framework called RLIFE, whose key insight is modeling user behaviors from trip transactions is of great importance in predicting the Remaining LIFespan of shared E-scooters. Specifically, we introduce an unsupervised contrastive learning component to learn the e-scooters' status representation over time considering degradation, where user preferences are served as a status reflector; We further design an LSTM-based recursive component to dynamically predict uncertain future usage, upon which we fuse the current status and predicted usage of the e-scooter for its remaining lifespan prediction. Extensive experiments are conducted on large-scale, real-world datasets collected from an e-scooter company. It shows that RLIFE improves the baselines by 35.67% and benefits from the learned user preferences and predicted future usage.
AB - Shared electric scooters (e-scooters) have been increasingly popular because of their characteristics of convenience and eco-friendliness. Due to their shared nature and widespread usage, e-scooters usually have a short lifespan (e.g., two to five months [2]), which makes it important to predict the remaining lifespan accurately, ensuring timely replacements. While several studies have focused on the lifespan prediction of various systems, such as batteries and bridges, they present a two-fold drawback. Firstly, they require significant manual labor or additional sensor resources to ascertain the explicit status of the object, rendering them cost-ineffective. Secondly, these studies assume that future usage is similar to historical usage. To solve these limitations, we aim at accurately predicting the remaining lifespan of e-scooters without extra cost, and its essence is to accurately represent its current status and anticipate its future usage. However, it is challenging because: i) lack of explicit rules for the e-scooters' status representation; and ii) e-scooters' future usage may significantly differ from their historical usage. In this paper, we design a framework called RLIFE, whose key insight is modeling user behaviors from trip transactions is of great importance in predicting the Remaining LIFespan of shared E-scooters. Specifically, we introduce an unsupervised contrastive learning component to learn the e-scooters' status representation over time considering degradation, where user preferences are served as a status reflector; We further design an LSTM-based recursive component to dynamically predict uncertain future usage, upon which we fuse the current status and predicted usage of the e-scooter for its remaining lifespan prediction. Extensive experiments are conducted on large-scale, real-world datasets collected from an e-scooter company. It shows that RLIFE improves the baselines by 35.67% and benefits from the learned user preferences and predicted future usage.
KW - Contrastive Learning
KW - Micro-mobility Transportation
KW - Remaining Lifespan Prediction
UR - https://www.scopus.com/pages/publications/85178113690
UR - https://www.scopus.com/pages/publications/85178113690#tab=citedBy
U2 - 10.1145/3583780.3615037
DO - 10.1145/3583780.3615037
M3 - Conference contribution
AN - SCOPUS:85178113690
T3 - International Conference on Information and Knowledge Management, Proceedings
SP - 3544
EP - 3553
BT - CIKM 2023 - Proceedings of the 32nd ACM International Conference on Information and Knowledge Management
PB - Association for Computing Machinery
Y2 - 21 October 2023 through 25 October 2023
ER -