**6. Related Work**

**Impacts of bike-sharing systems**. Many studies analyzed the impact of bike-sharing systems on different aspects of society. The work of [16] mentioned that bike-sharing programs have significantly positive externalities, including the economy, the environment, and health-related externalities. Moreover, introducing bike-sharing systems gives an opportunity to organize public transport interchanges better [17]. Shared bicycles facilitate allow getting to stops and stations for those who do not own a private bike. Additionally, bike-sharing gives more flexibility–shared bicycles users are not burdened with the threat of theft or an obligation to service the bicycle. The study of [18] developed a spatial Agentbased model to simulate the use of bike-sharing services and other transport modes in Taipei city. The simulation results indicate that free use of bike-sharing to connect the transit system can be more sustainable with 1.5 million US dollars in transportation damage cost saved per year and 22 premature deaths further prevented per year due to mode shift to cycling and walking based on the business. The work of [19] demonstrated the importance of user-interface (UI) design, social influence, and new media in affecting users' awareness of and attitude towards uncivilized behaviors, which in turn improve their intention of bike-sharing services use.

The emergence of dockless bike-sharing services has revolutionized bike-sharing markets in recent years. The work of [20] suggested that the dockless design of bikesharing systems significantly improves users' experiences at the end of their bike trips. However, the availability and usage rates of dockless bike-sharing systems imply that they may seriously affect individuals' subjective well-being by influencing their satisfaction with their travel experiences, health, and social participation, which requires further exploration. The work of [21] mentioned that, as Chinese enterprises already invest heavily in Europe, it is crucial for policymakers to introduce rules that would counteract potentially negative consequences of the introduction of a new system of bike-sharing and support positive effects.

**Behavior analysis in bike-sharing systems**. The behavior patterns of users in bikesharing systems are also worth exploring. The estimation results of [22] show that descriptive norm, conformity tendency, and past behavior are important factors that affect both e-bike riders' intention to violate traffic rules and accident proneness. The work of [23] found that perceived ease of use positively influences the attitude towards the systems and the use intention. Therefore, the bike-sharing operating companies should carefully design the usage procedures to make them as simple as possible. The work of [24] adopted machine learning to show that speed, travel distance, and the number of parks and recreational facilities seem to be critical spatial predicting factors of the travel choice in bike-sharing systems. Moreover, considering the impact of COVID-19 on bike-Sharing systems, the work of [25] indicated that usage bike-sharing is more likely to become a more preferable mobility option for people who were previously commuting with private cars as passengers and people who have already registered users in a bike-sharing system. The bike-sharing systems have proved in the study of [26] to be more resilient than the subway system, with a less significant ridership drop and an increase in its trips' average duration.

The work of [27] shows that a high availability rate, a low price, and a large difference in travel time between bike-sharing and other travel modes make potential customers more likely to use a bike-sharing program by modeling a different aspect of travel behavior: heterogeneous time-sensitive customers.

**Bike station deployment**. Research on bike-sharing systems is becoming more and more prevalent worldwide; topics covered range from site selection to rebalancing bike distribution. The works of [28,29] try to figure out the best locations for bike stations from candidate sites. The work of [30] proposes a mixed model to minimize fixed construction costs and variable operational costs. Research combining probability and simulation such as in [31] develops a probabilistic model to infer future demand, and the work of [32] adopts Monte Carlo to predict the over-demand probability in each bike station cluster. On the other hand, the works of [8,33–35] focus on bike imbalance and rebalancing problems, proposing methods to transfer bikes between stations.

**Bike demand analysis and prediction**. In all bike-related problems, the most widely studied is bike demand or traffic flow prediction. The studies of [22,36] have identified the importance of natural environmental factors such as temperature, precipitation, and humidity on cycling activities across different cities. At the feature level, studies [5,37] consider a single factor instead of multiple aspects features and thus may neglect representative elements. Other works collect historical data such as public transportation pattern records [38], crowd flow [39], meteorology data [7,8,40], and so on. Clustering methods applied to bike stations are more and more common in recent works since bike stations share partially similar regional characteristics and will reduce the variance and improve prediction accuracy. The difference between these works is what the cluster is based on. The works of [7,9,32,41] cluster stations according to bike transition pattern records, geographical locations, bike usage, etc. The study of [42] employs SimRank to calculate the similarities between stations and then adopts the density clustering algorithm OPTICS.

However, the works above are not applicable for our scenario since they rely on the historical mobility data and therefore are unavailable for batch prediction in newly established stations in expansion areas. Furthermore, they mostly aim to predict demand in a relatively short period from hourly [11,43], rush hours [9], to weekends and holidays [32], and thus cannot be applied to our long-term prediction.
