**2. Overview**

We propose a robust framework called LDA (Long-Term Demand Advisor) to predict long-term (e.g., six months) demand in newly established bike regions. We first extract spatial and temporal features from multi-source open data, then apply our proposed G-clustering algorithm to measure the geographical characteristics and urban correlations in a city. The G-clustering algorithm takes the surrounding locations of the target candidate location into consideration to make a better prediction. Moreover, we extract the urban factors correlated with the long-term demand of sharing bikes, such as POIs (Point of Interests), road structure, and time. On the other hand, features from existing neighbor stations and future stations that have an overlapping operating period are also applied to new bike stations predictions since they will influence the number of demands and transit behaviors.

Our work focuses on long-term prediction, e.g., six months, since the short-term prediction (e.g., one month) is too difficult to predict and not worth studying in practice due to initially unstable environments. Moreover, the long-term effectiveness of stations seems worth investigating to aid in the government's decision and urban planning. For the reasons above, we consider that the predictions of no less than six months are relatively appropriate for urban decision-making. Figure 1 shows our proposed LDA framework, which consists of two major components: data preprocessing and batch prediction.

**Data preprocessing**. We first collect governmen<sup>t</sup> open data and fetch others from Facebook Place API. We also record the latitude and the longitude of all bike stations. Next, we extract spatial features for each station, including nearby station features, seasons, number of POIs and number of check-ins, popular spots, number of intersections, and the length of bike routes based on the parameter *r* of the reachable station region. Finally, the proposed G-clustering algorithm is applied to cluster categories, and all of the extracted features are prepared to be fed into prediction models. Numerical data normalization, data cleaning, and missing data imputation are also applied to all features.

**Batch prediction.** We observe that new stations are sometimes constructed in batches in the real world. For example, the bike station deployment of New York from 2013 to 2017 can be mainly divided into four stages. Each stage contains at least 97 stations to be established in a newly expanded area. After data preprocessing, we split stations into original ones and the others in batches according to their month of establishment. From Batch 1 to Batch *n* (*n* = 3 for the NYC example) predictions, stations established before the corresponding period are set as training sets, and those in the period are testing sets. Finally, a strong prediction model can be applied to finish *n* batches of predictions.

**Figure 1.** The overview of LDA Framework.
