4.1.1. Baselines

The framework proposed in our work is denoted as **C**ategory **C**lustering applying e**X**treme**G**radient **B**oosting (**CC-XGB**). XGBoost [15] is regarded as one of the most powerful techniques in the public transportation domain.

Regressors such as **RF** (Random Forest), **LR** (Linear Regression), and **SVR** (Support Vector Regression) are used in comparison; **NN** (Neural Network) is also included as a predictor. Moreover, the following compared baselines according to historical average demand are used to verify the performance of our models.

**HA** (**H**istory **A**verage). History rental/drop-off average of stations whose established months are earlier than the predicted station *Si*.

**HSA** (**H**istory **S**imilarity **A**verage). History rental/drop-off average of stations whose established month is earlier and is in the top-five high cosine similarity with the predicted station *Si*.

**HSW** (**H**istory **S**imilarity **W**eight). Let *Si*,<sup>1</sup> ∼ *Si*,<sup>5</sup> be the top-five high cosine similarity stations to the predicted station *Si*.

$$\text{HSW}(S\_i) = \frac{\sum\_{k=1}^{S} (S\_i \cdot \text{rent}) \* CS(S\_{i\prime}, S\_{i,k})}{\sum\_{k=1}^{5} CS(S\_{i\prime}, S\_{i,k})} \tag{2}$$

**HSC** (**H**istory in the **S**ame **C**luster). History rental/drop-off average of stations whose established months are earlier in the same DBSCAN cluster with station *Si*.

**HNN** (**H**istory **N**earest **N**eighbors). History rental/drop-off average of stations whose established month is earlier and distance in the top-*k* nearest with the predicted station *Si*.

## 4.1.2. Evaluation Metric

Since bike demands vary dramatically due to many factors, RMSLE (**R**oot **M**ean **S**quared **<sup>L</sup>**ogarithmic **E**rror) is a more appropriate metric to adopt.

$$\sqrt{\frac{1}{N}\sum\_{i=1}^{N}\left(\log\left(\left(S\_{i'}\operatorname{return}/d\operatorname{drop}\right)-\left(\log\left(\left(S\_{i'}\operatorname{return'}/d\operatorname{drop'}\right)\right)\right)^{2}\right.\tag{3}}\tag{3}$$

*Si*·*rent*/*drop* is the ground truth of demand in six months of *Si*, and *Si*·*rent*/*drop* is the corresponding prediction result of the ground truth.

#### *4.2. Batch Prediction Results*
