TaxiHailer: A Situation-Specific Taxi Pick-Up Points Recommendation System(阅读笔记)20180114

《TaxiHailer: A Situation-Specific Taxi Pick-Up Points Recommendation System》2014 DASFAA

Abstract:

本文为寻找出租车的乘客构建了一个根据情景定制的推荐系统，根据给定的起点，终点和时间需求信息，该系统能够推荐一系列带有距离信息的上车点，这些上车点根据潜在的等待时间排序，从而设定优先级。
本文不同于现有的工作在于，本文考虑了三组特征来构建回归模型，同时也构建了泊松过程模型，用于道路分割聚类。
本文的亮点在于，在不同的道路分割聚类区域上的上车点，系统会评估和选择最合适该情景的推荐模型。
最后本文提出了一个筛选推荐点的小技巧，将推荐点加上驾驶方向信息，如果出租车驾驶方向和该方向相反，将这些点剔除。

MainWork:

1、系统架构

OFFLINE:

The offline modules re-build the waiting time prediction models using recent data periodically.

• filter and persist trajectories into repository（preparing）
• calculate the statistical features of road segments（MapReduce）（addtion feature :weather conditions.etc.）
• road segments are clustered into hundreds of groups to reflect different traffic situations（ road segments cluster）
• clustered historical pick-up points on segments with frequency and distance
  rules（pick-up points cluster）
• Then, one point in each cluster is chosen as potential pick-up point for recommendation（choose point in road segment cluster）
• build regression models and Poisson process models or each road segment cluster under different situations（build situation-specific models ）
• choose the best models for each cluster（choose the best one）

ONLINE:

In the online part, TaxiHailer processes queries and recommends pick-up points within the specified distance

• the origin of user query is mapped to the road network（mapping）
• select the prediction model according to the query context（select best model）
• a collection of result candidates can be fetched by distance limitation（calculation）
• use the route towards the destination to prune the candidate set（filter）
• re-rank the pruned candidates and return them with predicted waiting time（re-rank）

2、Situation-Specific Model Building

动机：

The waiting time for the same road segment may differ in different situations, such as weather, time period.
Also, social activities can cause traffic jams, leading to a longer waiting time there.

方法：

first train set of models for each road segment group.
then, the most proper waiting time prediction model for different situations can be selected by periodically evaluation.
generate sufficient test sets which are sampled on the roads and calculated from the historical trajectory data.

3、推荐点的筛选

动机：

the volumes of taxis driving on opposite side of the same road are quite different in most cases. It is undesirable to get recommended taxis that drive in the other side.

方法：

As we generate potential pick-up points including their directions, destination awareness is included in our query algorithm.

形象化表示：

如图所示，汽车行驶方向为右行，由于该道路分为左行和右行，而且汽车不便于变道，如果在改路段上有如图所示的潜在上车点（红色），那么分布在左行车道上的上车点需要被剔除（绿色圈代表剔除），最后将剩余的点进行排序。

思考：

1、基于场景的策略选择给我一个启示，我们在构建模型的时候应该考虑更多的特征，如果特征考虑的不够全面或者尚未发掘潜在的特征，可以考虑多策略模型。

2、本文提到的推荐点的筛选是很有用的一个技巧，能够合理化推荐结果，我们可以借鉴该方法。