论文阅读——Where to Go Next: A Spatio-Temporal Gated Network for Next POI Recommendation

00 文章基本信息

01 摘要

当前RNNs在处理neighbor check-ins时存在的弊端：

• rarely consider the spatio-temporal intervals between neighbor
  check-ins, which are essential for modeling user check-in behaviors
  in next POI recommendation.
• 换言之：现有的基于RNN的方法在对用户短期偏好建模时，要么忽略了用户的长期偏好，要么忽略了最近访问的poi之间的地理关系，使得我们的推荐结果并不可靠。

our contributions：

• we propose a new Spatio-Temporal Gated Network (STGN) by
  enhancing long-short term memory network;针对上述局限性，我们提出了一种新的长短期偏好建模方法（LSTPM）。
• two pairs of time gate and distance gate are designed to control the
  short-term interest and the long-term interest updates, respectively;
• we introduce coupled input and forget gates to reduce the number of
  parameters and further improve efficiency;
• we evaluate the proposed model using four real-world datasets from
  various location-based social networks.

该论文的核心贡献：
为什么要加入时间门和距离门，原因就在于最近一次访问的POI并不一定就具有很强的说服力，比如，最近一次的访问距离现在时间很长（用户已经改变了以前的行为习惯）或者最近一次访问的距离太长了（比如我家海南，去黑龙江玩了一次，现在再给我推荐黑龙江的POI就不合适了）。
改进之处就在于考虑最近一次访问的POI的时间和距离间隔，如下图：

1 Introduction

以往的工作：
以往的工作大都是用对待sequential data analysis and recommendation的方式来处理next POI recommandation问题，但是这种处理方法存在一个很大的弊端：

• 它们没有考虑time intervals and geographical distances between neighbor
  items，这也是next POI recommandation与other sequential tasks的不同之处所在。

因此，本文提出了a new Spatio-Temporal Gated Network by enhancing long short term memory, named STGN, to model users’ sequential visiting behaviors，该方法的优点如下：

• Time intervals and distance intervals of neighbor check-ins are
  modeled by time gate and distance gate;
• use the coupled input and forget gates to reduce the number of
  parameters, making our model more efficient

最后，总结本文的contributions：

• We propose an innovative gate mechanism way to model
  spatio-temporal intervals between check-ins under LSTM
  architecture to learn user’s visiting behavior for the next POI
  recommendation.
• A STGN model is proposed to incorporate carefully designed time gates
  and distance gates to capture the spatio-temporal interval
  information between check-ins. As a result, STGN well models user’s
  short-term and long-term interests simultaneously;
• Experiments on four large-scale real-world datasets are conducted to
  evaluate the performance of our proposed model. Our experimental
  results show that our method outperforms state-of-the-art methods.

2 Related Work

2.1 POI Recommendation

1.POI Recommendation

• Different from traditional recommendations (e.g., movie
  recommendation, music recommendation)
• POI recommendation is characterized by geographic information and
  no explicit rating information（不明确的评级信息）;
• additional information, such as social influence,temporal information, review information, and transition between POIs, has been leveraged for POI recommendation.

2.Next POI recommendation

• a natural extension of general POI recommendation;

• sequential influence between successive check-ins plays a crucial
  role in next POI recommendation;

2.2 Neural Networks for Recommendation

卷积神经网络CNN已经用于POI推荐系统。

3 Preliminaries

下面给出Next POI recommandation的形式化定义以及LSTM的简介。

3.1 Problem Formulation

Next POI recommandation的关键在于推荐的下一个POI跟时间有关，其实跟常规的POI recommandation相比，是换汤不换药的。
问题定义如下：

• 用户集：
  
• POI集：
  
• 用户u的历史POI访问序列：
  
  其中，
  表示用户u在时间ti时访问POI v。

Next POI recommandation问题的目标就是：
根据用户的历史访问记录，为用户推荐一个其在时间ti可能会访问的unvisited名单。

3.2 LSTM

过于复杂，不展开讲解，可简单参考https://blog.csdn.net/qq_33429968/article/details/114416796?spm=1001.2014.3001.5501。

4 Our Approach

4.1 Spatio-Temporal Gated Network

其核心改变就是在LSTM的基础上加了两对time和distance gates，用来控制最近一次访问的POI对下次推荐的影响。第一对time/distance gate控制short-term interest，第二对time/distance gate控制long-term interest。

为什么要加入时间门和距离门：
原因就在于最近一次访问的POI并不一定就具有很强的说服力，比如，最近一次的访问距离现在时间很长（用户已经改变了以前的行为习惯）或者最近一次访问的距离太长了（比如我家海南，去黑龙江玩了一次，现在再给我推荐黑龙江的POI就不合适了）。
改进之处就在于考虑最近一次访问的POI的时间和距离间隔，如下图：

4.2 Variation of Coupled Input and Forget Gates

4.3 Training

5 Experiments

6 Conclusions