1、旅行目的地预测飞猪行业智能算法团队李良玥|SIGIR22 When online meets offline:exploring periodicity fortravel destination predictionWanjie Tao,Liangyue Li,Chen Chen,Zulong Chen,Hong Wen01背景介绍背景介绍02相关工作相关工作03研究方案研究方案04实验结果与结论实验结果与结论目录目录CONTENT|背景介绍01|飞猪介绍|上海品茶酒店交通旅行场景的特质|行为属性用户旅行是超低频需求，行为数据稀疏用户访问频次低、间隔长，上一次的访问信息可能失效决策属性旅行具有明

2、显的行前-行中-行后的状态转移过程，并且不同状态下存在明显的差异用户对旅行的决策期较长，会有明显的前瞻规划需求旅行目的地预测应用场景|频道入口信息流推荐研究挑战|Offline spatial-temporal periodicity时空周期性：用户一般周末或者假期出行，更经常进行短途游，偶尔长途游Online multi-interest exploration多兴趣探索：用户会探索多个目的地的旅行商品，而且倾向于去没有访问过的城市相关工作02|Next-POI RecommendationWhere You Like to Go Next:Successive Point-of-Inter

3、est Recommendation,IJCAI,2013|Next-POI RecommendationGeography-Aware Sequential Location Recommendation,KDD,2020|Next-POI RecommendationGeography-Aware Sequential Location Recommendation,KDD,2020|Next-POI RecommendationSTAN:Spatio-Temporal Attention Network for Next Location Recommendation,WWW,2021|

4、Sequential RecommendationDeep Interest Network for Click-Through Rate Prediction,KDD,2018|Sequential RecommendationDeep Interest Network for Click-Through Rate Prediction,KDD,2018|Sequential RecommendationDeep Session Interest Network for Click-Through Rate Prediction,IJCAI,2019|Sequential Recommend

5、ationSparse-Interest Network for Sequential Recommendation,WSDM,2021|Sequential RecommendationSpatial-Temporal Deep Intention Destination Networks for Online Travel Planning,TIST,2021|解决方案03|我们的方案 Online-offline periodicity-aware information gain networkOOPIN|Offline Mobility Pattern Extractor|提取离线行

6、为的时空周期性离线行为矩阵基于CNN的时空周期性提取Periodicity-aware GRU Layer|提取离线行为的序列演变时空周期性感知的GRUDistance-aware Self-Attention Net(DSN)|用户线上探索的城市通常会呈现多个聚类例如长三角地区，或者京津冀地区Target-aware Attention Net(TAN)|从在线行为中提取出跟离线行为相似的信息Information Gain Net(IGN)|从在线行为中提取出信息增益即跟离线行为不一样的信息Final Prediction Layer|对每一个候选城市进行打分实验结果04|Datasets

7、|All the data are user authorizedStats:1.76M users,341 citiesavg length of online behavior seq is 47avg length of offline mobility seq is 18Process:label the cities the user visited as positive samples,while negative samples are randomly sampledBaselines|Sequential RecommendationsGRU4Rec ICLR 16SASR

8、ec ICDM 18DSIN IJCAI 19SINE WSDM 21Next POI recommendationsSTRNN AAAI 16STGN TKDE 20GeoSAN KDD 19STAN WWW 21Comparison Results|Observations:SINS STAN(online behavior is more predictive of next destinationAmong next-POIs,STAN is best since it explicitly considers the spatial-temporal preferences with

9、 a POI graph OOPIN is best by jointly modeling the spatial-temporal periodicity from the offline mobility data and the information gain from the online behavior data.Ablation StudiesEffectiveness of offline and online modules|Ablation StudiesEffectiveness of several key ingredients|Online A/B testin

10、g|Setting:Deploy OOPIN to“guess what you like”and assigns higher scores to OOPIN predicted destinations on top of the original ranking scores Results:During 10 days,consistently outperform baselineOn avg,3.73%lift on CTROther Works|-冷启动用户-SMINet:State-Aware Multi-Aspect Interests Representation Netw

11、ork for Cold-Start Users Recommendation.Thirty-Sixth AAAI Conference on Artificial Intelligence(AAAI),2022.-价格竞争力建模-Cheaper is Better:Exploring Price Competitiveness for Online Purchase Prediction.38th IEEE International Conference on Data Engineering(ICDE),2022.-酒店搜索异构实体间的地理关系-G2NET:A General Geography-Aware Representation Network for Hotel Search Ranking.KDD,2022.非常感谢您的观看|