1. LARS : A Location-Aware Recommender System ICDE ‘12 1

2. 1. Introduction Traditional recommender systems – triple(user, rating, item) – (user id U) + (limit K) return K recommended items to U Locations – destinations check-in(Facebook, Foursquare) – user zip code(MovieLens) 2

3. 1.1 Motivation: A Study of Location-Based Ratings Preference locality 3

4. 1.1 Motivation: A Study of Location-Based Ratings Travel locality 4

5. 1.2 LARS - A Location-Aware Recommender (user, ulocation, rating, item) (user, rating, item, ilocation) (user,ulocation, rating, item, ilocation) 5

6. 1.2 LARS - A Location-Aware Recommender next…… 2. an overview of LARS 3. spatial user ratings for non-spatial items 4. non-spatial user ratings for spatial items 5. spatial user ratings for spatial items 6. experimental analysis 6

7. 2.1 LARS Query Model (user id U) + (limit K) + (location L) ——> return K recommended items to U query ： – snapshot (one-time) queries – continuous queries 7

8. 2.2 Item-Based Collaborative Filtering Phase I: Model Building – 计算 item 间的相似度 sim – 对于每个 item 模型只会存储前 n 个相似度最高的 sim 值 n 为 user 个数 Phase II: Recommendation Generation 8

9. 3 Spatial User Ratings For Non-spatial Items (user, ulocation, rating, item) requirements – Locality( 局部性 ) ：能对地点感知 – Scalability( 可扩展性 ) ：能够对大量的用户进行 运算 – Influence() ：用户能够改变感知的区域大小 9

10. 3.1 Data Structure partial pyramid structure— 局部锥形结构 10

11. 3.1 Data Structure Level 3 Level 2 Level 1 Level 0 天朝南广东珠三角粤东粤西粤北广西福建海南北中西四川云南西藏青海 11

12. 3.2 Query Processing query processing steps – 1. 从最底层找起 – 2. 如果没找到 去上一层找 – 3. 直到找到为止 12

13. 3.2 Query Processing Continuous queries – ( 一边移动 & 一边查询 ) – 1. 如果没有离开上一次查询时所在的 grid 还是原来熟悉的结果 – 2. 否则 去上一层找，找到为止 13

14. 3.3 Data Structure Maintenance 当有 new users,ratings,items 时 Trigger: N% ( 才会启动 Maintenance) The maintenance will be amortized( 均摊 ) Step I: Model Rebuild Step II: Merging( 合并 )/Splitting( 分裂 ) Maintenance 14

15. 3.3.1 Cell Merging Impoves scalability – storage less CF models size( 只需储存于高层，底层不储存 ) 主要标准 – computational overhead less maintenance computation 维护次数减少 less continuous query processing computation 次要标准 Hurts locality 15

16. 3.3.1 Cell Merging Two percentage values – locality_loss – scalability_gain A system parameter M ∈ [0,1] Merges if ： – M 越小，则越倾向于合并 16

17. 3.3.1 Cell Merging Calculating locality_Loss – Sample – Compare 17

18. 3.3.1 Cell Merging Calculating scalability_gain – ( child cells ) / ( child cells + parent cell ) – st 还是举之前的栗子 – scalability_gain 4 child cells == 2GB parent cell == 2GB scalability_gain=50% 18

19. 3.3.1 Cell Merging locality_loss=25% scalability_gain=50% Assuming M=0.7 but (0.3*50%)<(0.7*25%) will not merge 19

20. 3.3.2 Cell Splitting 其效用与 Cell Merging 相反 – Improves locality – Hurts scalability 计算与 Cell Merging 基本相同 – locality_gain – scalability_loss 20 MergingSplitting locality_losslocality_gain scalability_gainscalability_loss

21. 4 Non-spatial User Ratings For Spatial Items (user, rating, item, ilocation) travel locality travel penalty – – expensive computational overhead – so,employs “early termination” 21

22. 4.1 Query Processing Algorithm – 1. 找出全部 item 中， TravelPenalty 最小的 k 个 item ，将 k 个 item 按照 RecScore 从大到小排序，形成表 R – 2. 设 LowestRecScore 为 R 中最小的 ( 也就是第 K 个 ) RecScore 值 – 3. 找出剩余 item 中 TravelPenalty 最小的 item 4. 设 MaxPossibleScore = MAX_RATING – TravelPenalty 5.IF MaxPossibleScore <= LowestRecScore – 6. 不再找了，直接 return R 7. 算出此 item 的 RecScore 8.IF RecScore > LowestRecScore – RecScore 替换 LowestRecScore 进入 R – 重新找一个 LowestRecScore – 回到 3 22

23. 4.2 Incremental Travel Penalty Computation Incremental KNN – online – Exact – Expensive Penalty Grid – Offline – Less exact – Efficient 23

24. 5 Spatial User Ratings For Spatial Items (user, ulocation, rating, item, ilocation) user partitioning & travel penalty – can be used together – with very little change 24

25. 6 Experiment test recommendation quality – Foursquare ： real dataset – MovieLens ： real dataset test scalability and query efficiency – Synthetic ： synthetically generated dataset 25

26. 6 Experiment CF: item-based collaborative filtering LARS-T: LARS with only travel penalty LARS-U: LARS with only user partitioning LARS: LARS with both techniques default parameter – M == 0.3 – k == 10 – the number of pyramid levels (h) == 8 26

27. 6.1 Recommendation Quality for Varying Pyramid Levels 27 80% 训练， 20% 验证： Measure （ Quality ） – 统计预测的推荐结果 进入真实评分前 k( 默认 k=10) 的次数 层次分太细，每个 grid 中 rating 太少

28. 6.2 Recommendation Quality for Varying Values of k 28

29. 6.3 Storage Vs. Locality 29 Note ： M 越小 ，越倾向于合并 M 越大，越倾向于分裂

30. 6.4 Scalability 30 Default ： M=0.3 LARS is acceptable.

31. 6.5 Query Processing Performance 31 单次查询： LARS vs LARS-U LARS vs LARS-T 通过对比可以发现 之前两种技术所带来的时间上的优势 连续查询： CF 最快（那当然了 -_-# ） 除此之外， LARS 最快