Graph-based Point-of-interest Recommendation with Geographical and - PowerPoint PPT Presentation

Graph-based Point-of-interest Recommendation with Geographical and Temporal Influences Date : 2015/05/07 Author: Quan Yung, Gao Cong, Aixin Sun Source: ACM CIKM’14 Advisor: Jia-ling Koh Speaker: Han, Wang 1

Outline • Introduction • Approach • Experiment • Conclusion 2

Introduction Motivation: • The availability of user check-in data offers the opportunity to facilitate users’ travels and social interactions. Goal: Recommend POIs to users who have not visited before. 3

Introduction Two observation: • Geographical Influence • Temporal Influence 10 KM 2 KM 4

Introduction Framework: POIs Recommendation Step 1: Input: output: Construct Check- Input: GTAG ins, Top-k not target (Users, visited Step 2: Step 4: user, time, POIs Weight Breadth-first target POI) computation Preference time, of edges Propagation GTAG, k Step 3: Path Selection 5

Outline • Introduction • Approach • Experiment • Conclusion 6

Approach Temporal influence on check-ins: time difference � • daily periodic • � users tend to visit the same • POIs at close time 7

Approach Geographical-Temporal Influence Aware Graph(GTAG): 1. User’s interests vary with time. 2. The check-in interests of a user in the time closer to the target time are more relevant, and more important. 3. If two users have similar temporal interests(the POIs he visited in that time) in two time, they tend to visit the same POIs in the two time. 4. Users tend to visit nearby POIs. National Palace Museum correspond POI links to time slot time slot Silks Holly 20:00 Palace Check-in links 8

Approach Geographical-Temporal Influence Aware Graph(GTAG): 9

Approach Weight of edges: (initial = 1 ) # visit 1 L 1 S Ｕ 1 L 2 S 1,17:00 L 2 1 S Ｕ 1 1 S 2,18:00 L 3 Ｕ 2 1 S 2,21:00 L 4 target time = 18:00 H = 2 10

Approach Weight of edges: (initial = 1 ) L 1 Ｕ 1 1km 1 S 11 L L L 2 1/2 2km S 22 L 3 Ｕ 2 1/4 4km S 25 L 4 α = 1, β = -1 11

Approach Normalize Weight: L 1 0.14 Ｕ 1 0.71 0.76 0.23 S 11 1 0.46 L 2 0.14 0.46 0.14 0.07 Ex: 0.66 S 22 0.14 L 3 L 1 -> S 11 : 1/0.3*1+1 = 0.76 0.71 0.93 Ｕ 2 L 1 -> L 2 : 0.3*1/0.3*1+1 = 0.23 0.83 0.07 0.07 S 25 0.93 0.33 S 11 -> U 1 : 1/0.2*2+1 = 0.71 L 4 S 11 -> L 1 : 0.2/0.2*2+1 = 0.14 0.17 U 1 -> S 11 : / = 1 target time = 18:00 = 0.2, = 0.3 12

Approach Path Selection: Simple path, no repeated node in a path. • The path only contain one visited POI node and session node of the target • user, for avoiding long propagation paths. The path terminates when an unvisited POI node met. • The path always start form a target user, valid path length is 3,4 or 6. 13

Approach Breath-first Preference Propagation:(BPP) (target user: u1, k = 1) Q: U1 S11 L2 L3 r u1 :-0.71 r u1 :1 r L1 :1 L 1 0.14 Ｕ 1 r s11 :2 r s11 :1 0.71 0.76 0.23 S 11 N: 1 0.46 U1 S11 L2 L3 L 2 r L2 :1 0.14 0.46 r s22 :1 0.14 0.07 0.66 S 22 0.14 L 3 r L3 :1 � 0.71 0.93 Ｕ 2 rs11 = 1 + 1 * 1 = 2 0.83 � 0.07 0.07 S 25 ru1 = 0 - 1*1*0.71 = -0.71 0.93 0.33 L 4 0.17 Iteration 1 14

Approach Breath-first Preference Propagation:(BPP) (target user: u1, k = 1) Q: U1 S11 L1 L2 L3 r u1 :0.71 r u1 :-0.29 r L1 :1.28 r L1 :1 L 1 0.14 Ｕ 1 r s11 :-1.762 r s11 :2 0.71 0.76 0.23 S 11 N: 1 0.46 U1 S11 L2 L3 L 2 r L2 :1 0.14 r L2 :1.28 0.46 r s22 :1 0.14 0.07 0.66 S 22 0.14 L 3 r L3 :1 � � � 0.71 ru1 = -0.71 + 2 * 0.71 = 0.71 0.93 Ｕ 2 rL2 = 1 + 2 * 0.14 = 1.28 rL1 = 1 + 2 * 0.14 = 1.28 � 0.83 � � 0.07 0.07 rs11 = -0.3416 - 2*0.71*1 = -1.762 rs11 = -0.2128 - 2*0.46*0.14 = -0.3416 S 25 rs11 = 0 - 2*0.14*0.76 = -0.2128 0.93 0.33 L 4 0.17 Iteration 2 15

Approach Breath-first Preference Propagation:(BPP) Until Q is empty,stop. Output: L3(Recommendation)

Outline • Introduction • Approach • Experiment • Conclusion 17

Experiment Dataset: Foursquare vs Gowalla(shut down) Foursquare: 342,850 check-ins between Aug.2010 ~ Jul.2011 in Singapore • Gowalla: 736,148 check-ins between Feb.2009 ~ Oct.2010 in California and • Nevada Remove POIs which fewer than 5 times be checked and users who check-in • fewer than 5 times. 10% development set, 20% test set, 70% training set. • 18

Experiment Impact of Parameters: 3 0.08 Preference propagated 0.2 to user nodes H 0.01 Impact of session 270 nodes far from target time 0.3 290 Geographical influence 19

Experiment Temporal influence: 20

Experiment Geographical influence: 21

Experiment Effect of the Length time slot: 22

Outline • Introduction • Approach • Experiment • Conclusion 23

Conclusion Propose the GTAG to model check-in behaviors and a graph-based preference propagation algorithm for POI recommendation. The solution exploits both geographical and temporal influences. Conduct experiments over two real-world LBSN datasets, and the result outperforms than the others. 24