![OSN architecture
Undirected social network user topology
No node has information about entire graph
Identity of a node available to only friends
Each edge has an associated trust value
Each user v has profile attributes K vPAtt
Profile attributes built by set of keywords
Each keyword k has associated Policy(k)
Policy(k) = [D, T]
FAtt
Each user v has friendly attributes K v
6](https://image.slidesharecdn.com/searchtalk-090831013428-phpapp01/85/Information-Flow-and-Search-in-Unstructured-Keyword-based-Social-Networks-6-320.jpg)
Information Flow and Search in Unstructured Keyword based Social Networks
- 1. Information Flow and Search in Unstructured Keyword based Social Networks Ankush Garg, Prantik Bhattacharyya, Charles U. Martel, S. (Shyhtsun) Felix Wu SMW '09, August 30, Vancouver, Canada
- 2. Introduction How to search information in online social networks (OSNs) ? 2
- 3. Searching in OSNs Expectations in terms of relevancy of result results from direct friends results from trustworthy friends Challenges in the absence of structure high node degree high clustering low diameter 3
- 4. Search Problem Search a set of users with queried keyword as profile attributes Output Relevant results first Constraints Minimum resource usage 4
- 5. Outline Introduction OSN architecture Information Flow Model Search Algorithm Simulation Methodology Results Related Work Concluding Remarks 5
- 6. OSN architecture Undirected social network user topology No node has information about entire graph Identity of a node available to only friends Each edge has an associated trust value Each user v has profile attributes K vPAtt Profile attributes built by set of keywords Each keyword k has associated Policy(k) Policy(k) = [D, T] FAtt Each user v has friendly attributes K v 6
- 7. OSN architecture (more) D: maximum distance from where k is visible and v can be contacted from users T: min trust of each edge in social path required to let users contact v Each edge has trust > 0.5 Each oval shows how keywords will flow and conversely, can be searched Example: Keyword based social network 7
- 8. Information Flow Model Motivation Information maintenance and update Diffusion of non-inclusive keyword data Integral to search algorithm development Users diffuse keyword information Along with privacy consideration 8
- 9. Information Flow Model (more) Hr: Hops remaining Hc: Hops covered PID; propagation ID to avoid cycles Identity of source suppressed to non- direct friends Non-inclusive data propagation ∆Hr > 0 || ∆T < 0 || ∆Hc < 0 9
- 10. Information Flow Model (more) Keyword Forwarding Table, FTw F5 F9 Keyword Propagation Data Friends K1 {PID1,Hr1,Hc1,T1} (F1, F2,F3) W Keyword Received Table, RTw Keyword K1 Max. Hops HmaxK1 (Friend, Min. Hops) (F9, Hmin F ),, (F5, HminK ) K1 9 1 F1 F2 F5 Please see paper for details 10
- 11. Search Algorithm Fast lookup for boundary conditions No results possible (no entry exists for query keyword) Information provider are only the direct friends (Hmax == 1) Components: Metric to define value of a direct friend Threshold function for dynamic network pruning Query message processing algorithm 11
- 12. Search Algorithm (more) For user w Selecting Topologically Closer Nodes min k∈Sk Hmax k − max k∈Sk H min u k distance value, DV (u , Sk ) = min k∈Sk H max k Selecting Trustworthy Nodes Twu trust valu e, TV (u , Sk ) = Sk Tmax Value for keyword set for a direct friend V(u, Sk) = ρ × DV(u, Sk) + (1− ρ) ×TV(u, Sk), 0 ≤ ρ ≤1 12
- 13. Search Algorithm (more) Threshold Function Θ ( u , Q k ) = max V ( w , Q k ) − f ( N uQ k ) × w∈ N u Q k ( maxk V ( w , Q k ) − min k V ( w , Q k )) w∈ N u Q Q w∈ N u Pruning function f with properties: f (1) = 1 and limN Qk →∞ f ( Nu k ) = 0 Q u We use g(x) = x-p for p ≥ 0 as f p = 0, f ( N u k ) = mink V ( w, Q k )) Q Q w∈ N u Breadth First Search (BFS) 13
- 14. Search Algorithm (more) Query message structure <QID, Qk, Tmin, Hopsdone, Hopsleft> Starting Query Structure <qid, Qk, 1, 0, Hl> Query Serviced Table, QIDu: <QID, direct friend who forwarded the query> 14
- 15. Search Algorithm (more) Advantage of Dynamic Pruning Reduces number of messages sent in network Makes unstructured social network scalable Selects targets of high value topologically closer and trustworthy Returns a set of good results amongst all available obtainable results 15
- 16. Simulation Methodology Newman Watts Model Trust Distribution Five Categories of Trust From ‘Blind Trust’ (0.9) to ‘Don’t Know’ (0.1) Information propagation policy Restrictive policy Depth set to 2 i.e. friends of friends Liberal policy Depth set randomly between 1 and graph diameter 16
- 17. Performance at ρ values Helps us understand system resource consumption Shows significant reduction in message generation when compared to BFS (12 compared to 121) Analysis of network with restrictive propagation policy 17
- 18. Performance at ρ values (more) Helps us understand how many queries are successful to find results Significant improvement at ρ = 0.5 when compared to BFS Analysis of network with restrictive propagation policy 18
- 19. Performance at ρ values (more) Similar observation with higher levels of improvement Reduction in number of query messages generated from 17855 to only Analysis of network with liberal 100’s.. propagation policy 19
- 20. Performance at ρ values (more) Similar results with higher levels of performance Best results when hop and trust are considered together Results dip when either of the parameters are considered Analysis of network with liberal separately propagation policy 20
- 21. Performance at hop values Computed as a fraction of total number of results found by BFS Please see paper for more results Analysis of network (liberal policy) for various hops and at ρ = 0.5 21
- 22. Performance at hop values (more) Higher denominator in pruning function associated with increased number of successful query messages Dynamic pruning using friend values is an effective way Analysis of network with liberal propagation policy 22
- 23. Related Work Search in social networks Algorithms using structural knowledge of the network through geographical distance, organizational hierarchy, interest of users.. Search in decentralized and unstructured networks BFS, random BFS, Intelligent BFS, directed BFS, iterative BFS, random walks, k random walks, other variation of random walks.. 23
- 24. Concluding Remarks Developed a Search Algorithm Using an Information Flow model With focus on decentralization and privacy Concentrates on finding set of good results Dynamically prunes the network to search Improvement in orders of magnitude Next step Evaluation using larger graph topologies 24
- 25. Thanks! 25