1. Purnamrita Sarkar (UC Berkeley) Deepayan Chakrabarti (Yahoo! Research) Andrew W. Moore (Google, Inc.) 1

2.  Which pair of nodes {i,j} should be connected? Alice Bob Charlie Goal: Recommend a movie

3.  Which pair of nodes {i,j} should be connected? Goal: Suggest friends

4.  Predict link between nodes Connected by the shortest path With the most common neighbors (length 2 paths) More weight to low-degree common nbrs (Adamic/Adar) 8 followers 1000 followers Prolific common friends  Less evidence Less prolific  Much more evidence Alice Bob Charlie

5.  Predict link between nodes Connected by the shortest path With the most common neighbors (length 2 paths) More weight to low-degree common nbrs (Adamic/Adar) With more short paths (e.g. length 3 paths )  exponentially decaying weights to longer paths (Katz measure) …

6. RandomShortest Path Common Neighbors Adamic/AdarEnsemble of short paths Link prediction accuracy* *Liben-Nowell & Kleinberg, 2003; Brand, 2005; Sarkar & Moore, 2007 How do we justify these observations? Especially if the graph is sparse

7. 7 Unit volume universe Model: 1.Nodes are uniformly distributed points in a latent space 2.This space has a distance metric 3.Points close to each other are likely to be connected in the graph  Logistic distance function (Raftery+/2002)

8. 8 1 ½ Higher probability of linking radius r α determines the steepness The problem of link prediction is to find the nearest neighbor who is not currently linked to the node.  Equivalent to inferring distances in the latent space Model: 1.Nodes are uniformly distributed points in a latent space 2.This space has a distance metric 3.Points close to each other are likely to be connected in the graph  Logistic distance function (Raftery+/2002)

9. RandomShortest Path Common Neighbors Adamic/AdarEnsemble of short paths Link prediction accuracy *Liben-Nowell & Kleinberg, 2003; Brand, 2005; Sarkar & Moore, 2007 Especially if the graph is sparse

10.  Pr 2 (i,j) = Pr(common neighbor|d ij ) Product of two logistic probabilities, integrated over a volume determined by d ij As α  ∞ Logistic  Step function Much easier to analyze! i j

11. 11 Everyone has same radius r i j # common nbrs gives a bound on distance η =Number of common neighbors V(r)=volume of radius r in D dims Unit volume universe

12.  OPT = node closest to i  MAX = node with max common neighbors with i  Theorem: w.h.p Link prediction by common neighbors is asymptotically optimal d OPT ≤ d MAX ≤ d OPT + 2[ ε/V(1)] 1/D

13.  Node k has radius r k.  i  k if d ik ≤ r k (Directed graph)  r k captures popularity of node k  “Weighted” common neighbors:  Predict (i,j) pairs with highest Σ w(r)η(r) 13 i k j rkrk m Weight for nodes of radius r # common neighbors of radius r

14. Presence of common neighbor is very informative r is close to max radius Absence is very informative Adamic/Adar 1/r Real world graphs generally fall in this range i k j radius

15. RandomShortest Path Common Neighbors Adamic/AdarEnsemble of short paths Link prediction accuracy *Liben-Nowell & Kleinberg, 2003; Brand, 2005; Sarkar & Moore, 2007 Especially if the graph is sparse

16.  Common neighbors = 2 hop paths  For longer paths:  Bounds are weaker  For ℓ ’ ≥ ℓ we need η ℓ ’ >> η ℓ to obtain similar bounds  justifies the exponentially decaying weight given to longer paths by the Katz measure

17.  Three key ingredients 1. Closer points are likelier to be linked. Small World Model- Watts, Strogatz, 1998, Kleinberg 2001 2. Triangle inequality holds  necessary to extend to ℓ- hop paths 3. Points are spread uniformly at random  Otherwise properties will depend on location as well as distance

18. RandomShortest Path Common Neighbors Adamic/AdarEnsemble of short paths Link prediction accuracy* *Liben-Nowell & Kleinberg, 2003; Brand, 2005; Sarkar & Moore, 2007 The number of paths matters, not the length For large dense graphs, common neighbors are enough Differentiating between different degrees is important In sparse graphs, length 3 or more paths help in prediction.

20. 20 1 ½ Higher probability of linking Two sources of randomness Point positions: uniform in D dimensional space Linkage probability: logistic with parameters α, r α, r and D are known radius r α determines the steepness The problem of link prediction is to find the nearest neighbor who is not currently linked to the node.  Equivalent to inferring distances in the latent space

21. 1 ½ Factor ¼ weak bound for Logistic Can be made tighter, as logistic approaches the step function.

22. 22 Generative model Link Prediction Heuristics node a Most likely neighbor of node i ? node b Compare A few properties  Can justify the empirical observations  We also offer some new prediction algorithms

23.  Combine bounds from different radii  But there might not be enough data to obtain individual bounds from each radius  New sweep estimator  Q r = Fraction of nodes w. radius ≤ r, which are common neighbors.  Higher Q r  smaller d ij w.h.p

24.  Q r = Fraction of nodes w. radius ≤ r, which are common neighbors larger Q r  smaller d ij w.h.p  T R : = Fraction of nodes w. radius ≥ R, which are common neighbors.  Smaller T R  large d ij w.h.p

25. Q r = Fraction of nodes with radius ≤ r which are common neighbors T R = Fraction of nodes with radius ≥ R which are common neighbors Number of common neighbors of a given radius Large Q r  small d ij Small T R  large d ij r

26.  Which pair of nodes {i,j} should be connected?  Variant: node i is given Friend suggestion in Facebook Alice Bob Charlie Movie recommendation in Netflix

27. 27 Nodes are uniformly distributed in a latent space The problem of link prediction is to find the nearest neighbor who is not currently linked to the node.  Equivalent to inferring distances in the latent space Raftery et al.’s Model: Unit volume universe Points close in this space are more likely to be connected.

28. 28 1 ½ Higher probability of linking Two sources of randomness Point positions: uniform in D dimensional space Linkage probability: logistic with parameters α, r α, r and D are known radius r α determines the steepness

29. i j k η 1 ~ Bin[N 1, A(r 1, r 1, d ij )] η 2 ~ Bin[N 2, A(r 2, r 2, d ij )] Example graph:  N 1 nodes of radius r 1 and N 2 nodes of radius r 2  r 1 << r 2 Maximize Pr[ η 1, η 2 | d ij ] = product of two binomials w(r 1 ) E[ η 1 |d*] + w(r 2 ) E[ η 2 |d*] = w(r 1 ) η 1 + w(r 2 ) η 2 RHS ↑  LHS ↑  d* ↓

30. { Variance Jacobian Small variance  Presence is more surprising r is close to max radius Small variance  Absence is more surprising Adamic/Adar 1/r Real world graphs generally fall in this range

31.  Common neighbors = 2 hop paths  Analysis of longer paths: two components 1. Bounding E( η l | d ij ). [η l = # l hop paths]  Bounds Pr l (i,j) by using triangle inequality on a series of common neighbor probabilities. 2. η l ≈ E( η l | d ij ) Triangulation

32.  Common neighbors = 2 hop paths  Analysis of longer paths: two components 1. Bounding E( η l | d ij ) [η l = # l hop paths]  Bounds Pr l (i,j) by using triangle inequality on a series of common neighbor probabilities. 2. η l ≈ E( η l | d ij ) Bounded dependence of η l on position of each node  Can use McDiarmid’s inequality to bound | η l - E( η l | d ij )|

33. i j k η 1 ~ Bin[N 1, A(r 1, r 1, d ij )] η 2 ~ Bin[N 2, A(r 2, r 2, d ij )] Example graph:  N 1 nodes of radius r 1 and N 2 nodes of radius r 2 w(r 1 ) E[ η 1 |d*] + w(r 2 ) E[ η 2 |d*] = w(r 1 ) η 1 + w(r 2 ) η 2 (d * =MLE) Decreasing function of d * “Weighted” common neighbors Weights RHS ↑  d * ↓ Link prediction by weighted common neighbors is justified

34.  Node k has radius r k.  i  k if d ik ≤ r k (Directed graph)  r k captures popularity of node k 34 i k j rkrk m

35.  Node k has radius r k.  i  k if d ik ≤ r k (Directed graph)  r k captures popularity of node k 35 i k j Type 1: i  k  j riri rjrj A(r i, r j,d ij ) Type 2: i  k  j i k j rkrk rkrk A(r k, r k,d ij )

36. i j k Example graph:  N 1 nodes of radius r 1 and N 2 nodes of radius r 2  η 1 and η 2 common neighbors with these radii w(r 1 ) E[ η 1 |d*] + w(r 2 ) E[ η 2 |d*] = w(r 1 ) η 1 + w(r 2 ) η 2 (d * =MLE) Decreasing function of d * “Weighted” common neighbors Weights More “weighted” common neighbors  points are closer  Useful for link prediction

37.  Common neighbors = 2 hop paths  Analysis of longer paths: 1. Triangulation: ℓ-hop path as a sequence of common neighbors 2. “Metric” property: intermediate distances linked to d ij

38.  Bound d ij as a function of η ℓ  For ℓ ’ ≥ ℓ we need η ℓ ’ >> η ℓ to obtain similar bounds  justifies the exponentially decaying weight given to longer paths by the Katz measure  Also, we can obtain much tighter bounds for long paths if shorter paths are known to exist.