1. Recsplorer: Recommendation Algorithms Based on Precedence Mining Aditya Parameswaran Stanford University (Joint work with G. Koutrika, B. Bercovitz & H. Garcia-Molina) 1

2. Applications (Far too many!) 2

3. What’s New? Collaborative filtering Extracting patterns ~10 yrs But not used in recommendations! Challenge: Aggregation & Sparsity Sets, not Sequences Won’t need ratings! Lack of “similar people” 3

4. Motivating Example q1q2q3q4 A : 5B : 5D : 5- A : 1E : 2D : 4F : 3 G : 4H : 2E : 3F : 3 B : 2G : 4H : 4E : 4 A : 5G : 4E : 4- User u1 u2 u3 u4 u Target user G : 4E : 3H : 2 H : 4 G : 4E : 4 G : 4E : 4 H : 2 H : 4 H : 3 Ignore potentially useful information Exploit patterns only among similar users Sparsity of ratings, Few recommendations Recommend 4

5. Motivating Example (contd.) q1q2q3q4 A : 5B : 5D : 5- A : 1E : 2D : 4F : 3 G : 4H : 2E : 3F : 3 B : 2G : 4H : 4E : 4 A : 5G : 4E : 4- User u1 u2 u3 u4 u Target user A : 5D : 5 A : 1D : 4 AD AD AD E : 2F : 3 E : 3F : 3 EF GH D F H Recommend Mine a larger portion of user histories Exploit patterns across all users More and better recommendations User preferences, logical orders, interest evolution H : 2G : 4 H : 4G : 4 How to assign scores? 5

6. Goals Quality of recommendations Not enough! Coverage Goodness Unexpectedness Predictability Not covered in this talk Efficiency 6

7. Precedence Model A prediction problem using conditional probabilities Given A, what is the probability that X will followP[ X | A ] Incorrect! Contains - A X A  X X  A User Hist u1 u2 u3 u4 u5 P[ X | A ] = 1/3 P[ X | A with no X preceding ] = 1/2 P[X |A  X ] 7

8. Algorithm 1: Single Item Max-Confidence Current user’s history U D1D1 D2D2 D3D3 DmDm … X sup(D i, X)  θ P[X|D m  X ] score(X) = max i P[X | D i  X ] 8

9. Current user’s history U D1D1 D2D2 D3D3 DmDm … X score(X) = P[X | U  X ] Algorithm 2: Joint Probabilities 9

10. score(X) =P[X | U  X ] Current user’s history : U = {D 1, D 2, … D m } Approximating: score(X) = P[X | D 1  X  D 2  X  …  D m  X ] score(X)  P[X] × Π P[D i  X | X] D i in U Algorithm 2: Joint Probabilities (Contd.) 10

11. Current user’s history U D1D1 D2D2 D3D3 DmDm … X score(X)  P[X] × Π P[D i  X | X] Di in U Top Di in U Algorithm 3: Hybrid 11

12. Evaluation: Methodology Dataset: 7,500 Student transcripts from CourseRank Evaluation Methodology: Input: xHidden: r Metrics: precision@k = fraction of top-k recommendations in r coverage@k = number of users for whom an algorithm generates at least k recommendations System: CourseRank (an educational social site for Stanford) 12

13. Evaluation: Algorithms Popularity Reranked Hybrid Joint Probabilities Single Item Max Confidence Collaborative Filtering Not covered in this talk Joint Probabilities with Support 13

14. Evaluation Support θ =30, I =3 samples, x=14 14

15. Evaluation Support θ =30, I =3 samples, k=2 recommendations 15

16. Evaluation Support θ =30, I =3 samples, k=10 recommendations 16

17. Summary of Contributions Finer-grained precedence model to leverage collective wisdom Higher coverage + precision@k More in paper: other algorithms goodness / unexpectedness optimal thresholds user study 17