Kartik Hosanagar Joint work with Daniel Fleder

1. Blockbuster Culture’s Next Rise or Fall? The Effect of Recommender Systems on Sales Diversity Kartik HosanagarJoint work with Daniel Fleder Paper available on SSRN

2. Outline1. Introduction2. Problem Statement3. Simulations4. Results5. Future Work & Conclusions

3. Why the growth of recommender systems Value to consumers • Discover new products Sort through choices (Pham & Healey) Value to firms • Convert browsers to buyers Cross-sell Increase loyalty (Schafer et al.) Major firms (Amazon, NetFlix, Yahoo, iTunes) Example: “By 2010, 25 percent of online music store transactions will be driven directly from consumer-to-consumer taste-sharing apps” (Gartner-Berkman)

4. Related work Influence on choice • Individual-level effects (Senecal and Nantel, Cooke et al.) • Our focus: aggregate/market level Sales concentration • Power laws (Kohli & Sah) • Long tail and Internet (Brynjolfsson et al., Anderson) • Best-seller lists (Salganik et al.) Anecdotal views • Homogeneity (Mooney & Roy, 2000) • Diversity (Brynjolfsson et al 2006; Anderson 2004)

5. Outline1. Introduction2. Problem Statement3. Simulations4. Results5. Future Work & Conclusions

6. Context • Single market • One class of goods (e.g., movies) • Focus is sales diversity (i.e. supply is fixed)

7. Measuring diversity: Gini coefficient G = A/(A+B) Closer to 0= Equality Closer to 1 = Inequality

8. Problem statement: Set of recommender systems R = {r1,…, rk}G0 = Gini without recommendations Gi = Gini when ri is employed, all else equal Diversity Gi < G0 Bias of ri Concentration Gi > G0 Neutral Gi = G0 Prove or Test: H0: Gi = G0 vs. Ha: Gi≠ G0

9. Outline1. Introduction2. Problem Statement3. Simulations4. Results5. Future Work & Conclusions

10. Products and Consumer Ideal Points

11. Awareness

12. ≡ similarityij vij = – klog(distanceij) Choice model: Multinomial logit • Product j’s utility to i: uij = vij+ εij, • Using the map • If recommended, temporarily boost vij= – klog(distanceij)+  • Product (J +1) is outside product with fixed utility for all i • 5.

13. G0 = 0.72 Baseline: Consumer behavior without rec’s Cum % Sales Cum % Items

14. Recommender systems tested System r1 • “Standard” collaborative filter • Find n most similar users (cosine similarity) • Recommend most popular item among them Products p1 p2 p3 c1 c2 Consumers c3 c4 c5 System r2 • Discounted CF • Find n most similar users • Discount their purchases by each item’s overall popularity • Recommend most popular item among them

15. G0 = 0.72 G1 = 0.82 Experiment 1: Turn on r1 (user-to-user CF) Cum % Sales Cum % Items

16. What about r2: discount by popularity G0 = 0.72 G1 = 0.75 Cum % Sales G2 = 0.82 2 1 Cum % Items

17. Product level effects Y=X

18. Consumer level effects (r2) Before After Edge(i,j) if similarity(i,j) > 0  similarity(i,j) Line width

19. Understanding concentration bias

20. Welfare Sales ––R1 - - R2 Utility

21. An alternate interpretation Concentration No rec’s Rec’s Best-seller R0 R2 R7

22. Outline1. Introduction2. Problem Statement3. Simulations4. Results5. Future Work & Conclusions

23. Summary and additional results • Collaborative filters can help enhance sales diversity (e.g., by increasing awareness) but … a design feature, namely the use of sales data to recommend products, can often come in the way and drive up sales concentration • Individual diversity can increase (aware of more products or buy more unique products) even when aggregate diversity decreases • Basic design choices affect the outcome

24. Thank You Paper available on SSRNComments welcome

26. Sensitivity • 1. Alternate utility specifications • Variety seeking 2. Simulation parameters 3. Additional recommenders

27. Variety seeking vijt = -klog(distanceij) + I(i,j,t) + Xijt Xi,j,t = Smoothed indicators of i purchasing j =  Xi,j,t-1 + (1- )I(i bought j on t-1) • < 0 variety;  > 0 loyalty/inertia

28. Variety and Inertia

29. Different recommenders Delta = 5 Before After G r1 “Standard CF”: most popular among n most similar consumers 0.72 0.81 +.10 r2 “Discounted CF”: in r1, discount items by popularity before choosing “ 0.75 +.03 r3 “Discounted CF by TF-IDF CF: in r1, discount in user-similarity calculation “ 0.81 +.09 r4 Both types of discounting (r2 + r3) “ 0.74 +.02 r5 Least popular “ 0.47 -.25 r6 Median product “ 0.62 -.10 r7 Most popular “ 0.80 +.08 r8 Top 5 bestsellers “ 0.85 +.13

30. Outline1. Introduction2. Problem Statement3. Analytical Model4. Simulations5. Results6. Future Work

31. Future work Empirical analysis of real-world recommenders - Using data from a music recommendation service

32. Urn 1: Bernoulli • Assumptions • U = {b black ,w white} • Select ball and return • Result • P(white) = w/(b+w) = p • Outcomes independent

33. Urn 2: Polya • Assumptions • U = {1 Black, 1 Red} • Select and return • Add new ball that is the same color as the ball selected

34. Polya in Action

35. Polya in Action

36. Polya in Action

37. Polya Results • Any probability of white balls is an equilibrium and equally likely • In general, a beta distribution for 2 balls • Dirichlet for n balls • Canonical example of Path Dependence • But path dependence too strong?

38. Before After Effect is not only pronounced, but sudden

39. Product level view: Rich get richer Before After

40. Recommender systems tested “Standard” user-to-user CF “Standard” CF, but with inverse popularity weighting

41. Product level view (another run, r1)

42. Concentration for CD Albums G=.63 Cum % Purchases Cum % Items

43. Market shares by item Before After(Path A) After(Path B) Product level view (multiple runs, r1)

44. Size of similar user group nas percent of total customers

46. Similarity and k vij = -k·log(distanceij) • For G0, k controls preference breadth • For G1, this matches the model: weaker preferences allow the recommender to be more influential k

47. Salience parameter  vij = -k·log(distanceij) +  