THE LIMITS OF RECOMMENDERS AND THE ROLE OF SOCIAL TIES ALESSANDRO PANCONESI, SAPIENZA UNIVERSITY OF ROME JOINT WITH: MARCO BRESSAN, STEFANO LEUCCI, PRABHAKAR RAGHAVAN, ERISA TEROLLI THEGOODOLDDAYS RS’s are now quite standard on the web Ecommerce sites: try to optimize revenue by getting users to buy more. News and media sites: try to keep the users interacting with content, because it generates (advertising) revenue. Users would get more utility if they receive recommendations they find valuable and trustworthy. RECOMMENDER SYSTEMS, THE MAIN GATEWAY TO ONLINE CHOICES AS SUCH THEY ARE BECOMING POWERFUL ECONOMIC ACTORS TO WHAT EXTENT CAN A RECOMMENDATION SYSTEM ALTER A MARKET? DIGITAL FOODS: TURNING BITS INTO BITES Market shares, before RS KDD FOODS: TURNING BITS INTO BITES Market shares, after RS KDD FOODS: TURNING BITS INTO BITES Different approaches User studies / social psychology experiments Data mining with real data Mathematical modelling A MATHEMATICAL MODEL Main Ingredients the more successful an item, the higher the probability of being recommended social forces: recommendations are made by friends PRODUCTS USERS (or BUYERS) Users are connected via a network 0.4 0.6 0.6 0.3 0.8 0.7 1.0 0.5 0.1 0.3 0.1 0.7 0.5 0.1 0.3 0.7 Trust 0.49 0.01 0.48 0.01 0.01 Personal inclinations 0.2 0.1 0.05 0.4 0.1 0.25 The purchasing process Items are bought in a sequence, when a purchase is made a buyer is chosen at random Following recommendations Recommendation 1 ↵Shreck ↵Shreck fShreck Following recommendations 0.01 Shreck is the current buyer and he will follow Wile Coyote’s suggestion The weight of history bought at time t 0.01 Shreck is the current buyer and he will follow Wile Coyote’s suggestion The weight of history bought at time t 0.01 Shreck is the current buyer and he will follow Wile Coyote’s suggestion A natural condition: The past fades away If eventually we put zero probability on the item purchased at any fixed time <t, we say that the past fades away. E.g., uniform on all past purchases, or.. truncated beyond recent history (e.g. consider only last 10 items), or.. recency effect, weights are 1, 1/2, 1/4, ..etc. START THE SYSTEM. WHAT HAPPENS? MAIN RESULT If the past fades away, then the system will converge to a unique steady state MAIN RESULT …furthermore, the shape of the market at the limit is given in closed form A bit of notation t xu (p) fraction of times that u has bought p, up to time t From now on product p is fixed. We drop p for notational convenience t x := t t (x1 , . . . , xn ) Main result Fix a product p. If the past fades away, then t lim E[x ] = x t!1 1 THE SYSTEM ALWAYS CONVERGES TO A UNIQUE LIMIT! Main result Fix a product p. If the past fades away, then t lim E[x ] = x t!1 x L := (I 1 1 = Lb. AM ) 1 (I A) The matrix L captures market effects Corollaries x 1 = Lb. x1 i = 0.2 n X = Lij bj j=1 = 0.001 + 0.002 + . . . + 0.1 + . . . + 0.003 We can quantify the influence of each user on every other user Influence and distortion The relative contribution of each u to the overall sales distribution is an influence vector over users: := f L Define the market distortion as the ratio of the market share of p* with/without the RS: := ( · b)/(f · b) Some known precedents If all ↵u = ↵ , then turns out to be exactly the vector of personalized pageranks [HKJ2003] Consider a coalitional game where each user either joins the coalition for p* or sits out. Then the Shapley value of user u is u bu EXPERIMENTS Experiments How do influence and distortion play out in real social networks and RS? Public datasets drawn from Google+, Twitter, Slashdot, Yelp and Facebook. Set all 𝛼 to be 0.2, which is abnormally large, to see how much we can distort the market. Convergence Does the RS distort the market? “Real” social networks dampen influence pretty heavily, with 𝝙 always measured in the range 1 ± 0.002 On the other hand, planting an oligarchy (where everyone follows a few superstars) results in high 𝝙 Influence In these “real” social networks, it was difficult to build large influence. This holds even under “nonlinear” experiments where items were recommended with heavier probability than uniform sampling. Influencers Although in real-world graphs there are users whose influence is 10,000 times that of an average user, their global effect on the market is tiny Summary Fairly general model of RS with a view to studying market influence. Closed form for equilibrium, influence and market distortion. Experiments suggest that “real” social networks dampen the influence of RS. Future research We view this as a proof of concept. Try to extend approach to more realistic models Thanks
© Copyright 2026 Paperzz