Do Social Explanations Work? Studying and Modeling the Effects of Social Explanations in Recommender Systems

1. INTRODUCTION

• Social explanations are common, but their effects on recommendation decisions and the acceptability of disclosing others’ interests or past activity remain underexplored.Such explanations can imply endorsement of recommended items.Source: §1 · p. 1
• The authors distinguish pre-consumption likelihood of checking out an item from post-consumption evaluation, treating the ratings as measures of persuasiveness and informativeness.Persuasive explanations may raise likelihood, while informative explanations may make likelihood align with consumption ratings.Source: §1 · p. 1
• In the music study, explanation types affected likelihood, but their influence was secondary to users’ inherent expectations; trusted friends could help, while unknown or incompatible friends could hurt.Users’ comments suggested different strategies for interpreting social explanations, giving personalization potential value.Source: §1 · pp. 1–2
• The paper proposes a generative model of how explanations and inherent preferences shape likelihood ratings, with potential to generalize to other explanation sources.The authors also argue for modeling both click-through likelihood and consumption preferences rather than optimizing only one.Source: §1 · p. 2
• Low correlation between likelihood and consumption ratings for the same artist suggests persuasive explanations may not be informative about actual liking.The study’s second phase had participants listen to music about a week later; the effect of explanation type did not transfer to consumption ratings.Source: §1 · p. 2

2.1 Explanations in recommender systems

• Prior recommender-system research studied explanation interfaces and found that a histogram of similar users’ ratings was most persuasive for likelihood of seeing a movie.Other work distinguishes promotion—persuading someone to try an item—from satisfaction—helping someone judge its quality accurately.Source: §2.1 · p. 2
• A key challenge is separating explanation effects on likelihood and consumption from item merits and other interface information; this study compares social strategies and delays consumption ratings.The delay is intended to help distinguish lasting effects from short-term influence.Source: §2.1 · p. 2

2.2 Social information for recommendation

• Social networks provide information for recommendation, and users have been reported to prefer known friends over algorithmically selected similar neighbors [3].Despite practical use of friends, egocentric networks, and overall popularity in explanations, their effects on likelihood and consumption decisions had received little study.Source: §2.2 · pp. 2–3

3. SOCIAL EXPLANATIONS IN THE WILD

• Social explanations appear across social networks and search, presenting who or how many people have liked, watched, or rated an item.Examples include Facebook and social recommendation sites such as Getglue or Hunch.Source: §3 · p. 3
• Social explanations can show overall or friends’ Likes, name particular friends, or combine names and counts, drawing on social-proof, friendship, and trust ideas.Figure 2 presents the experimental strategies alongside artist information, balancing realistic recommendation presentation against an otherwise sparse interface.Source: §3 · p. 3, Figure 2 · p. 3
• The study asks how explanation strategies affect likelihood ratings, interact with users’ inherent preferences, can be modeled, and relate to later consumption ratings.The motivating system-design questions include choosing explanations and recommendations to support goals such as end-user satisfaction.Source: §3 · p. 3

4. EXPLOREMUSIC: A USER STUDY

• ExploreMusic used Facebook data to explain music recommendations; music made it practical to collect consumption ratings, while Facebook provided social connections and artist preferences.The study involved 237 users; 175 completed both phases, yielding 4,458 Phase I ratings and 835 Phase II ratings.Source: §4 · p. 3, §4.2 · p. 4

4.1 Experiment Design

• In a within-subjects Phase I design, participants saw selected unfamiliar artists randomly assigned to five strategies: overall popularity, friend popularity, random friend, good friend, and good friend plus count.Good friends were selected using interaction frequency as a proxy for tie strength; items without a nonzero-tie-strength friend were skipped in those conditions.Source: §4.1 · pp. 3–4
• For each recommendation, users rated how likely they were to check out the artist and how sure they were of that answer on an inclusive 0–10 Likert scale; artist order was randomized.The display showed the artist’s name, Facebook profile picture, and associated explanation.Source: §4.1 · p. 4
• After seeing the recommendations, users answered questions about which explanations were convincing or effective, how they used the information, and their feelings about using friends’ social data.Source: §4.1 · p. 4
• In Phase II, participants listened to three songs by two randomly selected artists per strategy, then rated liking and confidence on a 0–10 scale.The study required at least three days between phases, did not show explanations during listening, and participants averaged 5.2 days between phases.Source: §4.1 · p. 4

5.1 Are different social explanations more persuasive on average?

• GoodFrCount was significantly higher than RandFriend (p = 0.002) and FriendPop (p = 0.006), although the overall strategy difference was significant (F(4, 763) = 4.96, p = 0.0006).GoodFrCount and GoodFriend had relatively high mean likelihood ratings, while FriendPop and RandFriend had relatively low ones.Source: §5.1 · p. 4
• Participants especially valued named, close friends whose music tastes they knew, trusted, or considered similar to their own.Participants described friend names as more useful than a vague count, and close friends as more influential and informative.Source: §5.1 · pp. 4–5
• Popularity cues divided participants: many likes could signal quality, but friend counts lost value when friends’ tastes mismatched.Some participants found a large number of likes persuasive, while others said popularity mattered only when they identified with the crowd.Source: §5.1 · p. 5

5.2 How important are social explanations in decision-making?

• Influence varied substantially across people and depended on whether they expected the social information to be informative.Participants reported relying on trusted or familiar friends, or on a large number of friends who liked an artist.Source: §5.2 · p. 5, Table 3 · p. 5
• Even between GoodFriend and RandFriend, the effect was small (Cohen’s d ≈0.2), while users’ underlying likelihood model had stronger influence than explanations.The authors characterize explanations as a second-order effect, with high variability in likelihood ratings.Source: §5.2 · p. 6
• Preconceptions and artist cues also shaped judgments: genre, name, or image could reinforce or outweigh social endorsements, sometimes prompting rejection.Participants described combining social explanations with prior impressions of an artist, including using those cues to decide not to look up a recommendation.Source: §5.2 · pp. 5–6
• Most likelihood ratings were below 2 across strategies, consistent with an underlying decision tendency that leaned toward rejection.Figure 3 shows an overall distribution with a mode of 0.Source: §5.2 · p. 6, Figure 3 · p. 6

6. A GENERATIVE MODEL FOR LIKELIHOOD RATINGS

• Likelihood ratings are modeled as a mixture of independent distributions for inherent item expectations and social-explanation effects.The model treats ratings as continuous despite the discrete 0–10 data, and parameter a controls how strongly explanations affect decisions.Source: §6 · p. 6
• The base likelihood distribution includes users’ preferences and the artist’s name and photo, and is modeled as exponentially decaying for unfamiliar artists.The decay parameter α represents individual discernment; more discerning users tend to give fewer high ratings.Source: §6 · p. 6

6.1 Aggregate effects of explanation strategies

• GoodFrCount had the lowest rigidness, while GoodFrCount and GoodFriend had more ratings above 5, indicating these strategies were more consistently persuasive.Across strategies, the fitted α values were similar, supporting a shared inherent discernment parameter for the average user.Source: §6.1 · p. 7, Table 5 · p. 7, Figure 4 · p. 7

6.2 Different users, different models

• Users formed three clusters with increasing receptiveness to explanations, from little influence to explanations helping make the decision.Parameter a decreases across clusters; Cluster 1 is dominated by the exponential component, while Cluster 3 is highly Gaussian.Source: §6.2 · p. 7, Figure 5 · p. 8, Table 6 · p. 8
• The model suggests personalizing explanations by learning users’ reactions and identifying which explanation types suit them.The paper proposes starting with population-average parameters, then using explicit feedback or inferred reactions as data accumulates.Source: §6.2 · p. 8

7.1 Do explanations affect consumption ratings?

• F(4, 378) = 1.64, p = 0.2; consumption-rating differences across explanation strategies were not significant.OverallPop had the highest mean, whereas GoodFrCount performed best for likelihood; the strategy means were close apart from FriendPop.Source: §7.1 · p. 8, Table 7 · p. 8

7.2 Does likelihood predict consumption?

• r = 0.17; z-score-adjusted likelihood and consumption ratings had little correlation, supporting separate treatment of persuasiveness and informativeness.Consumption ratings were close to uniformly distributed across the 11-point scale, apart from very high ratings and the anomalous 5.Source: §7.2 · p. 8, Figure 6 · p. 8, §7.1 · p. 8

7.3 Modeling Likelihood and Consumption

• A sequential framework can maximize consumption rating R subject to likelihood rating L exceeding a user-specific threshold εu.The threshold can be adjusted to retrieve enough recommendations or reflect goals such as serendipity; when L and R are highly correlated, this reduces to maximizing R.Source: §7.3 · p. 9

• Social explanations, especially close-friend cues, are persuasive but secondary to item expectations, and expected liking is a poor proxy for actual artist liking.Source: §8 · p. 9
• Because reactions depend heavily on expected informativeness, interfaces could expose similarity or an explainer’s expertise to make social cues more informative.Source: §8 · p. 9
• Separating persuasiveness from informativeness gives designers flexibility to support different goals, including serendipity, but persuasion may erode trust when users consume items they dislike.Source: §8 · p. 9
• Explanations can use item tokens, item-attached data, people metadata, or algorithm information, but their relative informativeness and persuasiveness remain open questions.Source: §8 · pp. 9–10
• Personalizing explanation strategies may extend the model beyond social cues when effects are nearly Gaussian; finer variations within a strategy remain future work.Source: §8 · p. 10
• Privacy and transferability bound application: participants generally accepted disclosure in music, but the study sampled young university users, friends’ Likes, and sparse list interfaces.Source: §8 · p. 10

9. CONCLUSION

• Low correlation between likelihood and consumption ratings suggests that modeling persuasiveness alone may not capture informativeness.The paper proposes an optimization framework that considers both ratings.Source: §9 · p. 10
• The generative model explains much of the variation in likelihood ratings and can be personalized.The paper also proposes an optimization framework for considering likelihood and consumption ratings together.Source: §9 · p. 10
• Explicitly modeling explanations and other interface information could inform interface design and help improve perceived recommender quality.The conclusion identifies which explanations are both persuasive and informative, and for which users, as an open question.Source: §9 · p. 10