Source-linked AI summary
The Role of Social Networks in Online Shopping: Information Passing, Price of Trust, and Consumer Choice
Stephen Guo, Mengqiu Wang, Jure Leskovec
TL;DR
The paper asks how social networks shape online purchasing, an area not previously studied at large scale. Using Taobao’s integrated trade, messaging, contact, and reputation data, it analyzes information passing, the price of trust, and seller choice. The results show information passing and social-network structure are important, while seller ratings carry a small but super-linear price premium and social features predict seller choice substantially better than baseline.
Problem
Large-scale evidence on how social networks and commerce networks interact in online shopping remains limited.
Method
The study models Taobao trades, messages, and contacts as networks, analyzes directed triadic closure and seller ratings, and predicts consumer choices with machine learning.
Results
For ten possible sellers, the model predicts the correct seller 42% of the time, approximately four times better than baseline, while social-network features are most important.
Takeaways & Limitations
Buyer-buyer communication is a primary purchasing driver, and social-network structure is a stronger consumer-choice signal than seller reputation or product price.
Takeaways & Limitations
The study identifies future needs for browsing data and models that incorporate both buyer-buyer and buyer-seller trust relationships.
Abstract
from arXiv · showhide
While social interactions are critical to understanding consumer behavior, the relationship between social and commerce networks has not been explored on a large scale. We analyze Taobao, a Chinese consumer marketplace that is the world's largest e-commerce website. What sets Taobao apart from its competitors is its integrated instant messaging tool, which buyers can use to ask sellers about products or ask other buyers for advice. In our study, we focus on how an individual's commercial transactions are embedded in their social graphs. By studying triads and the directed closure process, we quantify the presence of information passing and gain insights into when different types of links form in the network. Using seller ratings and review information, we then quantify a price of trust. How much will a consumer pay for transaction with a trusted seller? We conclude by modeling this consumer choice problem: if a buyer wishes to purchase a particular product, how does (s)he decide which store to purchase it from? By analyzing the performance of various feature sets in an information retrieval setting, we demonstrate how the social graph factors into understanding consumer behavior.
1. INTRODUCTION
The paper studies how social networks shape online commerce, using Taobao to examine information passing, trust premiums, and seller choice. It combines network analysis with transaction, reputation, and machine-learning evidence.
- Information passing: Information passing is defined as a buyer purchasing a product, messaging a friend, and the friend subsequently purchasing it.
- Approach: The study analyzes one million Taobao users by modeling trades, messages, and contacts as interconnected network edges.
- Information passing: Information passing rises with message strength, falls with product price and recommendation delay, and varies substantially across product categories.
- Price of trust: Across more than 10,000 products, seller ratings have a small but super-linear association with the price premium sellers can charge.
- Consumer choice: For ten possible sellers, the consumer-choice model predicts the correct seller 42% of the time, approximately four times better than baseline, with social-network features most important.
- Research gap: Prior e-commerce research examined social networks mainly through recommendations, recommendation engines, or limited datasets.
2. TAOBAO NETWORK
The Taobao dataset combines transactions, messages, and contacts for one million users during a two-month observation period. Its network distinguishes directed trades and messages from undirected contacts, revealing different reciprocity patterns.
- Dataset: Taobao is a large electronic marketplace with an integrated instant-messaging platform connecting users during commerce.
- Dataset: The dataset contains all observed transactions, contact lists, and message timestamps for the first one million transacting users from September 1 through October 28, 2009.
- Network representation: The network is modeled as a multigraph containing directed trade edges, directed message edges, and undirected contact edges.
- Network statistics: 1,337,497 directed trade edges correspond to 1,336,502 undirected trade edges because purchases are almost never reciprocated.
- Node roles: Buyer and seller labels describe a user's role in a particular transaction rather than a permanent node attribute.
3. DYADIC RELATIONSHIPS
Dyadic analysis finds that messaging and trading are positively related, with messaging patterns varying by product price and transaction timing. These relationships support the view that commercial activity is embedded in Taobao’s social networks.
- Messaging volume rises with trade volume across user pairs, with an even stronger relationship among pairs that both message and trade.
- Message counts remain relatively constant below 100 CNY, then increase logarithmically for more expensive products.The pattern is consistent with messaging helping buyers minimize transaction risk.
- Most buyer-seller messages occur on the transaction date, while post-trade messages are significantly more common than pre-trade messages.
4. INFORMATION PASSING
The study finds non-random information passing in Taobao: communication between buyers is associated with subsequent purchases from the same seller, with success shaped by timing, price, and product category. Directed triadic analysis further characterizes how social and commercial links close.
- Information-passing evidence: More mutual contacts increased the likelihood that two users engaged in a commercial transaction, especially among users who had exchanged messages.Trades were more likely to be embedded in dense communication subgraphs.
- Information-passing evidence: 0.00203 versus 0.00006: information-passing success was two orders of magnitude more likely than in the rewired random baseline.The comparison covers 3,906,354 node-pair instances and indicates non-random recommendation success.
- Influences on information passing: Stronger buyer-buyer communication increased the probability that the second buyer purchased from the first buyer’s seller.The pattern remained similar, though slightly lower, when restricting the second buyer to a first purchase from that seller.
- Influences on information passing: Information-passing success decreased as the time between the initial purchase and recommendation increased, and decreased with price from 1 to 15 CNY before rising slightly.The time pattern held for both first-time and unrestricted trades.
- Influences on information passing: Recommendation success varied strongly by product category, exceeding 20% for women’s clothing but reaching only 1.47% for home decorations.The authors summarize information passing as proportional to communication strength and inversely proportional to time difference and product price.
- Directed triadic closure: Directed configuration sets represent two time-ordered edges into a middle node followed by a possible triad-closing edge, enabling comparison of closure types and hypothesized buyer or seller roles.The analysis finds little brokerage or reselling, while seller-role configurations occur approximately 100 times more often than buyer-role configurations.
5. PRICE OF TRUST
The paper measures the price of seller trust by comparing prices for matched products across sellers with different ratings. Higher ratings are associated with price premiums, although the small elasticity suggests that other factors, including social-network context, also matter.
- Data preparation: Seller ratings were used as a proxy for reputation, and matched product listings were grouped into high-precision product clusters for price comparisons.The study compares sellers offering the same product rather than unrelated listings.
- Quantifying trust: R2 of 0.80: a power function fit the relationship between seller rating and per-listing price deviation from the cluster median.A 97.1% seller rating corresponded to the median cluster price.
- Quantifying trust: Higher seller ratings were associated with selling at a premium relative to most peers for the same product.The association was observed in the per-item comparison against median cluster prices.
- Quantifying trust: R2 of 0.87: aggregating sellers’ average price differences across items produced another strong power-function fit.Across sellers, seller rating had a small positive elasticity with respect to product price.
- Interpretation and boundary: The small elasticity indicates that purchasing decisions are likely influenced by variables beyond seller ratings, including the social network surrounding purchases.The authors also discuss higher seller costs as an alternative explanation for rating-related price premiums.
6. CONSUMER CHOICE PREDICTION
The paper models online seller selection as a ranking problem and finds that social-network and communication features substantially outperform simple price, rating, and random baselines. Performance is especially strong when buyers choose among many sellers.
- Problem and method: The task ranks up to 10 sellers offering the same product to predict which seller each buyer chooses.The data contain 9,950 buyer-seller clusters, restricted to clusters with 2–10 sellers.
- Problem and method: The model combines 23 product, buyer, and seller metadata features with buyer-seller interaction and network features.Prediction uses SVM-rank because each buyer has positive and negative seller examples that form sets.
- Overall performance: The all-feature model improves P@1 by 79% over Random, 13% over MostMsg, and 93% over MinPrice, while outperforming all three baselines across metrics.Communication and social-graph information remain important after discarding communication on and after the trade date.
- Overall performance: 42.1% P@1 is achieved with 10 sellers, and the improvement over MostMsg rises from 4.5% with 2 sellers to 39.5% with 10 sellers.All models decline as the number of sellers increases, but the full model’s advantage widens on harder choices.
- Feature analysis: Graph features perform only slightly below the full model, whereas metadata-only features remain close to Random, indicating that seller and product information alone is inadequate.The comparison supports incorporating social and trade networks around buyers and sellers.
- Feature analysis: Direct network features outperform indirect features, while message-network features are strongest and trade-network features slightly exceed contact-network features.Message volume indicates social relationships or prior queries, while trade volume indicates loyalty and trust.
- Feature analysis: Category-specific SVMs slightly improve aggregated performance, reaching P@1 of 0.58 versus 0.56 for the single Full SVM.The category models segment historical transactions by product category.
- Conclusion: The study concludes that social activity is the primary influence on consumer choice, with the social graph the strongest feature for predicting transaction partners.The result does not identify buyers’ precise reasoning, but it identifies social-network position as the most useful predictive signal.
7. CONCLUSION
Using one million Taobao users, the paper finds implicit information passing and shows that buyer communication drives purchasing activity. It also develops a model in which the social network is the most important feature for predicting transaction partners, while identifying future research directions.
- Main findings: The study analyzes one million Taobao users and empirically verifies implicit information passing through directed closure rules.It also examines how link formation depends on buyers’ and sellers’ network roles.
- Main findings: Buyer communication is identified as a fundamental driver of purchasing activity.This conclusion follows the paper’s analysis of information passing and directed triadic closure.
- Future directions: Future work includes browsing-based consumer-choice models, trust-aware information-passing analysis, and viral marketing in social commerce.The authors also encourage established e-commerce companies to add social-networking features.