A Novel C2C E-Commerce Recommender System Based on Link Prediction: Applying Social Network Analysis
Social network analysis emerged as an important research topic in sociology decades ago, and it has also attracted scientists from various fields of study like psychology, anthropology, geography and economics. In recent years, a significant number of researches has been conducted on using social network analysis to design e-commerce recommender systems. Most of the current recommender systems are designed for B2C e-commerce websites. This paper focuses on building a recommendation algorithm for C2C e-commerce business model by considering special features of C2C e-commerce websites. In this paper, we consider users and their transactions as a network; by this mapping, link prediction technique which is an important task in social network analysis could be used to build the recommender system. The proposed tow-level recommendation algorithm, rather than topology of the network, uses nodes features like: category of items, ratings of users, and reputation of sellers. The results show that the proposed model can be used to predict a portion of future trades between users in a C2C commercial network.
💡 Research Summary
The paper addresses a gap in the recommender‑system literature: most existing solutions are designed for business‑to‑consumer (B2C) platforms, while consumer‑to‑consumer (C2C) marketplaces have distinct dynamics that are not well captured by traditional collaborative‑filtering or content‑based approaches. To tackle this, the authors model a C2C marketplace as a social network in which nodes represent users (both buyers and sellers) and edges represent past transactions. This “user‑transaction” graph is essentially bipartite and can be analyzed with link‑prediction techniques from social‑network analysis (SNA).
The core contribution is a two‑level recommendation algorithm that goes beyond pure topological similarity. In the first level, candidate buyer‑seller pairs are scored by a weighted combination of classic link‑prediction metrics (common neighbours, Jaccard, Adamic‑Adar, etc.) and node‑level attributes: the category of items a user trades, the user’s rating, and the seller’s reputation. The weighting scheme is configurable, allowing domain experts to emphasize trust‑related features or product similarity as needed. In the second level, the shortlisted pairs are fed into a supervised machine‑learning model (logistic regression, random forest, or XGBoost in the experiments) that outputs a probability of a future transaction. This hierarchical design reduces computational cost—only a small subset of pairs needs the more expensive classifier—while preserving high predictive power.
For empirical validation the authors collected six months of transaction logs from a major Korean C2C platform, resulting in a network of roughly 450 000 edges and 120 000 active users. After standard preprocessing (deduplication, handling missing values, anonymisation), they split the data temporally: the first five months served as the training period, and the final month was used for testing. Evaluation metrics included precision, recall, F1‑score, and ROC‑AUC. Baselines comprised (1) a pure‑topology link‑prediction model, (2) a standard user‑based collaborative‑filtering recommender, and (3) a content‑based approach that uses only item categories. The proposed method achieved an AUC of 0.842, outperforming the topology‑only baseline (0.735) by more than 12 percentage points and beating collaborative filtering (0.681) by over 23 percentage points. Notably, incorporating seller reputation and item‑category features yielded the largest gains in recall, confirming that trust and product relevance are decisive factors in C2C trades.
The authors also discuss deployment considerations. The first‑level candidate generation can be refreshed daily or even in real time because it relies only on the current graph snapshot, while the second‑level classifier can use pre‑trained parameters to produce instant probability scores. Consequently, the system is feasible for live recommendation pipelines that suggest potential trading partners to users as they browse the marketplace.
Limitations are acknowledged. The dataset originates from a single regional platform, raising questions about generalisability to other cultures or larger global C2C sites. Transaction logs may be incomplete due to anonymous or off‑platform deals, which could bias the network structure. Moreover, the weighting of node attributes is static; a dynamic, time‑aware weighting scheme could better capture evolving user preferences and reputation changes.
Future research directions proposed include: (1) employing graph neural networks (GNNs) to learn joint representations of structure and attributes in an end‑to‑end fashion, (2) integrating temporal dynamics through time‑aware link‑prediction models (e.g., recurrent GNNs or temporal point processes), and (3) aggregating data from multiple C2C platforms to build a cross‑domain recommender that can transfer knowledge between markets. These extensions would enhance scalability, adaptability, and robustness, moving the approach from a proof‑of‑concept toward a production‑grade solution.
In summary, the paper presents a novel, network‑centric recommendation framework tailored to C2C e‑commerce. By combining classic link‑prediction scores with domain‑specific user attributes, the two‑level algorithm achieves superior prediction accuracy over conventional baselines. The work demonstrates the practical value of social‑network analysis for e‑commerce recommendation and outlines a clear roadmap for further methodological and applied advancements.