Learning consumer and producer embeddings for user-generated content recommendation

Wang-Cheng Kang, Julian J. McAuley

TL;DR

In User-Generated Content (UGC) platforms like Reddit or Pinterest, users aren't just consumers—they are also the producers. This paper introduces CPRec, a recommendation model that captures this "prosumer" dynamic by learning dual embeddings (consumer and producer roles) for every user, significantly improving recommendation accuracy, especially for "cold-start" users with few interactions.

The Prosumer Challenge in UGC

Traditional recommendation systems, such as those used by Netflix or Amazon, typically operate in a centralized domain where producers and consumers are distinct groups. Models like Matrix Factorization focus almost exclusively on the relationship between a consumer and an item.

Comparison between Professionally Generated Content (PGC) and User-Generated Content (UGC) ecosystems.

However, in UGC communities, the loop is closed: users create the content that other users consume. The authors argue that a user’s decision to engage with a post is driven not just by the content itself, but also by their "appreciation" for the producer of that content. Standard models often ignore this ternary relationship (Consumer-Item-Producer) or treat the producer as a simple categorical feature, failing to capture the complex, asymmetric "follow" relationships inherent in social platforms.

Methodology: The CPRec Model

The authors propose CPRec (Consumer and Producer based Recommendation). The core innovation lies in how they represent users through dual roles:

  1. Core Embeddings: Each user $u$ is assigned a single "core" latent embedding ($\gamma_u$) representing their global interests.
  2. Role Projection: To account for different behaviors when creating vs. consuming, the model uses two transformation matrices ($W^c$ and $W^p$) to project the core embedding into two specialized roles:
  3. $\gamma^c_u$ (Consumer role): How the user acts when looking for content.
  4. $\gamma^p_u$ (Producer role): The "vibe" or style of the content the user creates.
  5. Ternary Interaction Score: The predicted preference score for a user $u$ toward item $i$ (produced by user $p_i$) is calculated as the sum of two parts:
  6. Consumer-Item Preference: The standard inner product between the consumer role and the item embedding.
  7. Consumer-Producer Appreciation: An inner product between the consumer role of the user and the producer role of the creator.
  8. Learning: The model is trained using the Bayesian Personalized Ranking (BPR) framework, which optimizes for ranking observed interactions higher than non-observed ones.

Datasets and Evaluation

The model was evaluated on two large-scale UGC datasets:

  • Pinterest: A dataset of "pins" and "likes" from January 2013. It features a high "prosumer ratio" (74.42%), meaning most users both upload and like content.
  • Reddit: A crawl of submissions and comments from March 2017. This dataset has an even higher prosumer ratio (86.85%) and shows a stronger tendency for users to repeatedly consume content from the same producers.

Key Results

CPRec consistently outperformed standard Matrix Factorization (BPR), Factorization Machines (FM), and the state-of-the-art artistic recommendation model (Vista) across both datasets.

Performance comparison of CPRec against baselines on Pinterest and Reddit datasets.
  • Significant Gains: On average, CPRec achieved a 13.9% AUC improvement over standard BPR for all users.
  • Cold-Start Resilience: The improvement was even more pronounced for cold-start users (fewer than five interactions), with an 18.6% AUC gain over BPR. By linking roles to a single core embedding, the model shares information: knowing what a user produces helps predict what they will consume.
  • Platform Differences: The model performed better on Reddit than Pinterest. The authors attribute this to Reddit users being more inclined to "repeatedly follow" specific producers, making the producer-consumer signal stronger.

Conclusion and Future Work

CPRec demonstrates that modeling the dual nature of users in UGC platforms is crucial for accurate recommendations. By accounting for the asymmetric "appreciation" between consumers and producers, the model provides a more nuanced understanding of user behavior.

Future work includes incorporating explicit social signals (actual "follow" graphs), temporal dynamics, and content-based features (like image or text analysis) to further refine the recommendation engine.

Summary updated

BibTeX
@inproceedings{kang2018learning,
  author = {Wang-Cheng Kang and Julian J. McAuley},
  title = {{Learning consumer and producer embeddings for user-generated content recommendation}},
  booktitle = {RecSys},
  year = {2018},
  doi = {10.1145/3240323.3240394},
}
Figure