Learning consumer and producer embeddings for user-generated content recommendation

Wang-Cheng Kang, Julian J. McAuley

TL;DR

This paper introduces CPRec, a novel recommendation system for User-Generated Content (UGC) platforms that explicitly models users' dual roles as both consumers and producers. By learning separate "consumer" and "producer" embeddings for each user, derived from a shared core embedding, CPRec captures the nuanced ternary relationship between a consumer, an item, and its producer, leading to improved recommendation accuracy.

What Problem Does This Paper Address?

Traditional recommender systems, common in e-commerce or media streaming, typically assume a clear separation between content producers and consumers. However, User-Generated Content (UGC) platforms, like Reddit or Pinterest, operate differently. Here, users are often both creators and consumers of content, leading to a "closed loop" system. This duality means that a user's behavior as a consumer might be influenced by their own experiences as a producer, or their appreciation for another user's production. Existing recommendation methods, which primarily focus on the consumer-item interaction, fail to capture this richer, ternary relationship (consumer-item-producer), thus limiting their effectiveness on UGC platforms.

Methods

The core of CPRec lies in its innovative approach to user embeddings. Instead of a single embedding per user, CPRec learns a "core" embedding for each user and then uses two distinct transformation matrices to project this core embedding into two "role" embeddings: one for the user as a consumer and another for the user as a producer. This allows the model to capture asymmetric relationships, as a user might behave differently when consuming versus producing. The model then predicts a user-item score by considering the user's consumer-item preference and their "appreciation" for the item's producer. This is achieved by factoring the prediction into a consumer-item interaction term and a consumer-producer appreciation term, both leveraging the learned role embeddings. The model is trained using a Bayesian Personalized Ranking (BPR) framework to optimize for ranking observed items higher than unobserved ones.

Datasets

The authors evaluated CPRec on two large-scale, publicly available UGC datasets:

  • Pinterest: A content discovery platform where users pin, like, and repin images. The dataset used includes approximately 0.89 million users, 2.4 million images, and 56 million actions ('save' and 'like') from January 2013. Both 'like' and 'repin' actions were treated as implicit feedback.
  • Reddit: A discussion website where users submit content and comment. The dataset comprises data from March 2017, including 1.3 million users, 9.6 million submissions, and 48 million comments. Submissions were treated as items, and commenting actions as implicit feedback.

For evaluation, users were split into training, validation, and testing sets, with one action per user held out for validation and another for testing.

Key Insights and Ideas

  • The "Prosumer" Dilemma: The most significant insight is the critical need to model users' dual roles in UGC platforms. Ignoring this "prosumer" aspect leads to suboptimal recommendations.
  • Asymmetric Relationships Matter: The paper highlights that user relationships in UGC contexts can be asymmetric. A user's appreciation for a producer might not be reciprocated in the same way, and a single user embedding cannot capture these distinct behaviors.
  • Ternary Interactions are Key: Explicitly modeling the interaction between the consumer, the item, and the item's producer is crucial for improving recommendation quality in UGC environments.
  • Efficiency of Role Embeddings: Deriving role embeddings from a shared core embedding via projection matrices is an efficient way to introduce asymmetry and capture role-specific behaviors without a prohibitive increase in parameters, thus mitigating overfitting.

Key Results

CPRec demonstrated superior performance across both datasets compared to various baselines, including standard collaborative filtering (BPR) and methods that incorporate producer information (Factorization Machines, Vista).

  • Overall Improvement: CPRec achieved significant AUC improvements. On average, it provided a 13.9% AUC increase over BPR for all users and a 18.6% increase for cold users.
  • Outperforming Baselines: CPRec consistently outperformed FM and Vista, achieving an average improvement of 2.8% over the strongest baseline.
  • Robustness to Latent Dimensionality: The performance gap between CPRec and other methods widened as the latent dimensionality (K) increased, indicating CPRec's scalability and effectiveness with more complex models. Notably, CPRec achieved satisfactory performance even with a relatively small K (K=20).
  • Dataset Differences: The study observed better overall performance on Reddit than Pinterest, potentially due to Reddit users being more inclined to repeatedly follow the same producers, making their behavior more predictable.

Limitations and Future Work

  • Limitations: The paper notes that the Vista method was evaluated with a reduced model that only considered ownership information, as its full model also incorporates visual and temporal aspects. The authors also mention that the behance dataset was unsuitable due to its near-zero prosumer ratio, indicating that most users there had only one role.
  • Future Work: The authors propose several avenues for future research:
  • Incorporating more context information (e.g., temporal dynamics, explicit social networks).
  • Exploring additional side features to enrich the recommendation model.
  • Further investigating the problem of incorporating explicit social network information.

Summary updated

Abstract

User-Generated Content (UGC) is at the core of web applications where users can both produce and consume content. This differs from traditional e-Commerce domains where content producers and consumers are usually from two separate groups. In this work, we propose a method CPRec (consumer and producer based recommendation), for recommending content on UGC-based platforms. Specifically, we learn a core embedding for each user and two transformation matrices to project the user's core embedding into two 'role' embeddings (i.e., a producer and consumer role). We model each interaction by the ternary relation between the consumer, the consumed item, and its producer. Empirical studies on two large-scale UGC applications show that our method outperforms standard collaborative filtering methods as well as recent methods that model producer information via item features.
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