Pinterest Board Recommendation for Twitter Users

Xitong Yang, Yuncheng Li, Jiebo Luo

Abstract

Pinboard on Pinterest is an emerging media to engage online social media users, on which users post online images for specific topics. Regardless of its significance, there is little previous work specifically to facilitate information discovery based on pinboards. This paper proposes a novel pinboard recommendation system for Twitter users. In order to associate contents from the two social media platforms, we propose to use MultiLabel classification to map Twitter user followees to pinboard topics and visual diversification to recommend pinboards given user interested topics. A preliminary experiment on a dataset with 2000 users validated our proposed system.

Introduction

When a user joins a new social media platform like Pinterest, the service faces a significant "cold start" problem. Without a history of "pins" or "likes," the platform struggles to provide relevant content. However, many new users possess a rich history on other platforms, such as Twitter.

The primary challenge in cross-platform recommendation is the modality gap. Twitter is predominantly text-based and conversational, whereas Pinterest is visual and organized into curated collections known as "pinboards." This research investigates bridging this gap by translating textual interests from Twitter into visually diverse pinboard recommendations on Pinterest.

System Architecture and Methodology

The proposed system utilizes a multi-stage pipeline to connect the two platforms, focusing on mapping textual features to a hierarchical topic ontology.

The proposed system architecture for cross-network pinboard recommendation.

1. Ontology Construction

A "Pinterest Ontology" was developed by pruning Wikipedia categories into 471 nodes, including 20 root categories such as "Fashion," "Food," and "DIY." This structure provides a map for potential user interests.

2. User Modeling

  • Twitter Side: Features are extracted from user timelines using standard text features (Feature Hashing and TF-IDF) and "Personality" features via LIWC (Linguistic Inquiry and Word Count) to capture psychological traits.
  • Pinterest Side: Existing users' boards are mapped to the ontology to create a multi-label representation of their interests.

3. Cross-Network Mapping

To predict Pinterest interests from Twitter data, the system employs Multi-label Classification. Specifically, it uses RAkEL (Random k-Labelsets), an ensemble method that considers subsets of labels together. This approach accounts for the correlation between interests (e.g., "Food" and "Recipes").

4. Visual Diversification

Once topics are predicted, the system applies visual diversification using a Convolutional Neural Network (AlexNet) to extract features from pins. An entropy-based ranking method ensures recommended boards are visually distinct, providing broader coverage of a topic rather than repetitive imagery.

Experimental Evaluation

The researchers curated a linked dataset of 2,265 users active on both platforms, including 4.9 million tweets and 25.5 thousand pinboards. The system's performance was evaluated using the F1-macro score.

Performance comparison of the RAkEL algorithm with varying subset sizes (k) against Binary Relevance (BR).

Key Results

  • Classification Accuracy: The fusion of text features and multi-label classification achieved an F1 macro-ex score of approximately 0.142. While seemingly low, this significantly outperforms random selection across 471 labels.
  • Label Dependencies: The study found that capturing relationships between categories (via RAkEL) improved performance compared to treating topics as isolated questions (Binary Relevance).
  • Feature Efficacy: Raw text features proved to be much better indicators of interest than LIWC personality features.

Conclusion

This research demonstrates that cross-network data can effectively mitigate the cold start problem. By leveraging multi-label classification to handle topic dependencies and visual entropy to ensure recommendation diversity, the system provides a more engaging discovery experience for new Pinterest users coming from Twitter. Future work will focus on modeling the temporal evolution of interests to provide more dynamic recommendations.

Summary updated

Abstract

Pinboard on Pinterest is an emerging media to engage online social media users, on which users post online images for specific topics. Regardless of its significance, there is little previous work specifically to facilitate information discovery based on pinboards. This paper proposes a novel pinboard recommendation system for Twitter users. In order to associate contents from the two social media platforms, we propose to use MultiLabel classification to map Twitter user followees to pinboard topics and visual diversification to recommend pinboards given user interested topics. A preliminary experiment on a dataset with 2000 users validated our proposed system.
BibTeX
@inproceedings{yang2015pinterest,
  author = {Xitong Yang and Yuncheng Li and Jiebo Luo},
  title = {{Pinterest Board Recommendation for Twitter Users}},
  booktitle = {ACM Multimedia},
  year = {2015},
  url = {https://openreview.net/forum?id=8vXeU_Kf5zO},
}
Figure