Pinned it! A Large Scale Study of the Pinterest Network

Sudip Mittal, Neha Gupta, Prateek Dewan, Ponnurangam KumaraguruCODS 2014 · DOI

TL;DR

This paper presents a large-scale study of the Pinterest network, characterizing user behavior, content, and network structure. A key contribution is the development and evaluation of methods for predicting user gender on Pinterest, achieving a notable accuracy of 86.18% by combining profile, network, and content features with name-based analysis.

What Problem Does This Paper Address?

Online social networks (OSNs) have exploded in popularity, with image-based platforms like Pinterest gaining significant traction. Despite this growth, research on Pinterest has been limited. This paper aims to fill this gap by providing a comprehensive characterization of the Pinterest network, addressing fundamental questions about user behavior, content trends, geographical distribution, and the feasibility of inferring private attributes like gender. Understanding these aspects is crucial for platform developers, researchers, and businesses leveraging the platform.

Methods

The authors employed a multi-pronged approach:

  • Data Collection: A breadth-first search (BFS) crawler was developed to collect data from Pinterest. This involved an initial "snowball sampling" technique starting with the top 5 most followed users to gather a large set of user handles. This was followed by collecting detailed user profiles, pin information, board details, and image metadata.
  • User Characterization: Analysis of user profiles, including descriptions, connections (followers/followees), and linked social media accounts, was performed.
  • Content Analysis: Topical analysis of user-generated textual content (pin descriptions, user profiles, comments) was conducted using word frequency analysis and tag clouds.
  • Gender Prediction: Multiple machine learning algorithms (J48 Decision Tree, Naive Bayesian, Random Forest) were applied to predict user gender. They experimented with various feature sets, including Pinterest-specific features (profile, network, content) and a name-based feature.
  • Privacy Analysis: An initial exploration of privacy implications was conducted by searching for publicly shared Personally Identifiable Information (PII) like email addresses and phone numbers.

Datasets

  • User Handles Dataset: 17,964,574 unique user handles.
  • User Profile Dataset: 3,323,054 user profiles, containing details like display name, description, follower/followee counts, and linked social media accounts.
  • Boards Dataset: 777,748 boards with associated details like category and number of pins.
  • Pin Dataset: 58,896,156 unique pins, including descriptions, likes, comments, and repins.
  • Images Dataset: A random sample of 498,433 images from the pins.
  • Labeled Gender Dataset (for USA): 6,309 male users and 60,047 female users, with gender information extracted from linked Facebook profiles. This dataset was balanced for training by sampling.

Key Insights and Ideas

  • Dominance of "Fancy" Topics: Pinterest content (pins, boards, user descriptions) is heavily skewed towards topics like fashion, design, food, travel, and home decor, suggesting a focus on lifestyle and aesthetics.
  • Content Curation over Creation: A significant majority (over 95%) of images on Pinterest are pinned from pre-existing web sources, highlighting its role as a discovery and curation platform rather than a primary content creation hub.
  • Network Structure: The Pinterest network exhibits a comparable number of followers and followees for most users, with only a small fraction having highly skewed ratios. This suggests a relatively balanced social interaction pattern.
  • Gender Imbalance: The network is heavily dominated by female users, with a ratio of approximately 87% female to 7% male users in the analyzed data.
  • Name as a Strong Gender Predictor: Usernames, particularly first names, are a surprisingly strong indicator of gender, outperforming many Pinterest-specific features when used alone.
  • Privacy Concerns: Publicly available user descriptions can inadvertently reveal PII like email addresses and phone numbers, posing privacy risks.
  • Malware Presence: While Pinterest has measures against malicious content, a small percentage of pins were found to link to malware, though a significant portion of these were removed or no longer existed in subsequent crawls.

Key Results

  • Topical Analysis: The most common topics identified in user profiles and pin descriptions were design, fashion, photography, food, and travel.
  • Content Source: Users themselves are the largest contributors of original images (4.7%), followed by Google and Etsy as sources for pinned images.
  • Geographical Distribution: The USA, Canada, and the UK are the top countries for Pinterest users, with a significant concentration in major metropolitan cities.
  • Gender Prediction Accuracy:
    • Using only Pinterest-specific features (profile, network, content), a J48 Decision Tree classifier achieved an accuracy of 73.17%.
  • By adding the user's first name as a feature, the accuracy increased significantly to 86.18%.
  • Using only the name feature, an accuracy of 83.64% was achieved.
  • Feature Importance: The "Name" feature was the most informative for gender prediction, followed by board names and the number of pins. Follower and followee counts were the least informative.

Limitations and Future Work

  • Sampling Bias: The data collection started with the top 5 most followed users, which introduces a bias and means the sampled subgraph might not be fully representative of the entire Pinterest network.
  • Partial Crawl: Only partial subgraphs were crawled for the seed users and their immediate connections, limiting the scope of the network analysis.
  • Generalizing Gender Prediction: The current gender prediction is limited to USA users. Future work aims to generalize this to the entire global Pinterest population.
  • Expanding Feature Sets: The authors suggest incorporating more features for gender prediction, including natural language analysis of "about" fields and comments, and network centrality measures.
  • Image Spam and Copyright: Further analysis of image spam and copyright violations is planned.
  • Malware Persistence: Investigating the persistence of malware and malicious user profiles on the platform is a future direction.
  • Largest Connected Component (LCC) Analysis: The authors plan to re-analyze their data on the LCC of the Pinterest network for a more comprehensive view.

Summary updated

BibTeX
@inproceedings{mittal2014pinned,
  author = {Sudip Mittal and Neha Gupta and Prateek Dewan and Ponnurangam Kumaraguru},
  title = {{Pinned it! A Large Scale Study of the Pinterest Network}},
  booktitle = {CODS},
  year = {2014},
  doi = {10.1145/2567688.2567692},
}
Figure