"What's ur type?" Contextualized Classification of User Types in Marijuana-related Communications using Compositional Multiview Embedding

Ugur Kursuncu, Manas Gaur, Usha Lokala, Anurag Illendula, Krishnaprasad Thirunarayan, Raminta Daniulaityte, Amit Sheth, I. Budak ArpinararXiv 2018 · DOI · arXiv

  • Title: "What's ur type?" Contextualized Classification of User Types in Marijuana-related Communications using Compositional Multiview Embedding
  • Abstract: With 93% of pro-marijuana population in US favoring legalization of medical marijuana, high expectations of a greater return for Marijuana stocks, and public actively sharing information about medical, recreational and business aspects related to marijuana, it is no surprise that marijuana culture is thriving on Twitter. After the legalization of marijuana for recreational and medical purposes in 29 states, there has been a dramatic increase in the volume of drug-related communication on Twitter. Specifically, Twitter accounts have been established for promotional and informational purposes, some prominent among them being American Ganja, Medical Marijuana Exchange, and Cannabis Now. Identification and characterization of different user types can allow us to conduct more fine-grained spatiotemporal analysis to identify dominant or emerging topics in the echo chambers of marijuana-related communities on Twitter. In this research, we mainly focus on classifying Twitter accounts created and run by ordinary users, retailers, and informed agencies. Classifying user accounts by type can enable better capturing and highlighting of aspects such as trending topics, business profiling of marijuana companies, and state-specific marijuana policymaking. Furthermore, type-based analysis can provide more profound understanding and reliable assessment of the implications of marijuana-related communications. We developed a comprehensive approach to classifying users by their types on Twitter through contextualization of their marijuana-related conversations. We accomplished this using compositional multiview embedding synthesized from People, Content, and Network views achieving 8% improvement over the empirical baseline.

TL;DR

Researchers developed a "Compositional Multiview Embedding" (CME) approach to classify Twitter users in the marijuana space into three categories: Personal, Retail, and Informed Agency. By integrating text, emojis, profile images, and network interaction data, the model achieved a 96% F-score, significantly outperforming text-only baselines.

What Problem Does This Paper Address?

As marijuana legalization spreads across the U.S., social media has become a primary hub for discourse, marketing, and information sharing. However, analyzing this data is difficult because "marijuana Twitter" is not a monolith. A tweet from a casual user, a dispensary (retailer), or a news organization (informed agency) carries very different weight for policy analysts and public health researchers.

The core challenge is user classification in a highly imbalanced and noisy environment. Retailers and agencies are rare compared to individual users, and short-form text alone often lacks the context needed to distinguish between them. The authors aim to bridge this gap by looking beyond just the words in a tweet.

Methods

The authors propose a Compositional Multiview Embedding (CME) framework. Instead of relying on a single data source, they synthesize three "views" of a user:

  1. People View: Includes user descriptions, names (checked against a lexicon), and profile pictures. They used the Clarifai API to generate text tags for profile images, which were then converted into embeddings.
  2. Content View: Covers tweet text and emojis. They utilized EmojiNet to map emojis to their semantic meanings, ensuring that a "leaf" emoji is treated as a meaningful token rather than just a symbol.
  3. Network View: Captures user interactions (mentions and retweets). Because raw interaction matrices are massive and sparse, they applied truncated Singular Value Decomposition (t-SVD) to create dense, 300-dimensional network embeddings.
The overall architecture shows how People, Content, and Network views are pre-processed and combined for multi-class classification.

The "Compositional" aspect comes from using weighted vector addition to combine these disparate embeddings into a single, context-rich representation for each user.

This figure illustrates the vector addition process used to create CMEs for tweets, descriptions, and network data.

Datasets

The study utilized a massive corpus of 4.1 million tweets from over 1 million unique users, collected during the summer of 2017.

  • Training Set: A subset of 4,982 users was manually annotated by domain experts.
  • Class Distribution: The data was heavily skewed: 4,395 Personal accounts, 476 Informed Agencies, and only 111 Retail accounts.
  • Handling Imbalance: To prevent the classifier from simply ignoring the minority classes, the researchers used the SMOTE (Synthetic Minority Over-sampling Technique) algorithm.
A word cloud generated from Informed Agency tweets highlights common topics like "oil," "hash," and "cannabis."

Key Insights and Ideas

  • Visuals are Discriminative: One of the most significant findings was that including profile pictures drastically improved performance. Retailers often use logos, while individuals use selfies or scenery.
  • Network Homophily: Correlation analysis revealed that user descriptions and network interactions are highly correlated. This suggests that users on Twitter tend to interact with others who have similar profile characteristics (the "echo chamber" effect).
  • Emoji Semantics: Emojis are not just fluff; in the marijuana community, they serve as critical semantic markers. Integrating EmojiNet allowed the model to capture the "intent" behind pictorial communication.
Example profile pictures show clear visual distinctions between Personal (left), Retail (center), and Informed Agency (right) accounts.

Key Results

The full CME model achieved an Average F-score of 0.96, an 8% improvement over the empirical baseline.

  • Personal (P): 0.98 F-score.
  • Retail (R): 0.91 F-score.
  • Informed Agency (I): 0.73 F-score.

The inclusion of network features specifically boosted the F-score for the Informed Agency class by 6% and the Retail class by 4%, proving that who you talk to is just as important as what you say when identifying these specific roles.

Limitations and Future Work

The primary limitation remains the extreme data imbalance, particularly for the Retail class. While oversampling helped, the small sample size for retailers makes the model more sensitive to outliers in that category.

For future work, the authors plan to dive deeper into homophily analysis—specifically studying how these different user types form echo chambers and influence public opinion regarding marijuana legalization policies. They also intend to release their annotated dataset to the research community.

Summary updated

Abstract

With 93% of pro-marijuana population in US favoring legalization of medical marijuana, high expectations of a greater return for Marijuana stocks, and public actively sharing information about medical, recreational and business aspects related to marijuana, it is no surprise that marijuana culture is thriving on Twitter. After the legalization of marijuana for recreational and medical purposes in 29 states, there has been a dramatic increase in the volume of drug-related communication on Twitter. Specifically, Twitter accounts have been established for promotional and informational purposes, some prominent among them being American Ganja, Medical Marijuana Exchange, and Cannabis Now. Identification and characterization of different user types can allow us to conduct more fine-grained spatiotemporal analysis to identify dominant or emerging topics in the echo chambers of marijuana-related communities on Twitter. In this research, we mainly focus on classifying Twitter accounts created and run by ordinary users, retailers, and informed agencies. Classifying user accounts by type can enable better capturing and highlighting of aspects such as trending topics, business profiling of marijuana companies, and state-specific marijuana policymaking. Furthermore, type-based analysis can provide more profound understanding and reliable assessment of the implications of marijuana-related communications. We developed a comprehensive approach to classifying users by their types on Twitter through contextualization of their marijuana-related conversations. We accomplished this using compositional multiview embedding synthesized from People, Content, and Network views achieving 8% improvement over the empirical baseline.
BibTeX
@misc{kursuncu2018what,
  author = {Ugur Kursuncu and Manas Gaur and Usha Lokala and Anurag Illendula and Krishnaprasad Thirunarayan and Raminta Daniulaityte and Amit Sheth and I. Budak Arpinar},
  title = {{"What's ur type?" Contextualized Classification of User Types in Marijuana-related Communications using Compositional Multiview Embedding}},
  year = {2018},
  eprint = {1806.06813},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1806.06813},
}
Figure