[Paper Review] Twitter Hashtag Recommendation using Matrix Factorization

Hamidreza Alvari|arXiv (Cornell University)|May 30, 2017

Recommender Systems and Techniques20 references3 citations

TL;DR

This paper proposes a hashtag recommendation system using weighted matrix factorization (hWMF) that recommends hashtags based solely on users' historical hashtag usage, without relying on tweet content. By modeling hashtag correlations and using alternating least squares for optimization, the method achieves superior performance over baselines in recommending relevant, user-consistent hashtags.

ABSTRACT

Twitter, one of the biggest and most popular microblogging Websites, has evolved into a powerful communication platform which allows millions of active users to generate huge volume of microposts and queries on a daily basis. To accommodate effective categorization and easy search, users are allowed to make use of hashtags, keywords or phrases prefixed by hash character, to categorize and summarize their posts. However, valid hashtags are not restricted and thus are created in a free and heterogeneous style, increasing difficulty of the task of tweet categorization. In this paper, we propose a low-rank weighted matrix factorization based method to recommend hashtags to the users solely based on their hashtag usage history and independent from their tweets' contents. We confirm using two-sample t-test that users are more likely to adopt new hashtags similar to the ones they have previously adopted. In particular, we formulate the problem of hashtag recommendation into an optimization problem and incorporate hashtag correlation weight matrix into it to account for the similarity between different hashtags. We finally leverage widely used matrix factorization from recommender systems to solve the optimization problem by capturing the latent factors of users and hashtags. Empirical experiments demonstrate that our method is capable to properly recommend hashtags.

Motivation & Objective

• Address the challenge of recommending relevant hashtags on Twitter despite the lack of public tweet content due to privacy concerns.
• Overcome the limitations of content-based methods that require NLP and suffer from computational overhead.
• Formulate hashtag recommendation as an optimization problem incorporating user hashtag usage history and hashtag correlations.
• Develop a lightweight, content-free recommendation model that leverages user behavior patterns and hashtag co-occurrence.
• Demonstrate that users consistently adopt similar hashtags, validating the use of historical usage for effective recommendations.

Proposed method

• Model the user-hashtag interaction as a sparse, low-rank matrix X ∈ ℝ₊^{N×M}, where x_ij indicates if user i used hashtag j.
• Formulate the hashtag recommendation as an optimization problem that minimizes reconstruction error between observed and predicted user-hashtag interactions.
• Incorporate a weighted matrix factorization (hWMF) that assigns higher weights to training samples based on hashtag co-occurrence frequency.
• Use a hashtag correlation weight matrix W to reflect the similarity between hashtags, computed from their joint usage in tweets.
• Apply an alternating least squares (ALS) algorithm to iteratively optimize user and hashtag latent factor matrices.
• Learn low-dimensional latent factors for users and hashtags to capture underlying preferences and topical relationships without using tweet text.

Experimental results

Research questions

• RQ1 Do users tend to adopt similar hashtags over time, indicating consistent usage patterns?
• RQ2 Can hashtag co-occurrence patterns be effectively used to improve recommendation accuracy?
• RQ3 How does a content-free, history-based matrix factorization approach compare to content-dependent methods in hashtag recommendation?
• RQ4 To what extent does incorporating hashtag correlation enhance recommendation performance?
• RQ5 Can a low-rank matrix factorization model trained on user hashtag history alone achieve competitive results without relying on NLP or tweet similarity?

Key findings

• A two-sample t-test confirms that users are significantly more likely to adopt similar hashtags, validating the consistency of their hashtag usage behavior.
• The proposed hWMF model outperforms baseline methods in recommending accurate hashtags, demonstrating the effectiveness of incorporating hashtag correlation.
• The method achieves strong performance without requiring tweet content, making it suitable for private or non-public user timelines.
• Incorporating hashtag co-occurrence as a weight matrix improves model generalization and recommendation precision.
• The model is computationally efficient and scalable, avoiding the need for complex NLP pipelines or similarity computations.
• Empirical evaluation on a curated dataset of tweets with trending hashtags confirms the model's ability to recommend relevant hashtags aligned with user history.