💻 computer science

Examining hate speech infection in social media using range constraints in multidimensional data structures

This paper proposes a methodology that combines shapelet-based time series analysis with multidimensional data structures and range constraints to precisely identify the temporal windows and user connectivity patterns driving hate speech propagation on social media, thereby enabling more effective event detection and mitigation.

Original authors: Eleanna Kafeza, Christos Makris, Gerasimos Rompolas, Georgios Mavris

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eleanna Kafeza, Christos Makris, Gerasimos Rompolas, Georgios Mavris

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine social media as a giant, bustling party where everyone is shouting their thoughts. Sometimes, the vibe turns sour, and a specific kind of "toxic infection" (hate speech) starts spreading. The authors of this paper, Eleanna Kafeza and her team, wanted to figure out exactly when this infection starts and who is spreading it, rather than just trying to catch every single bad word as it flies by.

Think of a user's timeline not as a list of posts, but as a heartbeat monitor. Instead of looking at the words themselves, the team turned every user's emotional intensity into a squiggly line on a graph. If a user is posting angry, hateful things, their "heartbeat" spikes with negative energy.

The Detective's Toolkit: Shapelets and Suffix Trees
To find the bad actors, the researchers used a clever trick called shapelet classification. Imagine you are looking for a specific fingerprint in a pile of muddy footprints. A "shapelet" is like a tiny, distinctive chunk of that footprint that proves who made it. In this case, the "fingerprint" is a specific pattern of emotional spikes that only hate-speech users seem to have.

But checking every single footstep against every single fingerprint would take forever. So, the team built a super-fast library called a Suffix Tree. Think of this like a magical, organized filing cabinet where every possible pattern of words is sorted instantly. Instead of reading every book in the library to find a story, you just ask the cabinet, "Where is the story that starts with 'angry'?" and it points you right to the shelf. This made their search incredibly fast, turning a task that usually takes ages into something that happens in the blink of an eye.

The Two-Step Filter: Time and Influence
Finding the pattern was just step one. The real magic happened when they added two extra filters, like a bouncer checking two IDs at the door: Time and Influence.

The Time Window: The team realized that hate speech often explodes right after a specific event, like a news story breaking. They didn't just look for the bad pattern; they asked, "Did this pattern happen right now?" By zooming in on specific time windows (like a 13-hour window), they could pinpoint exactly when the "infection" started spreading.
The Influence Check: They also checked how "popular" or "connected" the user was. They found something surprising: the hate speech wasn't always coming from the biggest celebrities. In fact, they suspected that influential users get banned quickly, so the infection often spreads through regular, less-connected users. They used a math tool called a 2D Range Tree to check if a user's posts happened at the right time and if their level of influence fit a certain range.

What the Numbers Say
The team tested this on a massive dataset of tweets collected over one month (from February 1, 2022, to March 1, 2022). They looked at 3,912 users who posted hate speech and balanced it with a dataset of 46,557 tweets from 7,824 distinct users.

When they ran their experiments, the results were promising but specific to their setup:

Accuracy: Their best setup found the right hate-speech users with a mean accuracy of 0.72 and a mean F1-score of 0.84.
Speed: Their new method was wildly faster than the old, "naive" way of doing things. While the old method took about 0.12 × 10⁻² seconds for a small test, their new method did it in 2.15 × 10⁻⁶ seconds. As the data grew to 23,800 items, their method stayed fast (around 2.97 × 10⁻⁶ seconds), while the old method slowed down significantly.
The Sweet Spot: They found that the hate speech infection was most visible in a specific time window of [0, 13] (hours) and among users in the [0, 17.5] percentile of influence. This suggests the infection often starts with a small group of users who aren't necessarily the most famous, but are very active during a specific crisis.

What They Don't Claim
It's important to note what this paper doesn't say. The authors are careful to point out that they aren't claiming to have "solved" hate speech forever. They explicitly argue against relying only on simple word lists or just looking at the text without context. They also note that their method works best when you have specific conditions (like a known event time) to filter by; it's not a magic wand that finds hate speech in a vacuum without any extra data.

The Big Picture
By turning emotional chaos into organized data shapes and using a super-fast library to find them, the authors showed that we can catch the "infection" of hate speech more precisely. They found that by looking at when the anger spikes and who is spreading it, we can identify the source of the problem much faster than before. While they suggest this could help social media platforms take immediate action, they present this as a powerful new tool in the toolbox, not a final cure-all. The goal is to spot the spark before the whole forest burns down.

Technical Summary: Examining Hate Speech Infection in Social Media Using Range Constraints in Multidimensional Data Structures

Problem Statement
The paper addresses the challenge of detecting and analyzing the "infection" of hate speech on social media platforms like Twitter. The authors argue that hate speech is not merely a static classification problem of individual posts but a dynamic phenomenon driven by emotionally intense events and propagated through specific user networks. Existing literature often focuses on classifying hate messages in isolation or relies on deep learning models (e.g., BERT) that, while effective, can be difficult to interpret and may overfit data. Furthermore, traditional approaches often fail to integrate critical dimensions such as the temporal window of an event and the social influence of the users involved. The core problem is to identify not just who posts hate speech, but when and under what network conditions (e.g., influence levels) this behavior occurs to facilitate immediate mitigation.

Methodology
The proposed methodology treats hate speech detection as an integrated problem of "infection," modeling user behavior through time series analysis and utilizing efficient data structures for pattern matching and range querying.

Time Series Modeling of Emotion:
- User behavior is modeled as a time series representing the intensity of negative emotions over time.
- The NRC-Canada Emotion Intensity Lexicon (NRC-EIL) is used to identify negative words (anger, disgust, fear, sadness) and assign them real-valued intensity scores.
- For each user, a time series is constructed where each time frame's value is the sum of weights of negative words in their posts during that interval.
- To handle spelling variations, the Levenshtein distance is employed as a dissimilarity metric.
Symbolic Transformation and Dimensionality Reduction:
- The continuous time series are transformed into symbolic time series using Symbolic Aggregate approXimation (SAX).
- The process involves z-Normalization, Piece-wise Aggregation Approximation (PAA) to segment the series, and mapping segments to an alphabet (e.g., $\Sigma = \{a, b, c, d, e\}$ ) based on breakpoints. This reduces dimensionality while preserving the general shape of the emotional intensity curve.
Shapelet Classification:
- The study employs shapelet-based classification, where "shapelets" are discriminative subsequences of the time series that act as fingerprints for hate speech users.
- Instead of Euclidean distance (unsuitable for symbolic data), the authors use Alphabet-weight Edit Distance. This metric assigns weights to character substitutions based on their proximity in the alphabet, reflecting the intuition that similar emotional intensities are closer than distant ones.
- A Generalized Suffix Tree is constructed from the symbolic time series of all users to enable efficient searching for these shapelets.
Multidimensional Range Constraints:
- To refine the identification of hate speech users, the framework introduces post-processing constraints based on time and user influence.
- Time Constraint: The system searches for shapelet occurrences within specific time intervals (e.g., during a specific event).
- Influence Constraint: User influence is quantified using weighted degree centrality derived from the retweet graph. The model hypothesizes that less influential individual users may be more prone to posting hate speech (as influential users are often banned quickly).
- Data Structures:
  - A Balanced Binary Tree is integrated with the suffix tree (via Euler tour traversal) to handle 1D range queries on time efficiently.
  - A 2D Range Tree is added to handle simultaneous queries on time and influence.
- This results in a hybrid approach: a shapelet classifier identifies potential candidates, and the multidimensional data structures filter them based on temporal and influence constraints.

Key Contributions

Time Series Representation: The paper introduces the use of symbolic time series to represent the emotional intensity of users' social media presence for hate speech detection.
Shapelet Classification with Suffix Trees: It proposes a novel classification method that combines shapelet discovery with generalized suffix trees for fast retrieval, utilizing Alphabet-weight Edit Distance for symbolic series.
Multidimensional Constraint Framework: The authors introduce a framework that extends classification with post-processing search using multidimensional data structures (suffix trees, balanced binary trees, and range trees). This allows for the simultaneous filtering of users based on temporal windows and network influence metrics.
Interpretability: Unlike "black box" deep learning models, this approach provides interpretable "fingerprints" (shapelets) and allows for the exploration of specific time windows and user influence ranges that trigger hate speech diffusion.

Experimental Results

Dataset: The study utilized a dataset of tweets containing the keyword "@CNN" from February 2022, resulting in a balanced set of 46,557 tweets from 7,824 users.
Baseline Comparison: The proposed models were compared against naive shapelet classifiers (NSC) and various deep learning baselines (BERT-base-uncased, fBERT, HateBERT).
Performance:
- Accuracy: The optimal configuration (shapelet length 8, 3 shapelets, edit distance 2) achieved a mean accuracy of 0.72 and a mean F1-score of 0.84.
- Scalability: The proposed data structures (SSC, SSCAB, SSCAR) demonstrated significantly lower execution times compared to naive methods as dataset size increased. For instance, while the naive approach scaled linearly with data size, the suffix tree-based approaches maintained near-constant or logarithmic query times ( $O(n + \log m + k)$ ).
- Dimensional Analysis: The experiments identified that hate speech diffusion was most concentrated in specific time windows (e.g., [0, 13] hours) and among users in the lower percentiles of influence (0–17.5%), supporting the hypothesis that individual, less influential users drive the infection.
Visualization: Word cloud analysis of the identified hate speech users revealed topics related to politics, vaccines, and the Russia-Ukraine war, confirming the model's ability to capture contextually relevant negative discourse.

Significance and Claims
The paper claims that its significance lies in providing a hybrid approach that combines the pattern recognition power of shapelet classification with the efficiency of advanced data structures for multidimensional querying.

Efficiency: The use of suffix trees and range trees allows for the rapid identification of hate speech users that adhere to complex conditions (time and influence), which is computationally superior to brute-force methods.
Actionable Insights: By isolating the specific time windows and user influence levels associated with hate speech "infection," the methodology enables social media operators to take targeted mitigation actions (e.g., identifying the triggering event and the specific user groups involved) rather than relying solely on broad content filtering.
Interpretability: The approach offers a transparent mechanism for understanding why a user is flagged (via the specific shapelet pattern) and when the behavior occurred, addressing the "black box" limitations of deep learning models.
Extensibility: The framework is designed to be easily expanded to include additional dimensions (e.g., other network metrics or content features) without fundamentally altering the core classification logic.

The authors maintain a modest tone, acknowledging that while the results are promising, the approach relies on specific parameter tuning and that future work could explore more complex indexing structures (like property suffix trees) to further optimize space and time performance.

More like this