🤖 machine learning

Unsupervised Anomaly Detection of Information Operations Users via Behavioral and Language Patterns

This paper introduces TENSOR, a novel unsupervised framework that detects information operations users by modeling their coordinated temporal behaviors via Temporal Point Processes and refining these signals with quantitative scores derived from Large Language Model analysis of their textual content, thereby outperforming existing methods on real-world datasets.

Original authors: Sishun Liu, Sajal Halder, Ke Deng, Yan Wang, Xiuzhen Zhang

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sishun Liu, Sajal Halder, Ke Deng, Yan Wang, Xiuzhen Zhang

Original paper licensed under CC BY 4.0 (http://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 the internet as a massive, bustling town square where millions of people are chatting, sharing news, and arguing about politics. Usually, most people are just regular folks (let's call them "Control Users") living their lives. But sometimes, a small, secretive group of actors (called "Information Operation" or "IO" users) enters the square. They aren't there to chat; they are there to manipulate the conversation, spread lies, and stir up trouble, often working together like a well-rehearsed choir.

The problem is that these manipulators are very good at hiding. They try to look like normal people, but they have a secret rhythm and a specific way of talking that gives them away.

The Old Ways of Finding Them (And Why They Failed)

Previously, detectives tried to catch these manipulators in two ways:

The "Look for the List" Method (Supervised Learning): They trained computers using a list of known bad actors. But as soon as the bad actors changed their tactics (which they do very quickly), the computer got confused and couldn't spot the new tricks.
The "They Move Together" Method (Unsupervised Clustering): They assumed that bad actors always act in perfect unison, like a marching band. If two people posted the exact same thing at the exact same time, they were flagged. But in reality, bad actors are smarter than that; they don't always march in lockstep, so this method missed many of them.

The New Solution: TENSOR

The authors of this paper built a new detective tool called TENSOR. Think of TENSOR as a two-part security system that looks at two things simultaneously: when people post and what they say.

Part 1: The Rhythm Watcher (Temporal Point Process)

Imagine a security guard watching the town square who is obsessed with the timing of people's movements.

Normal people post at random times: maybe in the morning, maybe at lunch, maybe late at night. Their rhythm is messy and natural.
Bad actors often have a weird, coordinated rhythm. They might post too fast, or they might all post at the exact same second because they are following a script.

TENSOR uses a mathematical tool (called a Temporal Point Process) to learn the "heartbeat" of the square. It learns what a normal rhythm looks like. When it sees a rhythm that is too perfect or too strange, it raises an alarm.

The Catch: The security guard has a problem. The training data (the list of people the guard studied) is "contaminated." It's like trying to teach a guard what "normal" looks like, but the class includes a few bad actors who are pretending to be normal. The guard might accidentally learn the bad actors' weird rhythm as "normal."

Part 2: The Language Detective (The LLM)

To fix the guard's mistake, TENSOR brings in a second expert: a Large Language Model (LLM). Think of this as a super-smart literary critic who has read millions of books and knows the difference between a genuine human voice and a robotic script.

The LLM reads the actual words people post. It doesn't just look at the timing; it looks at the style.

Does this person talk about their cat, their job, and politics? (Likely a Control User).
Does this person only talk about one specific political topic, using strange, repetitive phrases, and ignoring everything else? (Likely an IO User).

Putting It Together: The "Evidence Function"

This is the magic sauce. TENSOR doesn't just let the two experts vote; it uses a special formula (called an Evidence Function) to combine their opinions.

The Rhythm Watcher says, "This person's timing is suspicious."
The Language Detective says, "This person's words sound like a script."
The Evidence Function takes the Language Detective's opinion and uses it to correct the Rhythm Watcher.

If the Rhythm Watcher was confused because it learned from bad actors in the training data, the Language Detective steps in and says, "Wait, look at the words. This is definitely a bad actor, even if the timing looked a bit normal." This allows TENSOR to spot the bad actors even when the training data was "dirty."

What They Found

The researchers tested TENSOR on real-world data from five different countries (Egypt, China, Iran, Russia, and the UAE). They compared it against other methods and found:

TENSOR was the best: It caught more bad actors and made fewer mistakes than the old methods.
Both parts are needed: If they removed the "Language Detective" (the LLM) or the "Rhythm Watcher" (the timing analysis), the system got much worse. It needs both the when and the what to work.
It works with different brains: They tested TENSOR with different AI models (like different types of "Language Detectives"), and it worked well with almost all of them.

The Bottom Line

TENSOR is a new way to find internet manipulators without needing a pre-made list of who they are. It works by noticing that bad actors have a weird "dance" (timing) and a fake "voice" (language). By using an AI to check the voice and correct the dance analysis, it can spot the manipulators even when they are trying to blend in with the crowd.

Technical Summary: TENSOR for Unsupervised IO User Detection

Problem Definition

The paper addresses the challenge of detecting Information Operation (IO) users on Social Networks (SNs). IOs are organized attempts to manipulate public opinion through disinformation and hate speech, often initiated by small groups of coordinated users. The primary difficulty lies in the generalization capability of existing detection methods:

Supervised and semi-supervised approaches rely on labeled datasets that lag behind the rapidly evolving nature of IO campaigns, making them ineffective against unseen IOs.
Existing unsupervised approaches often rely on oversimplified assumptions of coordination (e.g., clustering based on similarity) that may not hold in practice or fail to capture complex temporal dynamics.
Data Contamination: A critical issue is that training data for anomaly detection is often "contaminated," meaning the dataset intended to represent "normal" (control) users inevitably contains a small fraction of IO users. Training models on such data can lead to the model treating anomalies as normal, degrading performance.

The authors formulate IO user detection as an unsupervised anomaly detection problem, aiming to identify IO users (labeled 0) from control users (labeled 1) using only their post timelines, which consist of timestamps ( $t$ ) and content ( $c$ ).

Methodology: TENSOR

The proposed solution is Temporal-bEhavior-laNguage Signals for information Operation Recognition (TENSOR). It is a multimodal, unsupervised framework that combines behavioral patterns (temporal) and language patterns (textual) to detect anomalies. The architecture consists of two core components integrated via an adjustment framework:

1. Temporal Point Process (TPP) for Behavioral Patterns

Function: The TPP models the user's posting timeline as a stochastic process. It captures the conditional intensity function $\lambda^*(t)$ , representing the probability of an event occurring at time $t$ given the history.
Training: A TPP model (specifically the Self-attentive Hawkes Process, SAHP) is trained on the contaminated dataset $T$ (containing both control and IO users).
Mechanism: IO users often exhibit coordinated behaviors (e.g., posting at specific intervals or flooding content) that differ from organic users. The TPP calculates a score based on the negative log-likelihood of the observed timestamps: $f(x_i) = \exp(-\frac{1}{M_i} \log p(t_i))$ .
Limitation: Because the training data is contaminated, the TPP model may learn the IO patterns as part of the "normal" distribution, reducing its ability to distinguish anomalies.

2. Large Language Model (LLM) for Language Patterns

Function: An LLM is used to analyze the textual content of user posts ( $c_i$ ) alongside the timeline to generate an "evidence" score.
Prompting: The LLM is prompted with a zero-shot system instruction to act as a social media analyst, distinguishing between "IO accounts" (coordinated, state-sponsored) and "Control accounts" (organic).
Evidence Function: The LLM's raw text output is converted into a quantitative score $T(t_i, c_i)$ using a semantic similarity mapping strategy. The output is compared against reference texts ("IO account" vs. "Control account") to generate a probability distribution. To ensure robustness against LLM non-determinism, $N$ samples are drawn and averaged to produce $\bar{T}_N(t_i, c_i)$ .

3. Evidence-based Post-Hoc Adjustment (EPHAD)

To address the training set contamination problem, TENSOR integrates the TPP and LLM components using the Evidence-based Post-Hoc Adjustment Framework for Anomaly Detection (EPHAD).

Instead of retraining the TPP, the framework adjusts the TPP's output at test time using the LLM-derived evidence.
The final decision score is calculated by exponentially tilting the TPP score with the LLM evidence:
$\hat{f}(x) = f(x) \exp(\bar{T}_N(t_i, c_i) / \beta)$
where $\beta$ is a temperature parameter.
This allows the system to leverage the LLM's understanding of language patterns to correct the biases introduced by the contaminated training data in the TPP.

Key Contributions

Novel Formulation: The paper frames IO user detection as an unsupervised anomaly detection problem, moving away from the limitations of supervised methods that require lagging labeled data.
Multimodal Integration: TENSOR is the first approach to jointly leverage temporal behavioral patterns (via TPP) and language patterns (via LLM) for unsupervised IO detection.
Mitigation of Data Contamination: The paper introduces a novel application of the EPHAD framework to the specific domain of IO detection. It effectively mitigates the performance degradation caused by training on contaminated datasets (where IO users are present in the training set) by using an LLM-based evidence function to adjust the TPP inference.
Zero-Shot Capability: The method does not require labeled IO data for training the LLM component, relying instead on zero-shot prompting and semantic similarity.

Experimental Results

The authors evaluated TENSOR on five real-world IO datasets (Egypt, China_1, Iran_1, Russia_1, UAE) covering 26 labeled IOs across 15 countries. The datasets are highly imbalanced (few IO users, many control users).

Performance: TENSOR outperformed three baselines—User Embedding Clustering, Behavioral Languages for Online Characterization (BLOC), and a pure LLM-based zero-shot detector—across most metrics (Precision, Recall, F1, AUC, and AUPRC).
Ablation Studies:
- Removing the LLM evidence function significantly reduced performance, confirming the necessity of language patterns.
- Removing the TPP (using only LLM) resulted in poor performance, confirming that temporal dynamics are crucial.
- Using a random evidence function performed worse than the standard TPP, proving the LLM provides genuine signal.
Sensitivity Analysis:
- LLM Choice: TENSOR remained stable and effective across different LLMs (including reasoning models like gpt-oss-120B and non-reasoning models like llama3.3-70B), though larger reasoning models generally yielded slightly better results.
- Temperature ( $\beta$ ): The performance was sensitive to the temperature parameter $\beta$ . While the default $\beta=0.5$ was effective, the authors noted that optimal values vary by dataset (often around 0.3), suggesting potential for future unsupervised optimization of this hyperparameter.

Significance and Claims

The paper claims that TENSOR offers a resilient alternative to traditional supervised models by shifting the focus to the inherent friction between coordinated IO campaigns and organic user behaviors. Its primary significance lies in:

Overcoming the Generalization Gap: By being unsupervised, it can detect new, unseen IOs without requiring retraining on new labeled datasets.
Solving the Contamination Problem: It provides a practical mechanism to train anomaly detectors on real-world data (which is inevitably contaminated) without the model learning the anomalies as normal.
Practicality: The approach is "training-free" regarding the LLM (using frozen weights) and relies on a single TPP training step, making it adaptable to various existing TPP and LLM architectures.

The authors conclude that while unsupervised optimization of the temperature parameter $\beta$ remains a future direction, TENSOR currently establishes a new state-of-the-art baseline for unsupervised IO user detection in real-world, imbalanced scenarios.