Fred Morstatter

TMLR Journal 2025 Journal Article

Offset Unlearning for Large Language Models

• James Y. Huang
• Wenxuan Zhou
• Fei Wang
• Fred Morstatter
• Sheng Zhang
• Hoifung Poon
• Muhao Chen

Despite the strong capabilities of Large Language Models (LLMs) to acquire knowledge from their training corpora, the memorization of sensitive information in the corpora such as copyrighted, biased, and private content has led to ethical and legal concerns. In response to these challenges, unlearning has emerged as a potential remedy for LLMs affected by problematic training data. However, previous unlearning techniques are either not applicable to black-box LLMs due to required access to model internal weights, or violate data protection principles by retaining sensitive data for inference-time correction. We propose $\delta$-unlearning, an offset unlearning framework for black-box LLMs. Instead of tuning the black-box LLM itself, $\delta$-unlearning learns the logit offset needed for unlearning by contrasting the logits from a pair of smaller models. Experiments demonstrate that $\delta$-unlearning can effectively unlearn target data while maintaining similar or even stronger performance on general out-of-forget-scope tasks. $\delta$-unlearning also effectively incorporates different unlearning algorithms, making our approach a versatile solution to adapting various existing unlearning algorithms to black-box LLMs.

AAAI Conference 2021 Conference Paper

Exacerbating Algorithmic Bias through Fairness Attacks

• Ninareh Mehrabi
• Muhammad Naveed
• Fred Morstatter
• Aram Galstyan

Algorithmic fairness has attracted significant attention in recent years, with many quantitative measures suggested for characterizing the fairness of different machine learning algorithms. Despite this interest, the robustness of those fairness measures with respect to an intentional adversarial attack has not been properly addressed. Indeed, most adversarial machine learning has focused on the impact of malicious attacks on the accuracy of the system, without any regard to the system’s fairness. We propose new types of data poisoning attacks where an adversary intentionally targets the fairness of a system. Specifically, we propose two families of attacks that target fairness measures. In the anchoring attack, we skew the decision boundary by placing poisoned points near specific target points to bias the outcome. In the influence attack on fairness, we aim to maximize the covariance between the sensitive attributes and the decision outcome and affect the fairness of the model. We conduct extensive experiments that indicate the effectiveness of our proposed attacks.

IJCAI Conference 2019 Conference Paper

SAGE: A Hybrid Geopolitical Event Forecasting System

• Fred Morstatter
• Aram Galstyan
• Gleb Satyukov
• Daniel Benjamin
• Andres Abeliuk
• Mehrnoosh Mirtaheri
• KSM Tozammel Hossain
• Pedro Szekely

Forecasting of geopolitical events is a notoriously difficult task, with experts failing to significantly outperform a random baseline across many types of forecasting events. One successful way to increase the performance of forecasting tasks is to turn to crowdsourcing: leveraging many forecasts from non-expert users. Simultaneously, advances in machine learning have led to models that can produce reasonable, although not perfect, forecasts for many tasks. Recent efforts have shown that forecasts can be further improved by ``hybridizing'' human forecasters: pairing them with the machine models in an effort to combine the unique advantages of both. In this demonstration, we present Synergistic Anticipation of Geopolitical Events (SAGE), a platform for human/computer interaction that facilitates human reasoning with machine models.

JMLR Journal 2018 Journal Article

In Search of Coherence and Consensus: Measuring the Interpretability of Statistical Topics

• Fred Morstatter
• Huan Liu

Topic modeling is an important tool in natural language processing. Topic models provide two forms of output. The first is a predictive model. This type of model has the ability to predict unseen documents (e.g., their categories). When topic models are used in this way, there are ample measures to assess their performance. The second output of these models is the topics themselves. Topics are lists of keywords that describe the top words pertaining to each topic. Often, these lists of keywords are presented to a human subject who then assesses the meaning of the topic, which is ultimately subjective. One of the fundamental problems of topic models lies in assessing the quality of the topics from the perspective of human interpretability. Naturally, human subjects need to be employed to evaluate interpretability of a topic. Lately, crowdsourcing approaches are widely used to serve the role of human subjects in evaluation. In this work we study measures of interpretability and propose to measure topic interpretability from two perspectives: topic coherence and topic consensus. We start with an existing measure for topic coherence---model precision. It evaluates coherence of a topic by introducing an intruded word and measuring how well a human subject or a crowdsourcing approach could identify the intruded word: if it is easy to identify, the topic is coherent. We then investigate how we can measure coherence comprehensively by examining dimensions of topic coherence. For the second perspective of topic interpretability, we suggest topic consensus that measures how well the results of a crowdsourcing approach matches those given categories of topics. Good topics should lead to good categories, thus, high topic consensus. Therefore, if there is low topic consensus in terms of categories, topics could be of low interpretability. We then further discuss how topic coherence and topic consensus assess different aspects of topic interpretability and hope that this work can pave way for comprehensive measures of topic interpretability. [abs] [ pdf ][ bib ] &copy JMLR 2018. ( edit, beta )

AAAI Conference 2016 Conference Paper

Predicting Online Protest Participation of Social Media Users

• Suhas Ranganath
• Fred Morstatter
• Xia Hu
• Jiliang Tang
• Suhang Wang
• Huan Liu

IS Journal 2012 Journal Article

Opening Doors to Sharing Social Media Data

• Fred Morstatter
• Huan Liu
• Daniel Zeng

Research data sharing becomes increasingly difficult in the context of social media. Increasing restrictions from social media sites are creating an environment where data cannot be freely shared and as a result scientific claims cannot be verified. In this work, we present a novel approach to data sharing that does not require explicitly publishing a dataset. We create a framework where researchers systematically share the parameters they used to crawl the dataset along with the code used to collect the data, allowing the reader to re-assemble the dataset at a later time. While this approach is by no means a silver bullet, we seek to start a conversation for researchers to implement approaches to data sharing that can be embraced by the research community.