Homa Hosseinmardi

Amir Ghasemian, Homa Hosseinmardi, Upasana Dutta et al.

Recommender systems have grown from content-organization tools into sophisticated systems that shape daily behavior. By controlling what we see, they shape what we perceive, raising concerns about filter bubbles, radicalization, polarization, and social inequality. Large language models (LLMs) enable more powerful personalization, intensifying these dynamics. Yet most recommenders are tuned for engagement or limited accuracy metrics, with little attention to broader social implications, e.g. how personalization reshapes exposure in socially consequential domains. We investigate whether LLM-assisted reranking, while improving personalization, inadvertently amplifies exposure to ideologically extreme or conspiratorial political content, a risk theorized but not empirically characterized in news recommendation. Using real news-consumption histories, we rerank YouTube's sidebar candidates through zero-shot, instruction-based prompting. We compare a baseline prompt with a constrained variant that preserves topical relevance and broadens ideological exposure while reducing conspiratorial or extreme content. Without constraints, reranking strengthened personalization but increased exposure to conspiratorial and extremist material for users whose histories contained such content. Lightweight prompt-level regularization reduced promotion of extreme content and increased ideological diversity, with modest relevance loss. Synthetic experiments suggest that LLMs rerank via statistical regularities in language rather than semantic understanding of ideology, clarifying why naive prompts amplify these patterns and why regularization can reshape them. Together, our results highlight the power of LLMs to operationalize contextual nuance in high-stakes recommendation, and the need to evaluate LLM-assisted personalization beyond accuracy and treat prompt design as a value-laden rather than neutral default.

Pooriya Jamie, Amir Ghasemian, Homa Hosseinmardi

Recommender systems on social media increasingly mediate how users encounter mental health content, yet it remains unclear whether they distinguish help-seeking from distress expression. We conduct a controlled 7-day audit of TikTok's "For You" page using 30 fresh accounts and LLM-guided agents that vary initial search framing (distress- vs. help-initiated) and interaction strategy (engaged, avoidant, passive). Across 8,727 recommended videos, interaction behavior dominates exposure outcomes: engagement rapidly saturates feeds with mental health content (~45% of daily recommendations), while avoidance and passive viewing reduce but do not eliminate exposure (~11-20%). Search framing mainly shifts composition rather than volume--help-initiated searches yield more potentially supportive material, yet potentially harmful content persists at low but non-zero levels, including content in the Suicide/Self-Harm category. These findings suggest limited sensitivity to user intent signals in TikTok's recommendations and motivate context-aware safeguards for sensitive topics.

Guang Yang, Amir Ghasemian, Fengchen Liu et al.

Detecting unauthorized knowledge distillation from a deployed LLM API is hard because the defender controls neither the attacker's training pipeline nor the next-token logits. Existing defenses operate on the teacher's output tokens -- biasing the next-token distribution (green-list watermarks, cryptographic schemes, antidistillation sampling) or rewriting outputs after generation. Recent work shows a paraphrasing attacker can strip these signals without losing the underlying knowledge. We propose interaction-layer antidistillation watermarks, which move the trace one layer higher, into the teacher's interaction behavior: the defender wraps the teacher with a system prompt that intermittently induces a behavioral marker -- an explicit follow-up question, a low-frequency variant, or a declarative restatement. An oblivious distiller inherits the behavior, and the defender audits via black-box queries with a human-validated LLM-as-judge (Cohen's kappa = 0.84/0.78 on strong/style rubrics). Across 63 LoRA-distilled students under a Llama-3.3-70B-Instruct teacher (35,343 judged samples), behavioral watermarks transfer at 88.9% (Gemma) / 80.9% (OLMo) / 45.2% (Qwen) relative fidelity (H1, H2). Under non-adaptive DIPPER paraphrasing, robustness decomposes into a teacher-self ceiling (about 66.4%) and student-relative retention of 21-112%, with OLMo preserving the watermark above the teacher itself (H3, F-Amp). Low-density (about 20%) explicit and implicit declarative variants transfer above per-family baseline (H4, F-Style). An N=20 in-lab study (pre-registered Latin-square) shows all marker variants within 0.22 Likert step of baseline; TOST, Friedman, and Bonferroni-Wilcoxon support H5. The interaction layer is a viable design locus for antidistillation watermarking, complementary to token-, model-, and reasoning-trace-layer defenses.

Guang Yang, Amir Ghasemian, Ninareh Mehrabi et al.

The proliferation of deepfake audio challenges voice-based authentication systems; passive forensic detectors are sensitive to evolving generative models and to real-world channel distortions. We propose Asymmetric Phase Coding (APC), a training-free cryptographic signing layer for audio, designed as a compact and auditable provenance primitive that can stand alone or be stacked with learned watermarks. APC combines Ed25519 digital signatures (EdDSA, FIPS 186-5; 64-byte signatures) with Reed-Solomon error correction, pseudo-random STFT phase-bin selection, and a redundant quantization-index-modulation (QIM) code on log-magnitude differences of adjacent bin pairs, yielding a compact, non-repudiable, blind-extractable watermark. We evaluate APC on 1,000 LibriSpeech test-clean clips (10 s each, 44.1 kHz) under eight attack configurations -- identity, 10% end-cropping, 20% end-cropping, 8 kHz low-pass, 16 kHz round-trip resampling, FLAC re-encoding, MP3 at 128 kbps, and OGG-Vorbis at 128 kbps -- and achieve cryptographic verification rates between 97.5% and 98.3% on every condition at mean PESQ=3.02 and tens-of-milliseconds CPU latency. We explicitly compare APC against recent neural baselines (AudioSeal, WavMark, SilentCipher), detail the threat model (forgery resistance vs. erasure), characterize the dataset, define all metrics, quantify an adaptive white-box erasure attack, and release code, keys, and metadata for reproducibility.

Homa Hosseinmardi, Amir Ghasemian, Aaron Clauset et al.

Although it is under-studied relative to other social media platforms, YouTube is arguably the largest and most engaging online media consumption platform in the world. Recently, YouTube's scale has fueled concerns that YouTube users are being radicalized via a combination of biased recommendations and ostensibly apolitical anti-woke channels, both of which have been claimed to direct attention to radical political content. Here we test this hypothesis using a representative panel of more than 300,000 Americans and their individual-level browsing behavior, on and off YouTube, from January 2016 through December 2019. Using a labeled set of political news channels, we find that news consumption on YouTube is dominated by mainstream and largely centrist sources. Consumers of far-right content, while more engaged than average, represent a small and stable percentage of news consumers. However, consumption of anti-woke content, defined in terms of its opposition to progressive intellectual and political agendas, grew steadily in popularity and is correlated with consumption of far-right content off-platform. We find no evidence that engagement with far-right content is caused by YouTube recommendations systematically, nor do we find clear evidence that anti-woke channels serve as a gateway to the far right. Rather, consumption of political content on YouTube appears to reflect individual preferences that extend across the web as a whole.

Nazgol Tavabi, Homa Hosseinmardi, Jennifer L. Villatte et al.

Continuous collection of physiological data from wearable sensors enables temporal characterization of individual behaviors. Understanding the relation between an individual's behavioral patterns and psychological states can help identify strategies to improve quality of life. One challenge in analyzing physiological data is extracting the underlying behavioral states from the temporal sensor signals and interpreting them. Here, we use a non-parametric Bayesian approach to model sensor data from multiple people and discover the dynamic behaviors they share. We apply this method to data collected from sensors worn by a population of hospital workers and show that the learned states can cluster participants into meaningful groups and better predict their cognitive and psychological states. This method offers a way to learn interpretable compact behavioral representations from multivariate sensor signals.

Amir Ghasemian, Homa Hosseinmardi, Aram Galstyan et al.

Most real-world networks are incompletely observed. Algorithms that can accurately predict which links are missing can dramatically speedup the collection of network data and improve the validity of network models. Many algorithms now exist for predicting missing links, given a partially observed network, but it has remained unknown whether a single best predictor exists, how link predictability varies across methods and networks from different domains, and how close to optimality current methods are. We answer these questions by systematically evaluating 203 individual link predictor algorithms, representing three popular families of methods, applied to a large corpus of 548 structurally diverse networks from six scientific domains. We first show that individual algorithms exhibit a broad diversity of prediction errors, such that no one predictor or family is best, or worst, across all realistic inputs. We then exploit this diversity via meta-learning to construct a series of "stacked" models that combine predictors into a single algorithm. Applied to a broad range of synthetic networks, for which we may analytically calculate optimal performance, these stacked models achieve optimal or nearly optimal levels of accuracy. Applied to real-world networks, stacked models are also superior, but their accuracy varies strongly by domain, suggesting that link prediction may be fundamentally easier in social networks than in biological or technological networks. These results indicate that the state-of-the-art for link prediction comes from combining individual algorithms, which achieves nearly optimal predictions. We close with a brief discussion of limitations and opportunities for further improvement of these results.

Homa Hosseinmardi, Hsien-Te Kao, Kristina Lerman et al.

In recent years, the rapid growth in technology has increased the opportunity for longitudinal human behavioral studies. Rich multimodal data, from wearables like Fitbit, online social networks, mobile phones etc. can be collected in natural environments. Uncovering the underlying low-dimensional structure of noisy multi-way data in an unsupervised setting is a challenging problem. Tensor factorization has been successful in extracting the interconnected low-dimensional descriptions of multi-way data. In this paper, we apply non-negative tensor factorization on a real-word wearable sensor data, StudentLife, to find latent temporal factors and group of similar individuals. Meta data is available for the semester schedule, as well as the individuals' performance and personality. We demonstrate that non-negative tensor factorization can successfully discover clusters of individuals who exhibit higher academic performance, as well as those who frequently engage in leisure activities. The recovered latent temporal patterns associated with these groups are validated against ground truth data to demonstrate the accuracy of our framework.

Homa Hosseinmardi, Amir Ghasemian, Shrikanth Narayanan et al.

Today's densely instrumented world offers tremendous opportunities for continuous acquisition and analysis of multimodal sensor data providing temporal characterization of an individual's behaviors. Is it possible to efficiently couple such rich sensor data with predictive modeling techniques to provide contextual, and insightful assessments of individual performance and wellbeing? Prediction of different aspects of human behavior from these noisy, incomplete, and heterogeneous bio-behavioral temporal data is a challenging problem, beyond unsupervised discovery of latent structures. We propose a Supervised Tensor Embedding (STE) algorithm for high dimension multimodal data with join decomposition of input and target variable. Furthermore, we show that features selection will help to reduce the contamination in the prediction and increase the performance. The efficiently of the methods was tested via two different real world datasets.

Amir Ghasemian, Homa Hosseinmardi, Aaron Clauset

A common data mining task on networks is community detection, which seeks an unsupervised decomposition of a network into structural groups based on statistical regularities in the network's connectivity. Although many methods exist, the No Free Lunch theorem for community detection implies that each makes some kind of tradeoff, and no algorithm can be optimal on all inputs. Thus, different algorithms will over or underfit on different inputs, finding more, fewer, or just different communities than is optimal, and evaluation methods that use a metadata partition as a ground truth will produce misleading conclusions about general accuracy. Here, we present a broad evaluation of over and underfitting in community detection, comparing the behavior of 16 state-of-the-art community detection algorithms on a novel and structurally diverse corpus of 406 real-world networks. We find that (i) algorithms vary widely both in the number of communities they find and in their corresponding composition, given the same input, (ii) algorithms can be clustered into distinct high-level groups based on similarities of their outputs on real-world networks, and (iii) these differences induce wide variation in accuracy on link prediction and link description tasks. We introduce a new diagnostic for evaluating overfitting and underfitting in practice, and use it to roughly divide community detection methods into general and specialized learning algorithms. Across methods and inputs, Bayesian techniques based on the stochastic block model and a minimum description length approach to regularization represent the best general learning approach, but can be outperformed under specific circumstances. These results introduce both a theoretically principled approach to evaluate over and underfitting in models of network community structure and a realistic benchmark by which new methods may be evaluated and compared.

Homa Hosseinmardi, Sabrina Arredondo Mattson, Rahat Ibn Rafiq et al.

Cyberbullying is a growing problem affecting more than half of all American teens. The main goal of this paper is to investigate fundamentally new approaches to understand and automatically detect and predict incidents of cyberbullying in Instagram, a media-based mobile social network. In this work, we have collected a sample data set consisting of Instagram images and their associated comments. We then designed a labeling study and employed human contributors at the crowd-sourced CrowdFlower website to label these media sessions for cyberbullying. A detailed analysis of the labeled data is then presented, including a study of relationships between cyberbullying and a host of features such as cyberaggression, profanity, social graph features, temporal commenting behavior, linguistic content, and image content. Using the labeled data, we further design and evaluate the performance of classifiers to automatically detect and pre- dict incidents of cyberbullying and cyberaggression.