Can a machine be surprised?

We explore how artificial intelligence can learn not just from data, but from the unexpected. Our method, Learning via Surprisability (LvS), teaches AI systems to form expectations and adjust when reality diverges, just as humans do.

Watch my short talk — Teaching AI the Power of Surprise

By learning from surprise, AI becomes more efficient, adaptable, and explainable. We’ve applied LvS to social media, political discourse, and biomedical data, revealing how absence and deviation can be as meaningful as presence.

If you’d like to support this research and help advance more human-like, transparent AI learning, visit: Support Learning via Surprisability (LvS)

The paper, “PAIRSAT: Integrating Preference-Based Signals for User Satisfaction Estimation in Dialogue Systems,” tackles a central challenge in conversational AI: how to measure user satisfaction in a way that is both accurate and scalable.

Most current systems rely on explicit satisfaction labels (“satisfied”, “neutral”, “dissatisfied”), but these are expensive to collect and often domain-specific. Eran’s work introduces PAIRSAT, a new model that combines these scarce labels with the abundant preference data generated when users choose one AI response over another. By reframing satisfaction prediction as a bounded regression task and integrating pairwise ranking loss, PAIRSAT is able to capture both the nuance of absolute labels and the richness of relative feedback.

The results are compelling: across multiple datasets, PAIRSAT demonstrates strong and robust performance, showing that preference data can be a powerful complement to traditional labels.

This work is part of our lab’s broader agenda on human-centered AI and recommender systems, and we are delighted to see Eran’s contributions recognized by the RecSys community.

Congratulations to Eran on this achievement — and thank you to Adir and the team for their support and collaboration.

Read the paper: ACM RecSys 2025 Proceedings

“Consumer Sentiment and Hotel Aspect Preferences Across Trip Modes and Purposes”

In this study, I explore the nuances of online hotel reviews — specifically, how a traveler’s trip type (solo, couple, business) affects what they care about and how they describe their stay.

With over 137,000 reviews analyzed, this paper contributes to both text analytics methods and our understanding of user behavior in e-commerce settings.

The key message: understanding who is writing the review is just as important as what they’re writing.

Some details:

• Travelers don’t always “say” what they “rate” — textual sentiment often diverges from star ratings, especially for solo travelers.
• Using a graph-based clustering technique, I extracted hotel aspects (like food, room, location) directly from the review text, without pre-labeling.
• Different traveler types mention different things: business travelers talk more about service and room, couples focus on location, and solo travelers complain more — but mention food the least.

Here is a link to the discussion: https://youtu.be/U3aYXp8LT6Q

Being part of the University’s Ambassadors Program has been an incredible experience, showing me firsthand the value of connecting with researchers across different fields. These interdisciplinary conversations have truly expanded the way I think and work, and I’m excited to continue that exchange during this event.

Title: From Feelings to Fingerprints: New Approaches to Textual Communication

Abstract: In this talk, I will present two novel methods for analyzing human online discourse, each offering a new lens for understanding how people express themselves in written communication.

First, I will show how online reviews — typically viewed as reflections of consumer opinions — can also be mined for the emotional experiences users have while interacting with experience goods, such as films. By aggregating these emotional traces, we can construct emotional signatures for items, offering new insights into collective emotional responses and their role in decision-making and recommendation systems.

Second, I will introduce Learning via Surprisability (LvS), a method for identifying how individuals’ language deviates from domain norms. LvS allows us to extract distinctive personal signatures from text, which can be applied to explainable authorship attribution, the detection of online impersonation, and broader discourse analyses. I will also demonstrate how these tools can be applied to real-world contexts by analyzing the discourse surrounding the Russian attack on Ukraine.

I’m very much looking forward to the discussion and to sharing this work with colleagues and students. Hope to see you there!

My talk, “Using an Interaction-Driven Contagion Model to Understand the Temporal Dynamics of Disease Spread,” summarized three recently published papers. These works focused on a novel temporal model for analyzing the dynamic interplay between people’s interactions and the spread of airborne diseases.

Through unique temporal modeling, we showed that:

• The temporal dynamics of a community have a more significant effect on the spread of the disease than the characteristics of the spreading processes
• When considering airborne diseases, it is important to consider the duration of temporal meetings to model the spread of pathogens in a population
• The enigmatic nature of asymptomatic transmission stems from the latent effect of the network density, and has a substantial impact only in sparse communities

Many thanks to my former lab members Alex Abbey and Yanir Marmor, and our collaborator Yuval Shahar.

I am thankful for the opportunity to share and engage in such a vibrant scientific community!

Maybe it is a sudden dip in your fitness tracker stats, an unusual spike in your energy bill, or a strange shift in market trends. Our brains are wired to notice unexpected changes — we instinctively focus on what surprises us. Osnat Mokryn noticed this and asked herself (and Teddy Lazebnik with Hagit Ben-Shoshan) to imagine applying that same intuition to analyzing complex time-series data.

In our latest work, we introduce Learning via Surprisability (LvS) — a novel approach inspired by how humans detect anomalies. Instead of treating all deviations equally, LvS prioritizes the most surprising changes, preserving critical context and making patterns easier to interpret.

Why does this matter? Traditional methods often drown in high-dimensional data, flagging anomalies without explaining why they matter. LvS takes a more human-like approach, focusing on what truly stands out.

What can LvS do? We tested it on:

• Sensor data with hidden anomalies
• Global mortality trends over multiple years
• Two centuries of U.S. State of the Union Addresses

The result? LvS efficiently highlights the most important outliers, helping us see the bigger picture instead of just noise.

If you’re working with time-series data and looking for a better way to detect and understand anomalies, LvS might be the breakthrough you need.

Read more: ArXiv paper

Analysis of causes of death over 30 years: (a) A comparison of the original vectors representing causes of death over a 30-year period. Each square corresponds to the comparison of probability distribution vectors between two years, with lighter colors indicating greater similarity. The figure clearly shows that any two consecutive years exhibit a high degree of similarity in the underlying causes of death worldwide. (b) A comparison of the Surprisal Profile vectors created by LvS. Notable anomalies are clear in the years 1994, 2004, 2008, and 2010. The values within the SP vectors enable interpretation of the most surprising elements causing the anomaly in these years.

Special thanks to Alex Abbey for his help with the code in the early stages of the work.

This post shares our methodology, findings, and implications for the fields of Human-Computer Interaction (HCI) and creative design practice, as detailed in our paper: AI-Augmented Brainwriting: Investigating the use of LLMs in group ideation.

The intersection of AI and group creativity

The premise of our research was to investigate the role of LLMs in the divergent phase of group brainwriting — specifically, whether AI could enrich the process of generating a diverse array of ideas. To do this, we employed GPT-3, exploring its capacity to contribute unique perspectives and detailed insights to a group’s creative output.

Key findings: the impact of GPT-3 on idea generation

Our findings revealed a nuanced picture. Approximately half of the participants found GPT-3 to be a valuable ally in the ideation process, noting its ability to expand on problem statements with unique or technically detailed perspectives. However, a significant portion also observed GPT-3’s tendency toward redundancy and questioned its overall creativity. This feedback underscores the importance of effective prompt engineering to leverage AI’s capabilities fully.

Semantic divergence and convergence. Domain-based Latent Personal Analysis (LPA) was a key tool in our study, which we used to evaluate the semantic distribution of ideas generated by humans and GPT-3. LPA helps identify the unique terms in a document or idea relative to a larger corpus, effectively creating a “signature” for each set of ideas that highlights its distinctiveness or conformity. Our application of LPA and semantic clustering provided deep insights into the conceptual differences and overlaps between human and AI-generated ideas. While substantial overlap existed, the unique terminology used by GPT-3 pointed to its potential to augment human creativity meaningfully, without overshadowing it.

Using GPT-4 for evaluation. The convergence phase in our study was marked by a systematic approach to evaluating the plethora of ideas generated during the divergent phase. Here, GPT-4 played a central role, assessing ideas based on several criteria, including relevance to the problem statement, originality and creativity, and depth of understanding. On evaluating ideas during the convergent phase, GPT-4’s ratings aligned with the selections made by student teams, suggesting its viability in identifying promising ideas without prematurely discarding them. Yet the alignment between GPT-4 and expert evaluations wasn’t perfect, highlighting the complexities of AI-assisted idea evaluation.

How AI augments group creativity

Our findings reveal that AI, specifically GPT-3, can significantly augment the group creativity process by:

• Introducing new perspectives. GPT-3 often generated ideas that participants hadn’t considered, broadening the ideation scope and encouraging divergent thinking.
• Enhancing detail and technical insight. Ideas generated by GPT-3 included more technical details and usage insights, which helped in refining concepts and understanding their practical implications.
• Stimulating convergent thinking. By providing ideas that varied from human-generated concepts, GPT-3 supported the convergent thinking process, helping teams to develop and refine their ideas incrementally.

The role of effective prompt engineering

A critical aspect of maximizing the benefits of human-AI collaboration in the ideation process is effective prompt engineering. Crafting prompts that guide AI in generating useful and innovative ideas is crucial. Our study highlighted the need for prompts that challenge conventional thinking and stimulate creative responses from AI, enhancing the quality and diversity of the ideation output.

Implications for collaborative human-AI ideation

The integration of AI into the ideation process presents opportunities for HCI and creative design: a broader ideation scope, enhanced creativity and innovation by combining human creativity with AI’s computational abilities, and more efficient idea refinement.

By thoughtfully integrating AI into the creative process, we can unlock new levels of innovation, pushing the boundaries of what’s possible in design thinking and practice.

During lockdowns, Yuval — an IS professor and a medical doctor — was my go-to person for understanding viruses and their transmission. Thus, in the early days of COVID-19, we started formulating the disease progression, gathering all new evidence available. We analyzed an early outbreak in a Korean church and correlated viral load with exposure duration and intensity. We analyzed death rates at first outbreaks by age group. We devised a statistical model for early estimation of the prevalence and severity of an epidemic or pandemic from simple tests for infection confirmation.

Soon after, we were designing a methodology for quantifying the interplay of viral transmission in real-world societies while considering human behavior. COVID-19 was progressing in the world, country after country. Yet it became clear that the virus was progressing differently in different societies. The known R0 metric, describing the transmissibility of a disease, was not accurately capturing the viral transmission in different communities. Yuval and I hypothesized that the societal temporal interactions and exposure duration accounted for these differences.

At that stage, epidemiologic simulations were usually done on static models of communities. It was evident that many epidemiologic results would change under real-life conditions that take into account temporal characteristics such as temporal paths, and the mathematical characteristics of real-life interactions such as the rate with which people meet, the distribution of interactions, and the duration of meetings.

During the lockdown, recruiting research assistants and students was challenging. However, upon consulting Asaf Shapira, he recommended Yanir Marmor and Alex Abbey. After a brief conversation with each, I quickly realized that Asaf’s recommendations were excellent, leading them to join my lab as research assistants.

In the next two years, we devised a detailed model of the disease progression in humans, which considered different progressions for asymptomatic and symptomatic individuals; gathered COVID data from 15 countries (with the help of Yaniv Wisney); identified and analyzed a dataset of real-world interaction data; devised the first open-source temporal path-preserving Random Networks Generator; and devised the first interaction-driven modeling of viral transmission in societies using real-world interactions. Our model linked the infection probability with the interaction duration — as is the case for airborne viruses — and incorporated human behavior dynamics with the disease progression stages in humans. We made the open-source code of the Interaction-Driven Contagious Model with Individual Disease Progression Modeling freely available.

In a series of works, we demonstrated that the temporal dynamics of a community have a more significant effect on the spread of the disease than the characteristics of the airborne pathogen, supporting our initial hypothesis. The research culminated in several additional findings.

Effectiveness of social distancing strategies

Evaluating different social distancing strategies, we demonstrated the superiority of decreasing daily social interactions over partial isolation strategies, such as spatial distancing pods and a spatiotemporal distancing strategy.

We found that social distancing strategies are effective only at the beginning of the viral spread, when the vast majority of the population has not been infected yet.

These results appear in our Journal of Biomedical Informatics publication.

The effect of population dynamics on the spread of the virus

Highly active communities are more susceptible to viral infection. In highly active communities, slow-spreading pathogens (i.e., pathogens that require more prolonged exposure to infect) spread roughly at the same rate as fast-spreading ones. This result is surprising, since slow and fast pathogens follow different paths. A strategy to overcome this is to reduce the rate of daily interactions, thus reducing the spread of the disease.

These results appear in our 2024 Journal of Biomedical Informatics publication.

Predicting outcomes

We analyzed community and individual outcomes, and devised a prediction for the individual risk of engaging in social interactions as a function of the virus characteristics and its prevalence in the population.

Researching asymptomatic airborne transmission, we found that asymptomatic transmission accelerates viral transmission only in sparse communities, as the enigmatic nature of asymptomatic transmission stems from the latent effect of the temporal network density on transmission.

These results appear in our 2023 Scientific Reports publication.

Pathogen competition conditions

We identified the conditions under which the competition between several (two and three) competing airborne pathogens will result in the slower pathogen creating a second wave of infection that infects most of the population.

We then show that when the duration of the encounters is considered, the spreading dynamics change significantly. Our results indicate that when considering airborne diseases, it might be crucial to consider the duration of temporal meetings to model the spread of pathogens in a population.

These results appear in our 2022 Scientific Reports publication.

Registration: https://lnkd.in/emPKWid2