Large Language Models (LLMs) have been widely explored in educational scenarios.We identify a critical vulnerability in current educational LLMs, pedagogical jailbreaks, where students use answer-inducing prompts to elicit solutions rather than scaffolded instructions.To enable systematic study, we unify and formalize safe, helpful, and pedagogical behaviors with a knowledge-mastery graph and introduce SHAPE, a benchmark of 9,087 studentquestion pairs for evaluating tutoring behavior under adversarial pressure.We propose a graph-augmented tutoring pipeline that infers prerequisite concepts from queries, identifies mastery gaps, and routes generation between instructing and problem-solving via explicit gating.Experiments across multiple LLMs show that our method yields significantly improved safety under two pedagogical jailbreak settings, while maintaining near-ceiling helpfulness under the same evaluation protocol.

We introduce ImageGem, a dataset for studying generative models that understand fine-grained individual preferences. We posit that a key challenge hindering the development of such a generative model is the lack of in-the-wild and fine-grained user preference annotations. Our dataset features real-world interaction data from 57K users, who collectively have built 242K customized LoRAs, written 3M text prompts, and created 5M generated images. With user preference annotations from our dataset, we were able to train better preference alignment models. In addition, leveraging individual user preference, we investigated the performance of retrieval models and a vision-language model on personalized image retrieval and generative model recommendation. Finally, we propose an end-to-end framework for editing customized diffusion models in a latent weight space to align with individual user preferences. Our results demonstrate that the ImageGem dataset enables, for the first time, a new paradigm for generative model personalization.

Recommender Systems are built to retrieve relevant items to satisfy users' information needs. The candidate corpus usually consists of a finite set of items that are ready to be served, such as videos, products, or articles. With recent advances in Generative AI such as GPT and Diffusion models, a new form of recommendation task is yet to be explored where items are to be created by generative models with personalized prompts. Taking image generation as an example, with a single prompt from the user and access to a generative model, it is possible to generate hundreds of new images in a few minutes. How shall we attain personalization in the presence of "infinite" items? In this preliminary study, we propose a two-stage framework, namely Prompt-Model Retrieval and Generative Model Ranking, to approach this new task formulation. We release GEMRec-18K, a prompt-model interaction dataset with 18K images generated by 200 publicly available generative models paired with a diverse set of 90 textual prompts. Through a demo user interface based on the proposed framework, we illustrate the promise of Generative Model Recommendation as a novel personalization problem and highlight future directions. Our code and dataset are available at: https://github.com/MAPS-research/GEMRec.

Modern recommender systems have evolved rapidly along with deep learning models that are well-optimized for overall performance, especially those trained under Empirical Risk Minimization (ERM). However, a recommendation algorithm that focuses solely on the average performance may reinforce the exposure bias and exacerbate the “rich-get-richer” effect, leading to unfair user experience. In a simulation study, we demonstrate that such performance gap among various user groups is enlarged by an ERM-trained recommender in the long-term. To mitigate such amplification effects, we propose to optimize for the worst-case performance under the Distributionally Robust Optimization (DRO) framework, with the goal of improving long-term fairness for disadvantaged subgroups. In addition, we propose a simple-yet-effective streaming optimization improvement called Streaming-DRO (S-DRO), which effectively reduces loss variances for recommendation problems with sparse and long-tailed data distributions. Our results on two large-scale datasets suggest that (1) DRO is a flexible and effective technique for improving worst-case performance, and (2) Streaming-DRO outperforms vanilla DRO and other strong baselines by improving the worst-case and overall performance at the same time.

Impaired social functioning is a symptom of mental illness (e.g., depression, schizophrenia) and a wide range of other conditions (e.g., cognitive decline in the elderly, dementia). Today, assessing social functioning relies on subjective evaluations and self assessments. We propose a different approach and collect detailed social functioning measures and objective mobile sensing data from N=55 outpatients living with schizophrenia to study new methods of passively accessing social functioning. We identify a number of behavioral patterns from sensing data, and discuss important correlations between social function sub-scales and mobile sensing features. We show we can accurately predict the social functioning of outpatients in our study including the following sub-scales: prosocial activities (MAE = 7.79, r = 0.53), which indicates engagement in common social activities; interpersonal behavior (MAE = 3.39, r = 0.57), which represents the number of friends and quality of communications; and employment/occupation (MAE = 2.17, r = 0.62), which relates to engagement in productive employment or a structured program of daily activity. Our work on automatically inferring social functioning opens the way to new forms of assessment and intervention across a number of areas including mental health and aging in place.

The last year has been a breakout year for podcasts. There are now over 1 million podcast shows and over 64 million podcast episodes available through public RSS feeds. In the United States, 32% of all people listened to a podcast every month, and forecasts point to global podcast listenership to reach 2.2 billion monthly listeners by 2024. The workshop on Podcast Recommendations (PodRecs), collocated with RecSys 2020, introduces researchers in other domains of recommender systems to the special characteristics and challenges of podcast recommendations: how podcasts are as created, consumed, and how we might see algorithms being designed specifically for podcast content. We hope this workshop will help grow a community of researchers and foster an active research and innovation in the field.

Recommender systems play an important role in modern information and e-commerce applications. While increasing research is dedicated to improving the relevance and diversity of the recommendations, the potential risks of state-of-the-art recommendation models are under-explored, that is, these models could be subject to attacks from malicious third parties, through injecting fake user interactions to achieve their purposes. This paper revisits the adversarially-learned injection attack problem, where the injected fake user ‘behaviors’ are learned locally by the attackers with their own model – one that is potentially different from the model under attack, but shares similar properties to allow attack transfer. We found that most existing works in literature suffer from two major limitations: (1) they do not solve the optimization problem precisely, making the attack less harmful than it could be, (2) they assume perfect knowledge for the attack, causing the lack of understanding for realistic attack capabilities. We demonstrate that the exact solution for generating fake users as an optimization problem could lead to a much larger impact. Our experiments on a real-world dataset reveal important properties of the attack, including attack transferability and its limitations. These findings can inspire useful defensive methods against this possible existing attack.

Implicit feedback (e.g., clicks) is widely used in content recommendations. However, clicks only reflect user preferences according to their first impressions. They do not capture the extent to which users continue to engage with the content. Our analysis shows that more than half of the clicks on music and short videos are followed by skips from two real-world datasets. In this paper, we leverage post-click feedback, e.g. skips and completions, to improve the training and evaluation of content recommenders. Specifically, we experiment with existing collaborative filtering algorithms and find that they perform poorly against post-click-aware ranking metrics. Based on these insights, we develop a generic probabilistic framework to fuse click and post-click signals. We show how our framework can be applied to improve pointwise and pairwise recommendation models. Our approach is shown to outperform existing methods by 18.3% and 2.5% respectively in terms of Area Under the Curve (AUC) on the short-video and music dataset. We discuss the effectiveness of our approach across content domains and trade-offs in weighting various user feedback signals.

Traditionally, recommendation systems are built on the assumption that each service provider has full access to all user data generated on its platform. However, with increasing data privacy concerns and personal data protection regulation, service providers such as Google, Twitter, and Facebook are enabling their users to revisit, erase, and rectify their historical profiles. Future recommendation systems need to be robust to such profile modifications and user-controlled data filtering. In this paper, we explore how recommendation performance may be affected by time-sensitive user data filtering, that is, users choosing to share only recent "N days" of data. Using the MovieLens dataset as a testbed, we evaluated three widely used collaborative filtering algorithms. Our experiments demonstrate that filtering out historical user data does not significantly affect the overall recommendation performance, but its impact on individual users may vary. These findings challenge the common belief that more data is essential to better performance, and suggest a potential win-win solution for services and end users.

This tutorial reviews recent developments of deep neural network-based recommendation algorithms and demonstrates how to extend and adapt such algorithms for diverse application scenarios. The customization is supported by OpenRec framework that modularizes neural recommenders. The tutorial consists of a lecture and two hands-on sessions. It targets intermediate and advanced audiences who already possess knowledge of deep neural networks and are interested in applying those knowledge to the domain of recommendation. Materials are available at: http://openrec.ai/

Touch interaction on smartwatches suffers from the awkwardness of having to use two hands and the "fat finger" problem. We present Float, a wrist-to-finger input approach that enables one-handed and touch-free target selection on smartwatches with high efficiency and precision using only commercially-available built-in sensors. With Float, a user tilts the wrist to point and performs an in-air finger tap to click. To realize Float, we first explore the appropriate motion space for wrist tilt and determine the clicking action (finger tap) through a user-elicitation study. We combine the photoplethysmogram (PPG) signal with accelerometer and gyroscope to detect finger taps with a recall of 97.9% and a false discovery rate of 0.4%. Experiments show that using just one hand, Float allows users to acquire targets with size ranging from 2mm to 10mm in less than 2s to 1s, meanwhile achieve much higher accuracy than direct touch in both stationary (>98.9%) and walking (>71.5%) contexts.

Previous work on muscle activity sensing has leveraged specialized sensors such as electromyography and force sensitive resistors. While these sensors show great potential for detecting finger/hand gestures, they require additional hardware that adds to the cost and user discomfort. Past research has utilized sensors on commercial devices, focusing on recognizing gross hand gestures. In this work we present Serendipity, a new technique for recognizing unremarkable and fine-motor finger gestures using integrated motion sensors (accelerometer and gyroscope) in off-the-shelf smartwatches. Our system demonstrates the potential to distinguish 5 fine-motor gestures like pinching, tapping and rubbing fingers with an average f1-score of 87%. Our work is the first to explore the feasibility of using solely motion sensors on everyday wearable devices to detect fine-grained gestures. This promising technology can be deployed today on current smartwatches and has the potential to be applied to cross-device interactions, or as a tool for research in fields involving finger and hand motion.

In this paper, we investigate the effects of post-selection feedback for acquiring ultra-small (2-4mm) targets on touchscreens. Post-selection feedback shows the contact point on touchscreen after a user lifts his/her fingers to increase users' awareness of touching. Three experiments are conducted progressively using a single crosshair target, two reciprocally acquired targets and 2D random targets. Results show that in average post-selection feedback can reduce touch error rates by 78.4%, with a compromise of target acquisition time no more than 10%. In addition, we investigate participants' adjustment behavior based on correlation between successive trials. We conclude that the benefit of post-selection feedback is the outcome of both improved understanding about finger/point mapping and the dynamic adjustment of finger movement enabled by the visualization of the touch point.