Explore the vast range of titles available.

Recommender systems can benefit from a plethora of signals influencing user behavior such as her past interactions, her social connections, as well as the similarity between different items. However, existing methods are challenged when taking all this data into account and often do not exploit all available information. This is primarily due to the fact that it is non-trivial to combine the various information as they mutually influence each other. To address this shortcoming, here, we propose a ‘Fusion Recommender’ (FuseRec), which models each of these factors separately and later combines them in an interpretable manner. We find this general framework to yield compelling results on all three investigated datasets, Epinions, Ciao, and CiaoDVD, outperforming the state-of-the-art by more than 14% for Ciao and Epinions. In addition, we provide a detailed ablation study, showing that our combined model achieves accurate results, often better than any of its components individually. Our model also provides insights on the importance of each of the factors in different datasets.

The clinical and experimental data on seasonal variation in blood pressure is mainly from office and home blood pressure (BP) monitoring studies. There are few studies from temperate climates on seasonal changes with ambulatory blood pressure (ABP) monitoring and none from India. This is a prospective, observational study among healthy adults. ABP was measured in four different seasons in 28 subjects. Mean arterial pressure (MAP), ambulatory systolic blood pressure (SBP), and ambulatory diastolic blood pressure (DBP) were significantly higher in winter compared to summer season. 24-hour MAP was lowest in summer while highest MAP was recorded in winter (97.04 ± 8.30 and 103.89 ± 8.54, respectively). The mean difference was −6.86 mm Hg (95% CI: −10.74 to −2.97, p = 0.001). This difference was mainly due to increase in day time MAP. There was no difference in 24 h systolic and diastolic blood pressure between summer and winter. There was significant difference between summer and winter in the SBP (day time) [125.61 ± 11.44 and 131.93 ± 9.46, mean difference −6.32 (95% CI: −10.69 to −1.95, p = 0.005)] and DBP (day time) [79.57 ± 9.95 and 87.07 ± 9.9, mean difference −7.50 (95% CI: −12.49 to −2.51, p = 0.003)]. The night time systolic and diastolic BP was similar during winter and summer. Thus, BP increases significantly during winter compared to summer season. This change is primarily in the day time systolic, diastolic and mean blood pressures. Larger studies are required to further validate our findings.

The current research used human-coded Reddit posts categorized by already established food parenting concepts (coercive control, structure, autonomy support, recipes) as a basis for machine learning models, with the objective of providing insight into topics related to feeding children discussed on social media and to provide a way for future research to use our trained machine-learned model. Reddit posts from specific, parenting-related subreddits were collected and labeled as they related to aspects of child-feeding behavior. Posts were then put through text pre-processing, converted into TF-IDF vectors, and used to train support vector machine binary and multiclass classification models. Other classifiers and text-preprocessing steps were also tested. After training, the binary model was able to classify posts with 86.1% accuracy as being about child feeding or not, up from a baseline accuracy of 57.6%. The multiclass model yielded a 79.1% accuracy to classify posts related to four categories of child feeding concepts (coercive control, autonomy support, structure, recipes), up from a baseline of 51.9%. The comparison models were found to perform less favorably. The best performing binary model is publicly available for use via the Social Media Macroscope and we provide details on how to use this model. Information is presented such that other researchers and professionals interested in examining issues related to feeding children posted on social media could effectively utilize the same approach. Highlights Machine learning models based on human-coded Reddit posts were developed. The binary model can classify posts as being about child feeding with 86.1% accuracy. The multiclass model can classify child feeding concepts in posts with 79.1% accuracy. The best performing binary model is made available for use with instructions provided.

User behavior modeling has become an indispensable tool with the proliferation of socio-technical systems to provide a highly personalized experience to the users. These socio-technical systems are used in sectors as diverse as education, health, law to e-commerce, and social media. The two main challenges for user behavioral modeling are building an in-depth understanding of online user behavior and using advanced computational techniques to capture behavioral uncertainties accurately. This thesis addresses both these challenges by developing interpretable models that aid in understanding user behavior at scale and by developing sophisticated models that perform accurate modeling of user behavior. Specifically, we first propose two distinct interpretable approaches to understand explicit and latent user behavioral characteristics. Firstly, in Chapter 3, we propose an interpretable Gaussian Hidden Markov Model-based cluster model leveraging user activity data to identify users with similar patterns of behavioral evolution. We apply our approach to identify researchers with similar patterns of research interests evolution. We further show the utility of our interpretable framework to identify differences in gender distribution and the value of awarded grants among the identified archetypes. We also demonstrate generality of our approach by applying on StackExchange to identify users with a similar change in usage patterns. Next in Chapter 4, we estimate user latent behavioral characteristics by leveraging user-generated content (questions or answers) in Community Question Answering (CQA) platforms. In particular, we estimate the latent aspect-based reliability representations of users in the forum to infer the trustworthiness of their answers. We also simultaneously learn the semantic meaning of their answers through text representations. We empirically show that the estimated behavioral representations can accurately identify topical experts. We further propose to improve current behavioral models by modeling explicit and implicit user-to-user influence on user behavior. To this end, in Chapter 5, we propose a novel attention-based approach to incorporate influence from both user's social connections and other similar users on their preferences in recommender systems. Additionally, we also incorporate implicit influence in the item space by considering frequently co-occurring and similar feature items. Our modular approach captures the different influences efficiently and later fuses them in an interpretable manner. Extensive experiments show that incorporating user-to-user influence outperforms approaches relying on solely user data. User behavior remains broadly consistent across the platform. Thus, incorporating user behavioral information can be beneficial to estimate the characteristics of user-generated content. To verify it, in Chapter 6, we focus on the task of best answer selection in CQA forums that traditionally only considers textual features. We induce multiple connections between user-generated content, i.e., answers, based on the similarity and contrast in the behavior of authoring users in the platform. These induced connections enable information sharing between connected answers and, consequently, aid in estimating the quality of the answer. We also develop convolution operators to encode these semantically different graphs and later merge them using boosting. We also proposed an alternative approach to incorporate user behavioral information by jointly estimating the latent behavioral representations of user with text representations in Chapter 7. We evaluate our approach on the offensive language prediction task on Twitter. Specially, we learn an improved text representation by leveraging syntactic dependencies between the words in the tweet. We also estimate the abusive behavior of users, i.e., their likelihood of posting offensive content online from their tweets. We further show that combining the textual and user behavioral features can outperform the sophisticated textual baselines.

Self-supervised learning (SSL) has shown impressive results in downstream classification tasks. However, there is limited work in understanding their failure modes and interpreting their learned representations. In this paper, we study the representation space of state-of-the-art self-supervised models including SimCLR, SwaV, MoCo, BYOL, DINO, SimSiam, VICReg and Barlow Twins. Without the use of class label information, we discover discriminative features that correspond to unique physical attributes in images, present mostly in correctly-classified representations. Using these features, we can compress the representation space by up to 40% without significantly affecting linear classification performance. We then propose Self-Supervised Representation Quality Score (or Q-Score), an unsupervised score that can reliably predict if a given sample is likely to be mis-classified during linear evaluation, achieving AUPRC of 91.45 on ImageNet-100 and 78.78 on ImageNet-1K. Q-Score can also be used as a regularization term on pre-trained encoders to remedy low-quality representations. Fine-tuning with Q-Score regularization can boost the linear probing accuracy of SSL models by up to 5.8% on ImageNet-100 and 3.7% on ImageNet-1K compared to their baselines. Finally, using gradient heatmaps and Salient ImageNet masks, we define a metric to quantify the interpretability of each representation. We show that discriminative features are strongly correlated to core attributes and, enhancing these features through Q-score regularization makes SSL representations more interpretable.

Large language models show impressive results on few-shot NLP tasks. However, these models are memory and computation-intensive. Meta-training allows one to leverage smaller models for few-shot generalization in a domain-general and task-agnostic manner; however, these methods alone results in models that may not have sufficient parameterization or knowledge to adapt quickly to a large variety of tasks. To overcome this issue, we propose meta-training with demonstration retrieval, where we use a dense passage retriever to retrieve semantically similar labeled demonstrations to each example for more varied supervision. By separating external knowledge from model parameters, we can use meta-training to train parameter-efficient models that generalize well on a larger variety of tasks. We construct a meta-training set from UnifiedQA and CrossFit, and propose a demonstration bank based on UnifiedQA tasks. To our knowledge, our work is the first to combine retrieval with meta-training, to use DPR models to retrieve demonstrations, and to leverage demonstrations from many tasks simultaneously, rather than randomly sampling demonstrations from the training set of the target task. Our approach outperforms a variety of targeted parameter-efficient and retrieval-augmented few-shot methods on QA, NLI, and text classification tasks (including SQuAD, QNLI, and TREC). Our approach can be meta-trained and fine-tuned quickly on a single GPU.

Distilling conversational skills into Small Language Models (SLMs) with approximately 1 billion parameters presents significant challenges. Firstly, SLMs have limited capacity in their model parameters to learn extensive knowledge compared to larger models. Secondly, high-quality conversational datasets are often scarce, small, and domain-specific. Addressing these challenges, we introduce a novel data distillation framework named CoDi (short for Conversational Distillation, pronounced \"Cody\"), allowing us to synthesize large-scale, assistant-style datasets in a steerable and diverse manner. Specifically, while our framework is task agnostic at its core, we explore and evaluate the potential of CoDi on the task of conversational grounded reasoning for question answering. This is a typical on-device scenario for specialist SLMs, allowing for open-domain model responses, without requiring the model to \"memorize\" world knowledge in its limited weights. Our evaluations show that SLMs trained with CoDi-synthesized data achieve performance comparable to models trained on human-annotated data in standard metrics. Additionally, when using our framework to generate larger datasets from web data, our models surpass larger, instruction-tuned models in zero-shot conversational grounded reasoning tasks.

We hypothesize that a user's visual history with images reflecting their daily life, offers valuable insights into their interests and preferences, and can be leveraged for personalization. Among the many challenges to achieve this goal, the foremost is the diversity and noises in the visual history, containing images not necessarily related to a recommendation task, not necessarily reflecting the user's interest, or even not necessarily preference-relevant. Existing recommendation systems either rely on task-specific user interaction logs, such as online shopping history for shopping recommendations, or focus on text signals. We propose a novel approach, VisualLens, that extracts, filters, and refines image representations, and leverages these signals for personalization. We created two new benchmarks with task-agnostic visual histories, and show that our method improves over state-of-the-art recommendations by 5-10% on Hit@3, and improves over GPT-4o by 2-5%. Our approach paves the way for personalized recommendations in scenarios where traditional methods fail.

This paper investigates the relationship between graph convolution and Mixup techniques. Graph convolution in a graph neural network involves aggregating features from neighboring samples to learn representative features for a specific node or sample. On the other hand, Mixup is a data augmentation technique that generates new examples by averaging features and one-hot labels from multiple samples. One commonality between these techniques is their utilization of information from multiple samples to derive feature representation. This study aims to explore whether a connection exists between these two approaches. Our investigation reveals that, under two mild conditions, graph convolution can be viewed as a specialized form of Mixup that is applied during both the training and testing phases. The two conditions are: 1) \\textit{Homophily Relabel} - assigning the target node's label to all its neighbors, and 2) \\textit{Test-Time Mixup} - Mixup the feature during the test time. We establish this equivalence mathematically by demonstrating that graph convolution networks (GCN) and simplified graph convolution (SGC) can be expressed as a form of Mixup. We also empirically verify the equivalence by training an MLP using the two conditions to achieve comparable performance.

In this paper, we aim to discover archetypical patterns of individual evolution in large social networks. In our work, an archetype comprises of \\(\\textit{progressive stages}\\) of distinct behavior. We introduce a novel Gaussian Hidden Markov Model (G-HMM) Cluster to identify archetypes of evolutionary patterns. G-HMMs allow for: near limitless behavioral variation; imposing constraints on how individuals can evolve; different evolutionary rates; and are parsimonious. Our experiments with Academic and StackExchange dataset discover insightful archetypes. We identify four archetypes for researchers: \\(\\textit{Steady}\\), \\(\\textit{Diverse, Evolving and Diffuse}\\). We observe clear differences in the evolution of male and female researchers within the same archetype. Specifically, women and men differ within an archetype (e.g. Diverse) in how they start, how they transition and the time spent in mid-career. We also found that the differences in grant income are better explained by the differences in archetype than by differences in gender. For StackOverflow, discovered archetypes could be labeled as \\(\\textit{Experts, Seekers, Enthusiasts, and Facilitators}\\). We have strong quantitative results with competing baselines for activity prediction and perplexity. For future session prediction, the proposed G-HMM cluster model improves by an average of \\(32\\%\\) for different Stack Exchanges and \\(24\\%\\) for Academic dataset. Our model also exhibits lower perplexity than the baselines.