Search

Many existing benchmarks of large (multimodal) language models (LLMs) focus on measuring LLMs' academic proficiency, often with also an interest in comparing model performance with human test takers. While these benchmarks have proven key to the development of LLMs, they suffer from several limitations, including questionable measurement quality (e.g., Do they measure what they are supposed to in a reliable way?), lack of quality assessment on the item level (e.g., Are some items more important or difficult than others?) and unclear human population reference (e.g., To whom can the model be compared?). In response to these challenges, we propose leveraging knowledge from psychometrics - a field dedicated to the measurement of latent variables like academic proficiency - into LLM benchmarking. We make three primary contributions. First, we introduce PATCH: a novel framework for {P}sychometrics-{A}ssis{T}ed ben{CH}marking of LLMs. PATCH addresses the aforementioned limitations, presenting a new direction for LLM benchmark research. Second, we implement PATCH by measuring GPT-4 and Gemini-Pro-Vision's proficiency in 8th grade mathematics against 56 human populations. We show that adopting a psychometrics-based approach yields evaluation outcomes that diverge from those based on existing benchmarking practices. Third, we release 4 high-quality datasets to support measuring and comparing LLM proficiency in grade school mathematics and science against human populations.

Learning general-purpose user representations based on user behavioral logs is an increasingly popular user modeling approach. It benefits from easily available, privacy-friendly yet expressive data, and does not require extensive re-tuning of the upstream user model for different downstream tasks. While this approach has shown promise in search engines and e-commerce applications, its fit for instant messaging platforms, a cornerstone of modern digital communication, remains largely uncharted. We explore this research gap using Snapchat data as a case study. Specifically, we implement a Transformer-based user model with customized training objectives and show that the model can produce high-quality user representations across a broad range of evaluation tasks, among which we introduce three new downstream tasks that concern pivotal topics in user research: user safety, engagement and churn. We also tackle the challenge of efficient extrapolation of long sequences at inference time, by applying a novel positional encoding method., Comment: SIGIR 2024

Text-based personality computing (TPC) has gained many research interests in NLP. In this paper, we describe 15 challenges that we consider deserving the attention of the research community. These challenges are organized by the following topics: personality taxonomies, measurement quality, datasets, performance evaluation, modelling choices, as well as ethics and fairness. When addressing each challenge, not only do we combine perspectives from both NLP and social sciences, but also offer concrete suggestions. We hope to inspire more valid and reliable TPC research., Comment: Findings of ACL 2023. Long paper

Text embedding models from Natural Language Processing can map text data (e.g. words, sentences, documents) to supposedly meaningful numerical representations (a.k.a. text embeddings). While such models are increasingly applied in social science research, one important issue is often not addressed: the extent to which these embeddings are valid representations of constructs relevant for social science research. We therefore propose the use of the classic construct validity framework to evaluate the validity of text embeddings. We show how this framework can be adapted to the opaque and high-dimensional nature of text embeddings, with application to survey questions. We include several popular text embedding methods (e.g. fastText, GloVe, BERT, Sentence-BERT, Universal Sentence Encoder) in our construct validity analyses. We find evidence of convergent and discriminant validity in some cases. We also show that embeddings can be used to predict respondent's answers to completely new survey questions. Furthermore, BERT-based embedding techniques and the Universal Sentence Encoder provide more valid representations of survey questions than do others. Our results thus highlight the necessity to examine the construct validity of text embeddings before deploying them in social science research., Comment: Under review

A potentially powerful method of social-scientific data collection and investigation has been created by an unexpected institution: the law. Article 15 of the EU's 2018 General Data Protection Regulation (GDPR) mandates that individuals have electronic access to a copy of their personal data, and all major digital platforms now comply with this law by providing users with "data download packages" (DDPs). Through voluntary donation of DDPs, all data collected by public and private entities during the course of citizens' digital life can be obtained and analyzed to answer social-scientific questions - with consent. Thus, consented DDPs open the way for vast new research opportunities. However, while this entirely new method of data collection will undoubtedly gain popularity in the coming years, it also comes with its own questions of representativeness and measurement quality, which are often evaluated systematically by means of an error framework. Therefore, in this paper we provide a blueprint for digital trace data collection using DDPs, and devise a "total error framework" for such projects. Our error framework for digital trace data collection through data donation is intended to facilitate high quality social-scientific investigations using DDPs while critically reflecting its unique methodological challenges and sources of error. In addition, we provide a quality control checklist to guide researchers in leveraging the vast opportunities afforded by this new mode of investigation.

Electronic health records (EHRs) contain structured and unstructured data of significant clinical and research value. Various machine learning approaches have been developed to employ information in EHRs for risk prediction. The majority of these attempts, however, focus on structured EHR fields and lose the vast amount of information in the unstructured texts. To exploit the potential information captured in EHRs, in this study we propose a multimodal recurrent neural network model for cardiovascular risk prediction that integrates both medical texts and structured clinical information. The proposed multimodal bidirectional long short-term memory (BiLSTM) model concatenates word embeddings to classical clinical predictors before applying them to a final fully connected neural network. In the experiments, we compare performance of different deep neural network (DNN) architectures including convolutional neural network and long short-term memory in scenarios of using clinical variables and chest X-ray radiology reports. Evaluated on a data set of real world patients with manifest vascular disease or at high-risk for cardiovascular disease, the proposed BiLSTM model demonstrates state-of-the-art performance and outperforms other DNN baseline architectures., Comment: 8 pages, 2 figures

Clustering consists of a popular set of techniques used to separate data into interesting groups for further analysis. Many data sources on which clustering is performed are well-known to contain random and systematic measurement errors. Such errors may adversely affect clustering. While several techniques have been developed to deal with this problem, little is known about the effectiveness of these solutions. Moreover, no work to-date has examined the effect of systematic errors on clustering solutions. In this paper, we perform a Monte Carlo study to investigate the sensitivity of two common clustering algorithms, GMMs with merging and DBSCAN, to random and systematic error. We find that measurement error is particularly problematic when it is systematic and when it affects all variables in the dataset. For the conditions considered here, we also find that the partition-based GMM with merged components is less sensitive to measurement error than the density-based DBSCAN procedure.

Fair inference in supervised learning is an important and active area of research, yielding a range of useful methods to assess and account for fairness criteria when predicting ground truth targets. As shown in recent work, however, when target labels are error-prone, potential prediction unfairness can arise from measurement error. In this paper, we show that, when an error-prone proxy target is used, existing methods to assess and calibrate fairness criteria do not extend to the true target variable of interest. To remedy this problem, we suggest a framework resulting from the combination of two existing literatures: fair ML methods, such as those found in the counterfactual fairness literature on the one hand, and, on the other, measurement models found in the statistical literature. We discuss these approaches and their connection resulting in our framework. In a healthcare decision problem, we find that using a latent variable model to account for measurement error removes the unfairness detected previously., Comment: Online supplementary code is available at https://dx.doi.org/10.5281/zenodo.3708150

Combining data from varied sources has considerable potential for knowledge discovery: collaborating data parties can mine data in an expanded feature space, allowing them to explore a larger range of scientific questions. However, data sharing among different parties is highly restricted by legal conditions, ethical concerns, and / or data volume. Fueled by these concerns, the fields of cryptography and distributed learning have made great progress towards privacy-preserving and distributed data mining. However, practical implementations have been hampered by the limited scope or computational complexity of these methods. In this paper, we greatly extend the range of analyses available for vertically partitioned data, i.e., data collected by separate parties with different features on the same subjects. To this end, we present a novel approach for privacy-preserving generalized linear models, a fundamental and powerful framework underlying many prediction and classification procedures. We base our method on a distributed block coordinate descent algorithm to obtain parameter estimates, and we develop an extension to compute accurate standard errors without additional communication cost. We critically evaluate the information transfer for semi-honest collaborators and show that our protocol is secure against data reconstruction. Through both simulated and real-world examples we illustrate the functionality of our proposed algorithm. Without leaking information, our method performs as well on vertically partitioned data as existing methods on combined data -- all within mere minutes of computation time. We conclude that our method is a viable approach for vertically partitioned data analysis with a wide range of real-world applications., Comment: Fully reproducible code for all results and images can be found at https://github.com/vankesteren/privacy-preserving-glm, and the software package can be found at https://github.com/vankesteren/privreg

Latent variable models are well-known to suffer from rank deficiencies, causing problems with convergence and stability. Such problems are compounded in the "reduced-group split-ballot multitrait-multimethod model", which omits a set of moments from the estimation through a planned missing data design. This paper demonstrates the existence of rank deficiencies in this model and give the explicit null space. It also demonstrates that sample size and distance from the rank-deficient point interact in their effects on convergence, causing convergence to improve or worsen depending on both factors simultaneously. Furthermore, it notes that the latent variable correlations in the uncorrelated methods SB-MTMM model remain unaffected by the rank deficiency. I conclude that methodological experiments should be careful to manipulate both distance to known rank-deficiencies and sample size, and report all results, not only the apparently converged ones. Practitioners may consider that, even in the presence of nonconvergence or so-called "inadmissible" estimates, a subset of parameter estimates may still be consistent and stable.

Structural equation modeling (SEM) is a popular tool in the social and behavioural sciences, where it is being applied to ever more complex data types. The high-dimensional data produced by modern sensors, brain images, or (epi)genetic measurements require variable selection using parameter penalization; experimental models combining disparate data sources benefit from regularization to obtain a stable result; and genomic SEM or network models lead to alternative objective functions. With each proposed extension, researchers currently have to completely reformulate SEM and its optimization algorithm -- a challenging and time-consuming task. In this paper, we consider each SEM as a computation graph, a flexible method of specifying objective functions borrowed from the field of deep learning. When combined with state-of-the-art optimizers, our computation graph approach can extend SEM without the need for bespoke software development. We show that both existing and novel SEM improvements follow naturally from our approach. To demonstrate, we discuss least absolute deviation estimation and penalized regression models. We also introduce spike-and-slab SEM, which may perform better when shrinkage of large factor loadings is not desired. By applying computation graphs to SEM, we hope to greatly accelerate the process of developing SEM techniques, paving the way for new applications. We provide an accompanying R package tensorsem., Comment: R code and package are available online as supplementary material at https://github.com/vankesteren/sem-computationgraphs and https://github.com/vankesteren/tensorsem/tree/computationgraph, respectively

Social and behavioral scientists are increasingly employing technologies such as fMRI, smartphones, and gene sequencing, which yield 'high-dimensional' datasets with more columns than rows. There is increasing interest, but little substantive theory, in the role the variables in these data play in known processes. This necessitates exploratory mediation analysis, for which structural equation modeling is the benchmark method. However, this method cannot perform mediation analysis with more variables than observations. One option is to run a series of univariate mediation models, which incorrectly assumes independence of the mediators. Another option is regularization, but the available implementations may lead to high false positive rates. In this paper, we develop a hybrid approach which uses components of both filter and regularization: the 'Coordinate-wise Mediation Filter'. It performs filtering conditional on the other selected mediators. We show through simulation that it improves performance over existing methods. Finally, we provide an empirical example, showing how our method may be used for epigenetic research., Comment: R code and package are available online as supplementary material at https://github.com/vankesteren/cmfilter and https://github.com/vankesteren/ema_simulations

Binary data latent class models crucially assume local independence, violations of which can seriously bias the results. We present two tools for monitoring local dependence in binary data latent class models: the "Expected Parameter Change" (EPC) and a generalized EPC, estimating the substantive size and direction of possible local dependencies. The asymptotic and finite sample behavior of the measures is studied, and two applications to the U.S. Census estimation of Hispanic ethnicity and medical experts' ratings of x-rays demonstrate its value in arriving at a model that balances realism and parsimony., Comment: R code implementing our proposal and including both example datasets is available online as supplementary material

Over- and undertreatment harm patients and society and confound other healthcare quality measures. Despite a growing body of research covering specific conditions, we lack tools to systematically detect and measure over- and undertreatment in hospitals. We demonstrate a test used to monitor over- and undertreatment in Dutch hospitals, and illustrate its results applied to the aggregated administrative treatment data of 1,836,349 patients at 89 hospitals in 2013. We employ a random effects model to create risk-adjusted funnel plots that account for natural variation among hospitals, allowing us to estimate a measure of overtreatment and undertreatment when hospitals fall outside the control limits. The results of this test are not definitive, findings were discussed with hospitals to improve the model and to enable the hospitals to make informed treatment decisions.