Model Selection Strategy Research Articles

Morris Hansen Lecture 2023: Model Selection and Its Important Roles in Surveys

Abstract We discuss some fundamental issues regarding the roles that statistical models play in surveys. Many survey applications “borrow strength” from statistical models such as regression models and mixed effects models, even though these popular statistical models rarely hold exactly, if at all, in the real world. Yet, the idea of borrowing strength carries on, and the practice proves to be useful in many cases. We discuss basic ideas of borrowing strength via a statistical model and suggest practical criteria for evaluating a model, knowing that it is likely to be incorrect in a strict sense. Furthermore, we discuss a model selection strategy, known as the fence methods, which can incorporate practical interests into the selection criterion. Examples are used for illustration. Further remarks on challenging problems are offered.

Determination of soluble solids content in loquat using near-infrared spectroscopy coupled with broad learning system and hybrid wavelength selection strategy

Soluble solids content (SSC) serves as a crucial metric in assessing loquat quality. To achieve fast, non-destructive, and accurate detection of SSC, the combination of near-infrared (NIR) spectroscopy and broad learning system (BLS) model was developed in this study. Meanwhile, the influence of six spectral preprocessing methods and six wavelength selection algorithms on SSC prediction was explored, and the performance of the BLS model was compared with the widely used partial least squares regression (PLSR) and least squares support vector machine (LS-SVM) models. The results showed that Savitzky-Golay smoothing combined with standard normal transformation (SG-SNV) offered the optimal pretreatment, and the hybrid method, namely interval variable iterative space shrinking analysis and successive projections algorithm (iVISSA-SPA), emerged as the most effective wavelength selection method. The BLS model outperformed both PLSR and LS-SVM models. Specifically, the BLS model based on the iVISSA-SPA method achieved the optimal SSC prediction results, with RP2 = 0.8646, RMSEP = 0.5104, and RPD = 2.7481. Therefore, NIR spectroscopy coupled with the BLS model and hybrid wavelength selection strategy could rapidly, non-destructively and accurately detect the SSC of loquat, providing a viable alternative for SSC prediction in fruit.

Model selection of GLMMs in the analysis of count data in single-case studies: A Monte Carlo simulation.

Generalized linear mixed models (GLMMs) have great potential to deal with count data in single-case experimental designs (SCEDs). However, applied researchers have faced challenges in making various statistical decisions when using such advanced statistical techniques in their own research. This study focused on a critical issue by investigating the selection of an appropriate distribution to handle different types of count data in SCEDs due to overdispersion and/or zero-inflation. To achieve this, I proposed two model selection frameworks, one based on calculating information criteria (AIC and BIC) and another based on utilizing a multistage-model selection procedure. Four data scenarios were simulated including Poisson, negative binominal (NB), zero-inflated Poisson (ZIP), and zero-inflated negative binomial (ZINB). The same set of models (i.e., Poisson, NB, ZIP, and ZINB) were fitted for each scenario. In the simulation, I evaluated 10 model selection strategies within the two frameworks by assessing the model selection bias and its consequences on the accuracy of the treatment effect estimates and inferential statistics. Based on the simulation results and previous work, I provide recommendations regarding which model selection methods should be adopted in different scenarios. The implications, limitations, and future research directions are also discussed.

Implementation of deep learning models in predicting ESG index volatility

The consideration of environmental, social, and governance (ESG) aspects has become an integral part of investment decisions for individual and institutional investors. Most recently, corporate leaders recognized the core value of the ESG framework in fulfilling their environmental and social responsibility efforts. While stock market prediction is a complex and challenging task, several factors associated with developing an ESG framework further increase the complexity and volatility of ESG portfolios compared with broad market indices. To address this challenge, we propose an integrated computational framework to implement deep learning model architectures, specifically long short-term memory (LSTM), gated recurrent unit, and convolutional neural network, to predict the volatility of the ESG index in an identical environment. A comprehensive analysis was performed to identify a balanced combination of input features from fundamental data, technical indicators, and macroeconomic factors to delineate the cone of uncertainty in market volatility prediction. The performance of the constructed models was evaluated using standard assessment metrics. Rigorous hyperparameter tuning and model-selection strategies were implemented to identify the best model. Furthermore, a series of statistical analyses was conducted to validate the robustness and reliability of the model. Experimental results showed that a single-layer LSTM model with a relatively small number of neurons provides a superior fit with high prediction accuracy relative to more complex models.

Model selection for long-term load forecasting under uncertainty

Purpose The purpose of this study is to propose a systematic model selection procedure for long-term load forecasting (LTLF) for ex-ante and ex-post cases considering uncertainty in exogenous predictors. Design/methodology/approach The different variants of regression models, namely, Polynomial Regression (PR), Generalised Additive Model (GAM), Quantile Polynomial Regression (QPR) and Quantile Spline Regression (QSR), incorporating uncertainty in exogenous predictors like population, Real Gross State Product (RGSP) and Real Per Capita Income (RPCI), temperature and indicators of breakpoints and calendar effects, are considered for LTLF. Initially, the Backward Feature Elimination procedure is used to identify the optimal set of predictors for LTLF. Then, the consistency in model accuracies is evaluated using point and probabilistic forecast error metrics for ex-ante and ex-post cases. Findings From this study, it is found PR model outperformed in ex-ante condition, while QPR model outperformed in ex-post condition. Further, QPR model performed consistently across validation and testing periods. Overall, QPR model excelled in capturing uncertainty in exogenous predictors, thereby reducing over-forecast error and risk of overinvestment. Research limitations/implications These findings can help utilities to align model selection strategies with their risk tolerance. Originality/value To propose the systematic model selection procedure in this study, the consistent performance of PR, GAM, QPR and QSR models are evaluated using point forecast accuracy metrics Mean Absolute Percentage Error, Root Mean Squared Error and probabilistic forecast accuracy metric Pinball Score for ex-ante and ex-post cases considering uncertainty in the considered exogenous predictors such as RGSP, RPCI, population and temperature.

Temporal heterogeneity in the performance of machine learning models for PM2.5 concentration estimation

Machine learning (ML) methods have been applied extensively to simulate air pollutant concentrations and assess individual exposure in epidemiological studies. However, there is still a paucity of research on the temporal heterogeneity of ML model performance and the impact of dataset size. To explore the temporal heterogeneity in model performance when estimating daily concentrations of fine particulate matter (PM2.5) across China in 2021, we compared five decision tree-based ML models (Random Forest (RF), Categorical Boosting (CatBoost), Gradient Boost Regression Tree (GBRT), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)) across daily scales within three distinct timeframes. The performance of all models was evaluated using cross-validation. We observed that the performance of ML models varied with time, which showed a significant correlation with PM2.5 concentration. Among the 365 days in 2021, RF model performed best, the annual mean R2 was 0.86, with a minimum of 0.84 and a maximum of up to 0.95. For RF, we chose a cubic polynomial curve to fit the relationship between model performance and PM2.5 concentrations, and based on this, we devised a model selection strategy for different time scales, achieving an accuracy rate of up to 79.45 %, with the selected models having an average R2 of 0.85, and a maximum of up to 0.95. Additionally, we found that increasing the dataset size did not significantly improve model performance. Instead, it resulted in considerably longer runtime and increased memory usage. The methodology and findings of this study hold significant value for advancing the development of more efficient and precise modeling approaches for air pollutant concentrations. Furthermore, this research provides a foundation for regional air pollutant governance and future health-related research.

Ensemble distributional forecasting for insurance loss reserving

ABSTRACT Loss reserving generally focuses on identifying a single model that can generate superior predictive performance. However, different loss reserving models specialise in capturing different aspects of loss data. This is recognised in practice in the sense that results from different models are often considered, and sometimes combined. For instance, actuaries may take a weighted average of the prediction outcomes from various loss reserving models, often based on subjective assessments. In this paper, we propose a systematic framework to objectively combine (i.e. ensemble) multiple stochastic loss reserving models such that the strengths offered by different models can be utilised effectively. Our framework contains two main innovations compared to existing literature and practice. Firstly, our criteria model combination considers the full distributional properties of the ensemble and not just the central estimate – which is of particular importance in the reserving context. Secondly, our framework is that it is tailored for the features inherent to reserving data. These include, for instance, accident, development, calendar, and claim maturity effects. Crucially, the relative importance and scarcity of data across accident periods renders the problem distinct from the traditional ensembling techniques in statistical learning. Our framework is illustrated with a complex synthetic dataset. In the results, the optimised ensemble outperforms both (i) traditional model selection strategies, and (ii) an equally weighted ensemble. In particular, the improvement occurs not only with central estimates but also relevant quantiles, such as the 75th percentile of reserves (typically of interest to both insurers and regulators). The framework developed in this paper can be implemented thanks to an R package, ADLP, which is available from CRAN.

Experimental models for peri-implant diseases: a narrative review.

Peri-implant diseases, being the most common implant-related complications, significantly impact the normal functioning and longevity of implants. Experimental models play a crucial role in discovering potential therapeutic approaches and elucidating the mechanisms of disease progression in peri-implant diseases. This narrative review comprehensively examines animal models and common modeling methods employed in peri-implant disease research and innovatively summarizes the in vitro models of peri-implant diseases. Articles published between 2015 and 2023 were retrieved from PubMed/Medline, Web of Science, and Embase. All studies focusing on experimental models of peri-implant diseases were included and carefully evaluated. Various experimental models of peri-implantitis have different applications and advantages. The dog model is currently the most widely utilized animal model in peri-implant disease research, while rodent models have unique advantages in gene knockout and systemic disease induction. In vitro models of peri-implant diseases are also continuously evolving to meet different experimental purposes. The utilization of experimental models helps simplify experiments, save time and resources, and promote advances in peri-implant disease research. Animal models have been proven valuable in the early stages of drug development, while technological advancements have brought about more predictive and relevant in vitro models. This review provides clear and comprehensive model selection strategies for researchers in the field of peri-implant diseases, thereby enhancing understanding of disease pathogenesis and providing possibilities for developing new treatment strategies.

Do vocal mimics learn their mimetic songs from heterospecifics or conspecifics?

Abstract Vocal mimicry can develop either through imitation of heterospecific models (primary mimicry) or through imitation of conspecifics already producing mimetic song (secondary mimicry). Distinguishing primary versus secondary mimicry is important from ecological, evolutionary, and neurobiological perspectives. We outline four empirical strategies for researchers to detect the usage of secondary mimicry: (1) model selection strategy (heterospecific model commonly mimicked but rare or absent locally), (2) usage frequency strategy (syllable types commonly mimicked but rarely used by the model), (3) acoustic structure strategy (acoustic divergence in how a syllable type gets produced by mimics versus models), and (4) syntax strategy (consistent variation between mimic and model in the ordering of syllable types). We then use these strategies to produce evidence for secondary mimicry in northern mockingbirds (Mimus polyglottos), focusing mostly on the mimicked songs of northern cardinals (Cardinalis cardinalis) and tufted titmice (Baeolophus bicolor). As further evidence of the mechanisms by which secondary mimicry might occur, we also demonstrate that mockingbirds match countersing much more frequently in response to conspecifics versus heterospecifics. Deeper questions about the precise extent of primary versus secondary mimicry in mockingbirds and other mimicking species still need answers.

Text Classification in Natural Language Processing

This paper presents an overview of text classification techniques, focusing on the pre-processing steps, feature extraction methods, and model selection strategies employed in the process. Algorithms such as Naive Bayes, Support Vector Machines (SVM), logistic regression, and neural networks are used. Furthermore, recent advancements in deep learning models for text classification, including Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are used. Comprehensive understanding of text classification methodologies in NLP and insights into current trends and challenges in the field are mentioned

GlmmPen: High Dimensional Penalized Generalized Linear Mixed Models.

Generalized linear mixed models (GLMMs) are widely used in research for their ability to model correlated outcomes with non-Gaussian conditional distributions. The proper selection of fixed and random effects is a critical part of the modeling process, where model misspecification may lead to significant bias. However, the joint selection of fixed and random effects has historically been limited to lower dimensional GLMMs, largely due to the use of criterion-based model selection strategies. Here we present the R package glmmPen, one of the first to select fixed and random effects in higher dimension using a penalized GLMM modeling framework. Model parameters are estimated using a Monte Carlo expectation conditional minimization (MCECM) algorithm, which leverages Stan and RcppArmadillo for increased computational efficiency. Our package supports the Binomial, Gaussian, and Poisson families and multiple penalty functions. In this manuscript we discuss the modeling procedure, estimation scheme, and software implementation through application to a pancreatic cancer subtyping study. Simulation results show our method has good performance in selecting both the fixed and random effects in high dimensional GLMMs.

Innovative hybrid prediction method integrating wavelet threshold decomposition and entropy-based model selection strategy for building energy consumption prediction

Building energy prediction plays a pivotal role in the building energy management system, and an efficient prediction method is crucial to ensure the effective operation of building systems. Existing research enhances prediction accuracy by combining decomposition algorithms with prediction methods to separate non-stationary information from the original sequence. However, the decomposition process and the subsequent prediction of multiple decomposed subsequences result in significant time overhead, making it unsuitable for scenarios with high temporal requirements, such as short-term energy consumption forecasting. Additionally, a single prediction model struggles to adapt to the diverse features of decomposed subsequences, creating prediction accuracy and efficiency bottlenecks. Addressing these challenges, this study proposes a hybrid prediction method that combines wavelet threshold decomposition (WTD) with an entropy-based model selection strategy. The method initially utilizes the efficient WTD algorithm to decompose the original energy consumption data into three fixed subsequences. Subsequently, the permutation entropy for each subsequence is computed, and based on the comprehensive performance of four prediction models (ANN, LSTM, CNN-LSTM, and TCN) on sequences with different entropy values, the optimal prediction model is selected. Finally, the prediction results of the subsequences are aggregated to obtain the ultimate prediction outcome. Experimental results demonstrate that the proposed method significantly outperforms baseline methods in both prediction speed and accuracy across five real building energy consumption datasets with distinct operational modes.

Novel wind power ensemble forecasting system based on mixed-frequency modeling and interpretable base model selection strategy

Accurate wind power forecasting helps to maximize the utilization of wind energy resources, enhance wind power generation efficiency, and optimize grid operation. This study proposes an innovative mixed-frequency modeling and interpretable base model selection-based ensemble wind power forecasting system. Specifically, the data preprocessing module preprocesses wind speed and wind power data at different frequencies. The mixed-frequency modeling module then constructs 12 mixed-frequency and machine learning models to predict wind power, with a comprehensive evaluation metric to determine their optimal lags. Subsequently, the base model selection module effectively combines the elastic net and Shapley additive explanation methods to identify individual models that contribute significantly to the prediction target as base models. Finally, the ensemble module integrates the optimization algorithms with a machine learning model to ensemble the selected base models. The key findings are as follows: (1) mixed-frequency wind speed and wind power data effectively improve forecasting performance, and (2) the proposed base model selection strategy greatly enhances the accuracy and interpretability of the modeling process. This model could robustly predict two datasets from Inner Mongolian wind farms, with average absolute percentage errors of 2.4505% and 4.8270%, respectively, establishing this as a useful technique for wind power prediction.

Estimation of the Biogeochemical and Physical Properties of Lakes Based on Remote Sensing and Artificial Intelligence Applications

Lakes play a crucial role in the global biogeochemical cycles through the transport, storage, and transformation of different biogeochemical compounds. Their regulatory service appears to be disproportionately important relative to their small areal extent, necessitating continuous monitoring. This study leverages the potential of optical remote sensing sensors, specifically Sentinel-2 Multispectral Imagery (MSI), to monitor and predict water quality parameters in lakes. Optically active parameters, such as chlorophyll a (CHL), total suspended matter (TSM), and colored dissolved matter (CDOM), can be directly detected using optical remote sensing sensors. However, the challenge lies in detecting non-optically active substances, which lack direct spectral characteristics. The capabilities of artificial intelligence applications can be used in the identification of optically non-active compounds from remote sensing data. This study aims to employ a machine learning approach (combining the Genetic Algorithm (GA) and Extreme Gradient Boost (XGBoost)) and in situ and Sentinel-2 Multispectral Imagery data to construct inversion models for 16 physical and biogeochemical water quality parameters including CHL, CDOM, TSM, total nitrogen (TN), total phosphorus (TP), phosphate (PO4), sulphate, ammonium nitrogen, 5-day biochemical oxygen demand (BOD5), chemical oxygen demand (COD), and the biomasses of phytoplankton and cyanobacteria, pH, dissolved oxygen (O2), water temperature (WT) and transparency (SD). GA_XGBoost exhibited strong predictive capabilities and it was able to accurately predict 10 biogeochemical and 2 physical water quality parameters. Additionally, this study provides a practical demonstration of the developed inversion models, illustrating their applicability in estimating various water quality parameters simultaneously across multiple lakes on five different dates. The study highlights the need for ongoing research and refinement of machine learning methodologies in environmental monitoring, particularly in remote sensing applications for water quality assessment. Results emphasize the need for broader temporal scopes, longer-term datasets, and enhanced model selection strategies to improve the robustness and generalizability of these models. In general, the outcomes of this study provide the basis for a better understanding of the role of lakes in the biogeochemical cycle and will allow the formulation of reliable recommendations for various applications used in the studies of ecology, water quality, the climate, and the carbon cycle.

Residual LSTM-based short duration forecasting of polarization current for effective assessment of transformers insulation

The empirical application of polarization and depolarization current (PDC) measurement of transformers facilitates the extraction of critical insulation-sensitive parameters. This technique, rooted in time-domain dielectric response analysis, forms the bedrock for parameterization and insulation modeling. However, the inherently time-consuming nature of polarization current measurements renders them susceptible to data corruption. This article explores deep-learning-based short-duration techniques for forecasting polarization current to address this limitation. By incorporating spatial shortcuts, the residual long short-term memory (LSTM) network facilitates the seamless propagation of spatial and temporal gradients. Furthermore, the relative forecasting assessment of the proposed residual LSTM model’s performance is made against traditional LSTM, attention LSTM, gated recurrent units (GRU), and convolutional neural network (CNN) models. Thus, optimal model selection strategies are evaluated based on their capability to capture extended dependencies and short-term information present in the data. In addition, the Monte Carlo dropout prediction is employed to estimate uncertainty in polarization current forecasts. The findings demonstrate that the proposed residual LSTM network model for polarization current forecasting yields the lowest error metrics and maintains prediction consistency over the testing duration. Thus, the proposed approach significantly reduces PDC measurement time, providing an effective means to develop proactive maintenance strategies for evaluating the insulation condition of transformers.

Automatic detection of myocardial ischaemia using generalisable spatio-temporal hierarchical Bayesian modelling of DCE-MRI

Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE-MRI) can be used as a non-invasive method for the assessment of myocardial perfusion. The acquired images can be utilised to analyse the spatial extent and severity of myocardial ischaemia (regions with impaired microvascular blood flow). In the present paper, we propose a novel generalisable spatio-temporal hierarchical Bayesian model (GST-HBM) to automate the detection of ischaemic lesions and improve the in silico prediction accuracy by systematically integrating spatio-temporal context information. We present a computational inference procedure with an adequate trade-off between accuracy and computational efficiency, whereby model parameters are sampled from the posterior distribution with Gibbs sampling, while lower-level hyperparameters are selected using model selection strategies based on the Watanabe Akaike information criterion (WAIC). We have assessed our method on both synthetic (in silico) data with known gold-standard and 12 sets of clinical first-pass myocardial perfusion DCE-MRI datasets. We have also carried out a comparative performance evaluation with four established alternative methods: Gaussian mixture model (GMM), opening and closing operations based on Gaussian mixture model (GMMC&Omax), Markov random field constrained Gaussian mixture model (GMM-MRF) and model-based hierarchical Bayesian model (M-HBM). Our results show that the proposed GST-HBM method achieves much higher in silico prediction accuracy than the established alternative methods. Furthermore, this method appears to provide a more robust delineation of ischaemic lesions in datasets affected by spatially variant noise.

Deep Learning-Based Bitcoin Price Prediction Using Long Short-Term Memory Networks

Cryptocurrencies, notably Bitcoin, have surged in popularity and importance, attracting widespread attention due to their volatile nature and potential for significant financial gains. Predicting Bitcoin prices accurately has emerged as a challenging yet indispensable task for investors, traders, and policymakers seeking to navigate the complexities of the cryptocurrency market. Traditional forecasting methods often struggle to capture the intricate temporal patterns inherent in this domain, prompting the exploration of advanced techniques such as deep learning. This paper presents a comprehensive study on employing Long Short-Term Memory (LSTM) networks for Bitcoin price prediction, a variant of Recurrent Neural Networks (RNNs). Our research encompasses the entire process, from data collection to model evaluation, focusing on addressing the unique challenges posed by cryptocurrency market data. We begin by discussing the methodology involved in data acquisition, preprocessing, and feature engineering, which are crucial steps for ensuring the quality and relevance of input data. Next, we delve into the intricacies of LSTM model selection, hyperparameter tuning, and training strategies, aiming to optimize predictive performance while mitigating overfitting. Through extensive experimentation on real-world Bitcoin price datasets spanning multiple periods, we rigorously evaluate the effectiveness of LSTM networks in capturing long-term dependencies and forecasting future price movements. Comparative analyses against traditional time-series forecasting methods underscore the superior predictive capabilities of deep learning approaches in this context. Furthermore, we investigate the interpretability of LSTM-based models, examining the insights gained from analyzing learned representations and hidden states. These analyses provide valuable perspectives for market participants seeking to understand the driving factors behind Bitcoin price dynamics. Our findings demonstrate the practical utility of LSTM networks in Bitcoin price prediction and contribute to advancing the broader discourse on deep learning applications in financial forecasting. By shedding light on the potential of deep learning techniques to enhance predictive accuracy and robustness in cryptocurrency markets, our research offers valuable insights for stakeholders navigating this rapidly evolving landscape.

Unsupervised Spectral Demosaicing With Lightweight Spectral Attention Networks.

This paper presents a deep learning-based spectral demosaicing technique trained in an unsupervised manner. Many existing deep learning-based techniques relying on supervised learning with synthetic images, often underperform on real-world images, especially as the number of spectral bands increases. This paper presents a comprehensive unsupervised spectral demosaicing (USD) framework based on the characteristics of spectral mosaic images. This framework encompasses a training method, model structure, transformation strategy, and a well-fitted model selection strategy. To enable the network to dynamically model spectral correlation while maintaining a compact parameter space, we reduce the complexity and parameters of the spectral attention module. This is achieved by dividing the spectral attention tensor into spectral attention matrices in the spatial dimension and spectral attention vector in the channel dimension. This paper also presents Mosaic 25 , a real 25-band hyperspectral mosaic image dataset featuring various objects, illuminations, and materials for benchmarking purposes. Extensive experiments on both synthetic and real-world datasets demonstrate that the proposed method outperforms conventional unsupervised methods in terms of spatial distortion suppression, spectral fidelity, robustness, and computational cost. Our code and dataset are publicly available at https://github.com/polwork/Unsupervised-Spectral-Demosaicing.

Database Native Model Selection: Harnessing Deep Neural Networks in Database Systems

The growing demand for advanced analytics beyond statistical aggregation calls for database systems that support effective model selection of deep neural networks (DNNs). However, existing model selection strategies are based on either training-based algorithms that deliver high-performing models at the expense of high computational cost, or training-free algorithms that enhance computational efficiency with reduced effectiveness. These strategies often disregard computational cost and response time Service-Level Objectives (SLOs), which are of concern to average or budget-conscious machine learning users. In addition, they lack a well-designed integration of the model selection algorithms with DBMSs, which hinders efficient in-database model selection. This paper presents TRAILS, a resource-efficient and SLO-aware in-database model selection system. To leverage the strengths of both training-free and training-based model selection, we first characterize nine state-of-the-art training-free model evaluation metrics and propose a more effective one named JacFlow, and then, restructure the conventional model selection procedure into two phases: filtering and refinement. A novel coordinator is also introduced to strike a balance between the high efficiency of train-free algorithms and the high effectiveness of training-based algorithms, ensuring high-performing model selection while adhering to target SLOs. Moreover, we incorporate the proposed algorithm into PostgreSQL to develop TRAILS, thereby both enhancing resource efficiency and reducing model selection latency. This integration establishes a foundation for declarative model definition and selection within DBMSs. Empirical results demonstrate that our TRAILS reduces model selection time and computational expenses considerably by up to 24.38x and 29.32x respectively compared to existing model selection systems.

High-throughput target trial emulation for Alzheimer’s disease drug repurposing with real-world data

Target trial emulation is the process of mimicking target randomized trials using real-world data, where effective confounding control for unbiased treatment effect estimation remains a main challenge. Although various approaches have been proposed for this challenge, a systematic evaluation is still lacking. Here we emulated trials for thousands of medications from two large-scale real-world data warehouses, covering over 10 years of clinical records for over 170 million patients, aiming to identify new indications of approved drugs for Alzheimer’s disease. We assessed different propensity score models under the inverse probability of treatment weighting framework and suggested a model selection strategy for improved baseline covariate balancing. We also found that the deep learning-based propensity score model did not necessarily outperform logistic regression-based methods in covariate balancing. Finally, we highlighted five top-ranked drugs (pantoprazole, gabapentin, atorvastatin, fluticasone, and omeprazole) originally intended for other indications with potential benefits for Alzheimer’s patients.