Semi-synthetic Data Research Articles

A Machine Learning Framework for Assessing Experts’ Decision Quality

Expert workers make non-trivial decisions with significant implications. Experts’ decision accuracy is, thus, a fundamental aspect of their judgment quality, key to both management and consumers of experts’ services. Yet, in many important settings, transparency in experts’ decision quality is rarely possible because ground truth data for evaluating the experts’ decisions is costly and available only for a limited set of decisions. Furthermore, different experts typically handle exclusive sets of decisions, and thus, prior solutions that rely on the aggregation of multiple experts’ decisions for the same instance are inapplicable. We first formulate the problem of estimating experts’ decision accuracy in this setting and then develop a machine–learning–based framework to address it. Our method effectively leverages both abundant historical data on workers’ past decisions and scarce decision instances with ground truth labels. Using both semi-synthetic data based on publicly available data sets and purposefully compiled data sets on real workers’ decisions, we conduct extensive empirical evaluations of our method’s performance relative to alternatives. The results show that our approach is superior to existing alternatives across diverse settings, including settings that involve different data domains, experts’ qualities, and amounts of ground truth data. To our knowledge, this paper is the first to posit and address the problem of estimating experts’ decision accuracies from historical data with scarce ground truth, and it is the first to offer comprehensive results for this problem setting, establishing the performances that can be achieved across settings as well as the state-of-the-art performance on which future work can build. This paper was accepted by Anindya Ghose, information systems. Funding: T. Geva acknowledges research grants from the Jeremy Coller Foundation and from the Henry Crown Institute for Business Research. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2021.03357 .

Knowledge Transfer from High-Resource to Low-Resource Programming Languages for Code LLMs

Over the past few years, Large Language Models of Code (Code LLMs) have started to have a significant impact on programming practice. Code LLMs are also emerging as building blocks for research in programming languages and software engineering. However, the quality of code produced by a Code LLM varies significantly by programming language. Code LLMs produce impressive results on high-resource programming languages that are well represented in their training data (e.g., Java, Python, or JavaScript), but struggle with low-resource languages that have limited training data available (e.g., OCaml, Racket, and several others). This paper presents an effective approach for boosting the performance of Code LLMs on low-resource languages using semi-synthetic data. Our approach, called MultiPL-T, generates high-quality datasets for low-resource languages, which can then be used to fine-tune any pretrained Code LLM. MultiPL-T translates training data from high-resource languages into training data for low-resource languages in the following way. 1) We use a Code LLM to synthesize unit tests for commented code from a high-resource source language, filtering out faulty tests and code with low test coverage. 2) We use a Code LLM to translate the code from the high-resource source language to a target low-resource language. This gives us a corpus of candidate training data in the target language, but many of these translations are wrong. 3) We use a lightweight compiler to compile the test cases generated in (1) from the source language to the target language, which allows us to filter our obviously wrong translations. The result is a training corpus in the target low-resource language where all items have been validated with test cases. We apply this approach to generate tens of thousands of new, validated training items for five low-resource languages: Julia, Lua, OCaml, R, and Racket, using Python as the source high-resource language. Furthermore, we use an open Code LLM (StarCoderBase) with open training data (The Stack), which allows us to decontaminate benchmarks, train models without violating licenses, and run experiments that could not otherwise be done. Using datasets generated with MultiPL-T, we present fine-tuned versions of StarCoderBase and Code Llama for Julia, Lua, OCaml, R, and Racket that outperform other fine-tunes of these base models on the natural language to code task. We also present Racket fine-tunes for two very recent models, DeepSeek Coder and StarCoder2, to show that MultiPL-T continues to outperform other fine-tuning approaches for low-resource languages. The MultiPL-T approach is easy to apply to new languages, and is significantly more efficient and effective than alternatives such as training longer.

Enhancing Resilience in Biometric Research: Generation of 3D Synthetic Face Data Using Advanced 3D Character Creation Techniques from High-Fidelity Video Games and Animation.

Biometric authentication plays a vital role in various everyday applications with increasing demands for reliability and security. However, the use of real biometric data for research raises privacy concerns and data scarcity issues. A promising approach using synthetic biometric data to address the resulting unbalanced representation and bias, as well as the limited availability of diverse datasets for the development and evaluation of biometric systems, has emerged. Methods for a parameterized generation of highly realistic synthetic data are emerging and the necessary quality metrics to prove that synthetic data can compare to real data are open research tasks. The generation of 3D synthetic face data using game engines' capabilities of generating varied realistic virtual characters is explored as a possible alternative for generating synthetic face data while maintaining reproducibility and ground truth, as opposed to other creation methods. While synthetic data offer several benefits, including improved resilience against data privacy concerns, the limitations and challenges associated with their usage are addressed. Our work shows concurrent behavior in comparing semi-synthetic data as a digital representation of a real identity with their real datasets. Despite slight asymmetrical performance in comparison with a larger database of real samples, a promising performance in face data authentication is shown, which lays the foundation for further investigations with digital avatars and the creation and analysis of fully synthetic data. Future directions for improving synthetic biometric data generation and their impact on advancing biometrics research are discussed.

PathIntegrate: Multivariate modelling approaches for pathway-based multi-omics data integration.

As terabytes of multi-omics data are being generated, there is an ever-increasing need for methods facilitating the integration and interpretation of such data. Current multi-omics integration methods typically output lists, clusters, or subnetworks of molecules related to an outcome. Even with expert domain knowledge, discerning the biological processes involved is a time-consuming activity. Here we propose PathIntegrate, a method for integrating multi-omics datasets based on pathways, designed to exploit knowledge of biological systems and thus provide interpretable models for such studies. PathIntegrate employs single-sample pathway analysis to transform multi-omics datasets from the molecular to the pathway-level, and applies a predictive single-view or multi-view model to integrate the data. Model outputs include multi-omics pathways ranked by their contribution to the outcome prediction, the contribution of each omics layer, and the importance of each molecule in a pathway. Using semi-synthetic data we demonstrate the benefit of grouping molecules into pathways to detect signals in low signal-to-noise scenarios, as well as the ability of PathIntegrate to precisely identify important pathways at low effect sizes. Finally, using COPD and COVID-19 data we showcase how PathIntegrate enables convenient integration and interpretation of complex high-dimensional multi-omics datasets. PathIntegrate is available as an open-source Python package.

Deep learning models for crack segmentation in 3d images of concrete trained on semi-synthetic data

Deep learning models for crack segmentation in 3d images of concrete trained on semi-synthetic data

Semi-supervised ROC analysis for reliable and streamlined evaluation of phenotyping algorithms.

High-throughput phenotyping will accelerate the use of electronic health records (EHRs) for translational research. A critical roadblock is the extensive medical supervision required for phenotyping algorithm (PA) estimation and evaluation. To address this challenge, numerous weakly-supervised learning methods have been proposed. However, there is a paucity of methods for reliably evaluating the predictive performance of PAs when a very small proportion of the data is labeled. To fill this gap, we introduce a semi-supervised approach (ssROC) for estimation of the receiver operating characteristic (ROC) parameters of PAs (eg, sensitivity, specificity). ssROC uses a small labeled dataset to nonparametrically impute missing labels. The imputations are then used for ROC parameter estimation to yield more precise estimates of PA performance relative to classical supervised ROC analysis (supROC) using only labeled data. We evaluated ssROC with synthetic, semi-synthetic, and EHR data from Mass General Brigham (MGB). ssROC produced ROC parameter estimates with minimal bias and significantly lower variance than supROC in the simulated and semi-synthetic data. For the 5 PAs from MGB, the estimates from ssROC are 30% to 60% less variable than supROC on average. ssROC enables precise evaluation of PA performance without demanding large volumes of labeled data. ssROC is also easily implementable in open-source R software. When used in conjunction with weakly-supervised PAs, ssROC facilitates the reliable and streamlined phenotyping necessary for EHR-based research.

THE UTILIZATION OF SYNTHETIC AND SEMISYNTHETIC POINT CLOUDS AND IMAGES FOR TESTING NOVEL APPROACHES FOR CORRECTING LIDAR DATA

Abstract. The paper presents the application of lidar data and photo datasets, external orientation parameters (EOPs), ground control points (GCPs), and check points for testing new methods of geometric lidar data correction. These datasets are utilized to validate novel approaches such as altimetric deformation methods based on stereo models or lidargrammetric methods that utilize image matching and specialized lidar data formats. The paper presents specific use cases of these data as examples of two tested processes. After describing these processes, the methods of synthetic and semisynthetic data simulation are presented. The simulation is directed and subordinated to the aspects of the new method being tested. The data must be used for testing starting from basic functionality up to specific and untypical cases of new method application. By presenting specific cases of the application of synthetic and semisynthetic data, the paper introduces the general idea of benchmarking based on synthetic and semisynthetic data as another means of validating new methods. These artificially generated datasets provide a controlled environment for evaluating the effectiveness of new methods to be investigated.

Road safety evaluation with multiple treatments: A comparison of methods based on simulations

This paper focuses on ex-post road safety evaluation with multiple treatments. The potential outcome framework for causal inference is introduced to formalize the causal estimands of interest. Various estimation methods are compared via performing simulation experiments based on semi-synthetic data constructed from a London 20 mph zones dataset. The methods under evaluation include regressions, propensity score (PS) based methods, and a machine learning-based method termed generalized random forests (GRF). Both PS-based methods and GRF show higher flexibility with respect to functional specifications of outcome models. Moreover, GRF shows great superiority in the cases where road safety treatments are assigned following specific criteria and/or where there are heterogeneous treatment effects. Considering the ex-post evaluation of combined effects of multiple treatments has significant practical value, the potential outcome framework and the estimation methods presented in this paper are highly recommended for road safety studies.

Use of semi-synthetic data for catheter segmentation improvement.

In the era of data-driven machine learning algorithms, data is the new oil. For the most optimal results, datasets should be large, heterogeneous and, crucially, correctly labeled. However, data collection and labeling are time-consuming and labor-intensive processes. In the field of medical device segmentation, present during minimally invasive surgery, this leads to a lack of informative data. Motivated by this drawback, we developed an algorithm generating semi-synthetic images based on real ones. The concept of this algorithm is to place a randomly shaped catheter in an empty heart cavity, where the shape of the catheter is generated by forward kinematics of continuum robots. Having implemented the proposed algorithm, we generated new images of heart cavities with various artificial catheters. We compared the results of deep neural networks trained purely on real datasets with respect to networks trained on both real and semi-synthetic datasets, highlighting that semi-synthetic data improves catheter segmentation accuracy. A modified U-Net trained on combined datasets performed the segmentation with a Dice similarity coefficient of 92.6±2.2%, while the same model trained only on real images achieved a Dice similarity coefficient of 86.5±3.6%. Therefore, using semi-synthetic data allows for the decrease of accuracy spread, improves model generalization, reduces subjectivity, shortens the labeling routine, increases the number of samples, and improves the heterogeneity.

On the Estimation of Spatial Density From Mobile Network Operator Data

We tackle the problem of estimating the spatial distribution of mobile phones from Mobile Network Operator (MNO) data, namely CDR or signalling data. The process of transforming MNO data to a density map requires to perform geolocation of radio cells to determine their spatial footprint. Traditional geolocation solutions rely on Voronoi tessellations and approximate cell footprints by mutually disjoint regions. Recently, some pioneering work started to consider more elaborated geolocation methods with partially overlapping (non-disjont) cell footprints coupled with a probabilistic model for phone-to-cell association. Estimating the spatial density in such a probabilistic setup is currently an open research problem and is the focus of the present work. We start by reviewing three different estimation methods proposed in literature and provide novel analytical insights that unveil some key aspects of their mutual relationships and properties. Furthermore, we develop a novel estimation approach for which a closed-form solution can be given. Numerical results based on semi-synthetic data are presented to assess the relative accuracy of each method. Our results indicate that the estimators based on overlapping cells have the potential to improve spatial accuracy over traditional approaches based on Voronoi tessellations.

A Grey Literature Review on Data Stream Processing applications testing

Context:The Data Stream Processing (DSP) approach focuses on real-time data processing by applying specific techniques for capturing and processing relevant data for on-the-fly results, i.e. without necessarily requiring prior storage. Like in any other software, testing plays a vital role in the quality assurance of DSP applications. However, testing such kind of software is not a simple task. In this context, some factors that make challenging testing are message temporality, parallelism, data volume, complex infrastructure, variability, and speed of messages. Objective:This work aims to map and synthesize industry knowledge and experience regarding DSP application testing. Specifically, we want to know about challenges, test purposes, test approaches, test data sources, and adopted tools. Method:To achieve the objective, we performed a Grey Literature Review (e.g., blog posts, white papers, discussion lists, lecture themes at technical events, professional social networks, software repositories, and other web-published) on testing DSP applications. We searched the grey literature using Google’s regular search engine in addition to specific searches on technical software development content websites. The selected studies were analyzed using qualitative and quantitative techniques. Results:Results are based on evidence from 154 selected sources. The challenges for testing DSP applications are the complexity of DSP applications, test infrastructure complexity, timing, and data acquisition issues. The main test objectives identified are functional suitability, performance efficiency, reliability, and maintainability. The main test approaches reported: Performance Testing, Regression Testing, Property-Based Testing, Chaos Testing, and Contract/Schema Testing. The strategies adopted by practitioners to obtain test data: Historical Data, Production Data Mirroring, Semi-Synthetic Data, and Synthetic Data. We also report 50 tools used in various testing activities, which are used for: automating infrastructure, generating test data, test utilities, dealing with timing issues, load generation, simulation, and others. Furthermore, we identified gaps and opportunities for future scientific work. Conclusion:This work selected and summarized content produced by practitioners regarding DSP application testing. We identified that knowledge, techniques, and tools intrinsic to the practice were not present in the formal literature, so this study helps reduce the gap between industry and academia on this topic. The document has delivered benefits to industry practitioners and academic researchers.

A Framework to Maximize Group Fairness for Workers on Online Labor Platforms

As the number of online labor platforms and the diversity of jobs on these platforms increase, ensuring group fairness for workers needs to be the focus of job-matching services. Risk of discrimination against workers occurs in two different job-matching services: when someone is looking for a job (i.e., a job seeker) and when someone wants to deploy jobs (i.e., a job provider). To maximize their chances of getting hired, job seekers submit their profiles on different platforms. Similarly, job providers publish their job offers on multiple platforms with the goal of reaching a wide and diverse workforce. In this paper, we propose a theoretical framework to maximize group fairness for workers 1) when job seekers are looking for jobs on multiple platforms, and 2) when jobs are being deployed by job providers on multiple platforms. We formulate each goal as different optimization problems with different constraints, prove most of them are computationally hard to solve and propose various efficient algorithms to solve all of them in reasonable time. We then design a series of experiments that rely on synthetic and semi-synthetic data generated from a real-world online labor platform to evaluate our framework.

The K-mer antibiotic resistance gene variant analyzer (KARGVA).

Characterization of antibiotic resistance genes (ARGs) from high-throughput sequencing data of metagenomics and cultured bacterial samples is a challenging task, with the need to account for both computational (e.g., string algorithms) and biological (e.g., gene transfers, rearrangements) aspects. Curated ARG databases exist together with assorted ARG classification approaches (e.g., database alignment, machine learning). Besides ARGs that naturally occur in bacterial strains or are acquired through mobile elements, there are chromosomal genes that can render a bacterium resistant to antibiotics through point mutations, i.e., ARG variants (ARGVs). While ARG repositories also collect ARGVs, there are only a few tools that are able to identify ARGVs from metagenomics and high throughput sequencing data, with a number of limitations (e.g., pre-assembly, a posteriori verification of mutations, or specification of species). In this work we present the k-mer, i.e., strings of fixed length k, ARGV analyzer - KARGVA - an open-source, multi-platform tool that provides: (i) an ad hoc, large ARGV database derived from multiple sources; (ii) input capability for various types of high-throughput sequencing data; (iii) a three-way, hash-based, k-mer search setup to process data efficiently, linking k-mers to ARGVs, k-mers to point mutations, and ARGVs to k-mers, respectively; (iv) a statistical filter on sequence classification to reduce type I and II errors. On semi-synthetic data, KARGVA provides very high accuracy even in presence of high sequencing errors or mutations (99.2 and 86.6% accuracy within 1 and 5% base change rates, respectively), and genome rearrangements (98.2% accuracy), with robust performance on ad hoc false positive sets. On data from the worldwide MetaSUB consortium, comprising 3,700+ metagenomics experiments, KARGVA identifies more ARGVs than Resistance Gene Identifier (4.8x) and PointFinder (6.8x), yet all predictions are below the expected false positive estimates. The prevalence of ARGVs is correlated to ARGs but ecological characteristics do not explain well ARGV variance. KARGVA is publicly available at https://github.com/DataIntellSystLab/KARGVA under MIT license.

An improved semi-synthetic approach for creating visual-inertial odometry datasets

Capturing outdoor visual-inertial datasets is a challenging yet vital aspect of developing robust visual-inertial odometry (VIO) algorithms. A significant hurdle is that high-accuracy-ground-truth systems (e.g., motion capture) are not practical for outdoor use. One solution is to use a “semi-synthetic” approach that combines rendered images with real IMU data. This approach can produce sequences containing challenging imagery and accurate ground truth but with less simulated data than a fully synthetic sequence. Existing methods (used by popular tools/datasets) record IMU measurements from a visual-inertial system while measuring its trajectory using motion capture, then rendering images along that trajectory. This work identifies a major flaw in that approach, specifically that using motion capture alone to estimate the pose of the robot/system results in the generation of inconsistent visual-inertial data that is not suitable for evaluating VIO algorithms. However, we show that it is possible to generate high-quality semi-synthetic data for VIO algorithm evaluation. We do so using an open-source full-batch optimisation tool to incorporate both mocap and IMU measurements when estimating the IMU’s trajectory. We demonstrate that this improved trajectory results in better consistency between the IMU data and rendered images and that the resulting data improves VIO trajectory error by 79% compared to existing methods. Furthermore, we examine the effect of visual-inertial data inconsistency (as a result of trajectory noise) on VIO performance to provide a foundation for future work targeting real-time applications.

Optimizing the preventive maintenance frequency with causal machine learning

Maintenance is a challenging operational problem where the goal is to plan sufficient preventive maintenance (PM) to avoid asset overhauls and failures. Existing work typically relies on strong assumptions (1) to model the asset’s overhaul and failure rate, assuming a stochastic process with known hazard rate, (2) to model the effect of PM on this hazard rate, assuming the effect is deterministic or governed by a known probability distribution, and (3) by not taking asset-specific characteristics into account, but assuming homogeneous hazard rates and PM effects. Instead of relying on these assumptions to model the problem, this work uses causal inference to learn the effect of the PM frequency on the overhaul and failure rate, conditional on the asset’s characteristics, from observational data. Based on these learned outcomes, we can optimize each asset’s PM frequency to minimize the combined cost of failures, overhauls, and preventive maintenance. We validate our approach on real-life data of more than 4000 maintenance contracts from an industrial partner. Empirical results on semi-synthetic data show that our methodology based on causal machine learning results in individualized maintenance schedules that are more accurate and cost-effective than a non-causal approach that does not deal with selection bias and a non-individualized approach that prescribes the same PM frequency to all machines.

Test-data generation and integration for long-distance e-vehicle routing

Advanced route planning algorithms are one of the key enabling technologies for emerging electric and autonomous mobility. Large realistic data sets are needed to test such algorithms under conditions that capture natural time-varying traffic patterns and corresponding travel-time and energy-use predictions. Further, the time-varying availability of charging infrastructure and vehicle-specific charging-power curves may be necessary to support advanced planning. While some data sets and synthetic data generators capture some of the aspects mentioned above, no integrated testbeds include all of them. We contribute with a modular testbed architecture. First, it includes a semi-synthetic data generator that uses a state-of-the-art traffic simulator, real traffic volume distribution patterns, EV-specific data, and elevation data. These elements support the generation of time-dependent travel-time and energy-use weights in a road-network graph. The generator ensures that the data satisfies the FIFO property, which is essential for time-dependent routing. Next, the testbed provides a thin layer of services that can serve as building blocks for future advanced routing algorithms. The experimental study demonstrates that the testbed can reproduce travel-time and energy-use patterns for long-distance trips similar to commercially available services.

An improved semi-synthetic approach for creating visual-inertial odometry datasets

Capturing outdoor visual-inertial datasets is a challenging yet vital aspect of developing robust visual-inertial odometry (VIO) algorithms. A significant hurdle is that high-accuracy-ground-truth systems (e.g., motion capture) are not practical for outdoor use. One solution is to use a “semi-synthetic” approach that combines rendered images with real IMU data. This approach can produce sequences containing challenging imagery and accurate ground truth but with less simulated data than a fully synthetic sequence. Existing methods (used by popular tools/datasets) record IMU measurements from a visual-inertial system while measuring its trajectory using motion capture, then rendering images along that trajectory. This work identifies a major flaw in that approach, specifically that using motion capture alone to estimate the pose of the robot/system results in the generation of inconsistent visual-inertial data that is not suitable for evaluating VIO algorithms. However, we show that it is possible to generate high-quality semi-synthetic data for VIO algorithm evaluation. We do so using an open-source full-batch optimization tool to incorporate both mocap and IMU measurements when estimating the IMU’s trajectory. We demonstrate that this improved trajectory results in better consistency between the IMU data and rendered images and that the resulting data improves VIO trajectory error by 79% compared to existing methods. Furthermore, we examine the effect of visual-inertial data inconsistency (as a result of trajectory noise) on VIO performance to provide a foundation for future work targeting real-time applications.

Semi-synthetic data generation to fine-tune a convolutional neural network for retrieving Raman signals from CARS spectra

Coherent anti-Stokes Raman scattering (CARS) is a well-known third-order non-linear spectroscopic technique utilized to analyze the molecule structures. However, it has a coherent non-resonant background (NRB) alongside with Raman resonant, thus, it is difficult to extract the Raman part. In this work, the Raman signal is extracted from the CARS spectrum by using a convolutional neural network. The model architecture is adapted from the original SpecNet model. The model is pre-trained with synthetic data and fine-tuned with semi-synthetic data based on two sets of semi-synthetic spectra. The experimental results show that the model achieves 86% accuracy in predicting the Raman signal of semi-synthetic data. In addition, the sensitivity of the model performance to varying levels of noise is analysed. Based on standard metrics, the model performance decreases with the increasing level of noise in a non-linear manner. Finally, the prediction capability of the fine-tuned SpecNet model was evaluated on the four experimental CARS spectra and the results were found to be better compared to SpecNet.

The Exponomial Choice Model for Assortment Optimization: An Alternative to the MNL Model?

In this paper, we consider the yet-uncharted assortment optimization problem under the exponomial choice model, in which the objective is to determine the revenue-maximizing set of products that should be offered to customers. Our main algorithmic contribution comes in the form of a fully polynomial-time approximation scheme, showing that the optimal expected revenue can be efficiently approached within any degree of accuracy. We synthesize several ideas related to approximate dynamic programming, intended to construct a compact discretization of the continuous state space by keeping track of “key statistics” in rounded form and by operating with a suitable bit precision complexity. We complement this result by a number of NP-hardness reductions to natural extensions of this problem. Moreover, we conduct empirical and computational evaluations of the exponomial choice model and our solution method. Focusing on choice models with a simple parametric structure, we provide new empirical evidence that the exponomial choice model can achieve higher predictive accuracy than the multinomial logit (MNL) choice model on several real-world data sets. We uncover that this predictive performance correlates with certain characteristics of the choice instance—namely, the entropy and magnitude of choice probabilities. Finally, we leverage fully ranked preference data to simulate the expected revenue of optimal assortments prescribed using the fitted exponomial and MNL models. On semisynthetic data, the exponomial-based approach can lift revenues by 3%–4% on average against the corresponding MNL benchmark. This paper was accepted by Chung Piaw Teo, optimization. Funding: The research of Danny Segev on this project was supported by the Israel Science Foundation [Grants 148/10 and 1407/20]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/mnsc.2022.4492 .

Estimating the Individual Treatment Effect on Survival Time Based on Prior Knowledge and Counterfactual Prediction.

The estimation of the Individual Treatment Effect (ITE) on survival time is an important research topic in clinics-based causal inference. Various representation learning methods have been proposed to deal with its three key problems, i.e., reducing selection bias, handling censored survival data, and avoiding balancing non-confounders. However, none of them consider all three problems in a single method. In this study, by combining the Counterfactual Survival Analysis (CSA) model and Dragonnet from the literature, we first propose a CSA–Dragonnet to deal with the three problems simultaneously. Moreover, we found that conclusions from traditional Randomized Controlled Trials (RCTs) or Retrospective Cohort Studies (RCSs) can offer valuable bound information to the counterfactual learning of ITE, which has never been used by existing ITE estimation methods. Hence, we further propose a CSA–Dragonnet with Embedded Prior Knowledge (CDNEPK) by formulating a unified expression of the prior knowledge given by RCTs or RCSs, inserting counterfactual prediction nets into CSA–Dragonnet and defining loss items based on the bounds for the ITE extracted from prior knowledge. Semi-synthetic data experiments showed that CDNEPK has superior performance. Real-world experiments indicated that CDNEPK can offer meaningful treatment advice.