Search Results for « Uplift Modeling »

Uplift modeling (also known as heterogeneous treatment effect modeling) aims to predict the incremental impact of a treatment (e.g. a marketing offer or medical intervention) on an individual's outcome. The R package maq provides tools for model evaluation in uplift modeling, including Qini curves for policy analysis with multiple treatment arms (GitHub-grf-labs/maq: Qini curves for multi-armed treatment rules) (Qini curves: Automatic cost-benefit analysis • grf). However, maq currently relies on traditional models (like trees or forests) to estimate uplift. In this paper, we propose a deep learning extension that enhances all major functionalities of maq-modeling, treatment effect estimation, and evaluation-using PyTorch. We introduce new neural network architectures for uplift modeling that handle multiple treatment arms and improve individual treatment effect estimation. The models are implemented in Python/PyTorch and can interface with R for evaluation (via maq R package) to leverage Qini-based metrics. Experiments on synthetic data and real-world datasets (marketing campaign and healthcare trials) demonstrate that our deep uplift models achieve better targeting performance and more accurate treatment effect estimates than traditional methods. For example, on a synthetic uplift dataset, our approach achieves a 30% lower Precision in Estimation of Heterogeneous Effect (PEHE) and a higher Qini coefficient than both a two-model logistic regression and a causal random forest. We also illustrate Qini curves and AUUC (Area Under the Uplift Curve) to compare model performance. Results show that the deep learning extensions can more effectively identify persuadable individuals and yield policies with higher incremental gains. These findings extend the maq framework with state-of-the-art deep learning, bridging R's policy evaluation tools with powerful neural estimators of uplift.

Abstract Uplift modeling refers to individual level causal inference. Existing research on the topic ignores one prevalent and important aspect: high class imbalance. For instance in online environments uplift modeling is used to optimally target ads and discounts, but very few users ever end up clicking an ad or buying. One common approach to deal with imbalance in classification is by undersampling the dataset.In this work, we show how undersampling can be extended to uplift modeling.We propose three new undersampling methods for uplift modeling and one novel calibration method.We compare the proposed methods empirically and show when some methods have a tendency to break down.One key observation is that accounting for the imbalance is particularly important for uplift random forests, which explains the poor performance of the model in earlier works. Undersampling is also crucial for class-variable transformation based models.

Abstract Uplift modeling stands at the forefront of machine learning innovation, serving as a pivotal tool for quantifying the causal effect of marketing strategies on consumer decisions. This paper explores the intricacies of uplift modeling, examining its theoretical framework and the practical application of its principal methodologies: the one-model, two-model, and class transformation techniques. Through an empirical study of a Portuguese bank's direct phone marketing campaigns, we aim to discern the predictive capability of uplift modeling in determining client engagement with term deposit subscriptions. We present a detailed comparative analysis of the uplift modeling techniques, scrutinizing their effectiveness and limitations within the context of our dataset. The findings reveal that the class transformation approach, specifically using CatBoost, significantly outperforms its counterparts, providing a marked increase in predictive accuracy and customer conversion rates. This insight emphasizes the potential of uplift modeling to identify key customer segments for targeting, enhancing the precision and ROI of marketing initiatives. Our study contributes to the expanding literature on causal inference and targeted marketing, providing a pathway for businesses to fine-tune their marketing strategies. The application of the uplift models detailed in this paper transcends the scope of banking, offering a versatile framework for various industries to deploy data-informed decisions, thereby fostering revenue growth and optimizing marketing expenditures.

Quantifying the balance between uplift and erosion is essential for understanding mountain landscape evolution, and river longitudinal profiles provide an effective archive of this process. Clarifying the uplift history of the Malá Fatra Mountains is key to constraining the Cenozoic tectonic evolution of the Western Carpathians in Central Europe. Here, we integrate the stream power incision model with linear river-profile inversion, catchment-wide denudation rates from in-situ cosmogenic ¹⁰Be, and χ-plot analyses to reconstruct the uplift history of the Malá Fatra Mountains, comprising the SW Lúčanská Fatra and NE Krivánska Fatra. Corrected ¹⁰Be-based denudation rates range from 49.6 ± 9.0 to 163.0 ± 13.1 m Myr⁻¹, with a mean of 94.8 ± 7.2 m Myr⁻¹, indicating low to moderate long-term erosion. Using a mean erosion coefficient of K = (1.27 ± 0.13) × 10⁻⁶ m⁰·¹ yr⁻¹, modeling results show that the Lúčanská Fatra experienced a rapid uplift (~0.13 mm yr⁻¹) between ~6.3–3.5 Ma, whereas the Krivánska Fatra underwent a similar uplift (~0.14 mm yr⁻¹) between ~2.85–0.65 Ma. Knickpoint residuals yield ages consistent with these simulated phases, confirming the robustness of the approach. Our results reveal asynchronous uplift of the Malá Fatra Mountains since the Late Miocene, with an earlier phase linked to post-orogenic collapse near volcanic and attenuated lithosphere, and a later phase attributed to far-field compression from Pannonian Basin inversion. This integrated cosmogenic–geomorphic framework provides a transferable methodology for quantifying uplift histories and erosion in orogenic landscapes.