Quantum-inspired deep learning model claims 98.5% accuracy for cross-border e-commerce marketing analytics

Researchers have unveiled a hybrid quantum-classical deep learning framework that they report can predict the effectiveness of cross-border e-commerce digital marketing campaigns with 98.5% accuracy, outperforming a set of conventional baseline models in a simulator-based study published in the journal Discover Internet of Things.

The framework, called Quantum Fusion Deep Learning for Cross-Border Digital Marketing Analytics (QFDL-CBDMA), was developed by Xi Chen and combines deep learning with quantum-inspired computation. Its central component is a Quantum Fusion Neural Network that encodes multi-modal marketing data, including demographic indicators, consumer behaviour metrics, transaction history and logistics signals, into quantum states. By exploiting superposition and entanglement within parameterized quantum circuits, the model aims to capture high-order interactions among heterogeneous features that classical architectures struggle to represent without an exponential growth in parameters.

Training relies on a Quantum Gradient Descent optimizer that estimates gradients using the parameter-shift rule, requiring only two quantum circuit evaluations per parameter. The authors report that this approach converges faster and more stably than classical gradient descent in their experiments, avoiding some of the local-minimum trapping that afflicts high-dimensional optimization. The full pipeline includes two classical preprocessing layers with 256 and 128 neurons, a hybrid quantum module with 8 qubits and a circuit depth of six variational layers built from rotation gates and CNOT-based entanglement operations, and two prediction layers with 64 and 32 neurons. The quantum module contains 144 trainable parameters.

Importantly, the entire system was realized in a simulator-based hybrid environment rather than on physical quantum hardware. The experiments ran on a workstation with an Intel Core i9 processor, 64 GB of memory and an NVIDIA RTX 4090 graphics card, using Python libraries including TensorFlow, PyTorch, Qiskit and PennyLane. Training took roughly 3.8 hours, with average inference time per batch under 120 milliseconds. The authors explicitly state that no experimentally attested quantum speedup or hardware-level advantage is claimed; the study focuses instead on whether quantum-inspired representation and optimization mechanisms can improve predictive performance in a simulated setting.

The evaluation used the E-Commerce Transactions Dataset, a publicly available synthetic dataset of approximately 500,000 simulated transaction records. From this, a stratified analytical sample of 12,000 records was drawn to preserve the distributions of product categories, geographic regions and transaction values while keeping the computationally demanding quantum circuit simulations feasible. The data were split 70:15:15 into training, validation and test subsets, and all models were compared using identical partitions, preprocessing pipelines and hyperparameter optimization strategies. The target variable, marketing effectiveness, was defined as a binary label based on whether a customer completed a purchase within 30 days of campaign exposure with a transaction value above $100, yielding 3,540 high-effectiveness cases and 8,460 low-effectiveness cases.

Across 10-fold cross-validation, the framework achieved a mean accuracy of 98.5% with a standard deviation of 0.41 and a 95% confidence interval of 97.9% to 99.1%, statistically outperforming the baseline models AI-MPM, CBER-DL, DL-CBP and EL-ER according to paired t-tests reported at p < 0.05. The best classical baseline reached 93.0% accuracy over 100 training epochs. The framework also posted higher scores on return on investment, reaching 67.4% by the ninth month of a ten-month tracking period compared with 50.7% for AI-MPM, as well as on consumer engagement rate, average order value and a cross-cultural insight score that reached 96.7% and 97.4% for the China and India market segments respectively.

Two exploratory diagnostic metrics were introduced alongside the standard measures. The Quantum Feature Entanglement Score, which quantifies the strength of entangled representations between feature subsets from different modalities, ranged from 0.85 to 0.97 for the proposed model, above the 0.60 to 0.85 range observed for the baselines. The Quantum Complexity Gain Score, which relates performance to computational cost, rose from 0.92 to 1.00 over the ten-month evaluation. The authors caution that the latter metric is neither upper-bounded nor distribution-free and that its results are reported as exploratory indicators requiring further validation.

An ablation study supported the contribution of each component. Removing quantum encoding reduced the customer engagement ratio from 86.9% to 80.2%, a decline of 6.7 percentage points, while removing entanglement-based fusion lowered the cross-cultural insight score by 4.9 percentage points. The degradation patterns were more pronounced for engagement and cultural adaptation than for raw accuracy, suggesting that the quantum-inspired components contribute most to tasks that depend on sparse, high-dimensional interactions among cross-border data streams. The framework also remained robust under simulated conditions: accuracy stayed above 98.2% for network delays up to 200 milliseconds and 97.1% at 500 milliseconds, dropped only to 96.82% when 15% noise was injected into the simulated quantum circuits, and varied by less than 1.5% when data volume was scaled from 10,000 to 100,000 records.

The study acknowledges substantial limitations. The dataset is synthetic and contains no real user data, and it lacks direct variables for local currency, language preference or country-specific regulatory conditions, which the authors identify as priorities for future validation. Scaling to the full 500,000-record dataset would require roughly four million circuit evaluations per epoch under the parameter-shift rule. On near-term NISQ quantum devices, the same circuit would need eight physical qubits with gate fidelity exceeding 99.9% to preserve the observed accuracy; at typical current error rates, the effective circuit fidelity would fall below 95%, corresponding to an estimated accuracy degradation of 2.5 to 4.0 percentage points. The authors propose zero-noise extrapolation and randomized compilation as compatibility strategies for future hardware deployment.

The conclusion is carefully scoped: within the simulator-based environment and the dataset employed, the framework consistently outperformed the implemented baselines, but the observed gains are contingent on the specific dataset characteristics, baseline configurations and simulation conditions, and do not establish general superiority of quantum-classical hybrid architectures across all e-commerce contexts. The authors call for further testing with different datasets, real-world IoT deployments and physical quantum platforms before the approach can be considered validated beyond the simulated experimental setting.

Subject of Research: A hybrid quantum-classical deep learning framework for evaluating cross-border e-commerce digital marketing effectiveness

Article Title: A study on evaluating the effectiveness of cross-border e-commerce digital marketing based on deep learning

Article References: Chen, X. (2026). A study on evaluating the effectiveness of cross-border e-commerce digital marketing based on deep learning. Discover Internet of Things, 6(1), Article 225. https://doi.org/10.1007/s43926-026-00529-x

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00529-x

Keywords: quantum machine learning, deep learning, cross-border e-commerce, digital marketing analytics, quantum neural networks, parameter-shift rule, marketing effectiveness prediction, simulator-based study, synthetic dataset, consumer engagement, return on investment, hybrid quantum-classical computing