Cross-Attention Model Improves Post-Delivery E-Commerce Satisfaction Predictions in Brazilian Study

A computational framework combining multimodal cross-attention networks with heterogeneous ensemble modeling has achieved improved accuracy in predicting customer satisfaction from completed e-commerce orders, according to research conducted on a large-scale Brazilian marketplace dataset. The system, termed the Multi-modal Cross-attention Network, or MM-CrossNet, addresses a persistent challenge in digital commerce operations by integrating four distinct order-level data streams: statistical text review metrics, logistics tracking time series, payment and pricing figures, and geographical demographic profiles. Evaluated on the public Olist dataset comprising 96,358 order records across a two-year operational window, the architecture reached an Accuracy of 0.8425 and an F1-score of 0.9059. These scores outperformed nine established machine learning baselines, including gradient boosted decision trees and deep recommendation architectures. While the numerical margin over the strongest alternative, an optimized XGBoost model, remained modest at 0.0004 in Accuracy and 0.0002 in F1-score, rigorous paired testing confirmed that the performance difference was statistically significant. The findings establish a systematic engineering benchmark for post-fulfillment retrospective analytics, demonstrating that dynamic attention across heterogeneous data modalities captures predictive signals that conventional single-modality pipelines and feature concatenation techniques systematically miss.

The operational context surrounding online platform management highlights why accurate customer satisfaction modeling has become central to retail enterprise strategy. Rapid global expansion across e-commerce channels has heightened customer acquisition costs, elevating retention, repeat purchasing, and brand sentiment to primary determinants of platform competitiveness. Dissatisfied consumers exhibit pronounced behavioral asymmetries in digital feedback systems, demonstrating a documented tendency to submit written negative reviews far more frequently than contented buyers. In the Olist records examined by the researchers, orders accompanied by written commentary recorded an observed satisfaction rate of only 67.43 percent, compared to 86.77 percent among orders without written remarks, generating a substantial 19.34 percentage point disparity. Most existing commercial prediction systems rely predominantly on tabular transaction records or apply basic sentiment classification to customer feedback in isolation. By neglecting the nonlinear structural coupling among delivery timelines, freight charges, regional logistics disparities, and qualitative feedback, traditional systems struggle to balance class disparities or identify the operational inflection points that drive customer disaffection.

To build an operationally realistic prediction framework, the researchers strictly defined their modeling objective around a post-delivery retrospective analysis scenario rather than real-time pre-delivery tracking. In commercial practice, customer satisfaction is shaped across three distinct stages: pre-transaction product discovery, fulfillment and delivery, and post-sales customer service. The investigation concentrated specifically on fulfillment and delivery, as this phase contains complete order-level tracking timestamps, validated financial payments, and final customer feedback. The dataset features were systematically annotated according to their earliest temporal availability across three phases: order placement, shipment confirmation, and post-delivery completion. Out of 23 engineered features, 16 represent variables known prior to delivery finalization, while seven represent post-event observables, including final transit durations and review presence flags. Restricting target encoding computations exclusively to training folds prevented data leakage from future outcomes, ensuring that the retrospective scores could serve as reliable indicators for auditing service failures and guiding targeted customer outreach.

Feature engineering transformed raw relational tables into a structured 23-dimensional feature space covering five functional categories: logistics performance, pricing metrics, product attributes, regional geography, and textual sentiment proxies. Logistics attributes included normalized metrics such as absolute delivery deviation, calculated by subtracting actual transit duration from estimated arrival dates, alongside relative delivery ratios that normalize delays by total transit span. Financial indicators incorporated logarithmic price transformations and freight ratios representing shipping expenses relative to item value. Categorical variables with high cardinality, specifically merchant product categories and customer states, received out-of-fold target encoding to capture global relationships with satisfaction labels without causing label contamination. This rigorous schema ensured that the statistical inputs provided to the subsequent machine learning layers preserved clear physical interpretations, reflecting measurable logistics delays, price sensitivities, and regional infrastructure variations across 27 administrative territories throughout Brazil.

Rather than feeding raw unstructured data directly into a massive monolithic neural network, MM-CrossNet employs four dedicated modality encoders that project heterogeneous inputs into a unified 64-dimensional vector space. The textual semantic module utilizes four lightweight statistical linguistic markers: review character length, polarity density derived from the Portuguese SentiLex-PT lexicon, negation marker frequency, and exclamation point ratios. These statistical features allow the model to extract sentiment cues while gracefully handling the fact that only 40.58 percent of orders contained written customer reviews. The logistics encoder processes multidimensional temporal measurements through a two-layer multilayer perceptron with layer normalization, compensating for substantial variance across approval hours and dispatch spans. Pricing and payment variables pass through a single-layer perceptron tailored to homogeneous numerical inputs, while high-cardinality categorical variables are processed through a dual-path architecture combining 32-dimensional dense embeddings with one-dimensional target encoding, creating four structured modality tokens.

The central methodological contribution of the framework lies in its cross-attention fusion layer, which arranges the four modality embeddings as serialized tokens in a four-by-sixty-four matrix. Unlike conventional feature concatenation, which forces downstream classifiers to deduce inter-modal relationships without explicit architectural guidance, or average pooling, which obliterates modal identity, cross-attention enables dynamic, input-conditioned interaction. Utilizing four parallel attention heads, the network allows representations from one modality to query and attend to relevant features in other modalities, effectively permitting severe logistics delay signals to modulate how textual sentiment proxies are interpreted. The attended tokens pass through residual connections, layer normalization, and a two-layer feed-forward network before undergoing cross-token average pooling to produce a unified 64-dimensional latent representation vector. This vector encapsulates deep inter-modal dependencies and preserves essential structural contrasts before passing to downstream decision layers.

For final prediction generation, the architecture links the cross-attention representation to a heterogeneous ensemble combining gradient-boosted decision trees and a deep neural network. The ensemble utilizes LightGBM configured with 63 leaves and a maximum depth of eight, XGBoost configured with a maximum depth of seven, and a three-layer perceptron with hidden layer dimensions of 256, 128, and 64 units. The base models train independently on the fixed 64-dimensional latent representations extracted from the trained cross-attention encoder. A validation grid search identified optimal ensemble weights of 0.40 for LightGBM, 0.35 for XGBoost, and 0.25 for the multilayer perceptron. Model optimization utilized a weighted binary cross-entropy loss function that explicitly accounted for class imbalance, where satisfied ratings outnumber dissatisfied ratings by approximately four to one. This hybrid design exploits the complementary strengths of tree-based partitioning across structured feature splits and neural network flexibility across continuous latent spaces.

Extensive empirical evaluations conducted across five independent experimental runs verified the stability and predictive competitiveness of the unified model. The 96,358 cleaned records were split into 77,086 training samples and 19,272 testing samples using stratified sampling to preserve the underlying four-to-one satisfaction ratio. Across five iterations, MM-CrossNet achieved an average Accuracy of 0.8425 with a standard deviation of 0.0004, an F1-score of 0.9059 with a standard deviation of 0.0003, and an area under the receiver operating characteristic curve of 0.7829. While classical gradient boosted decision trees registered the highest standalone area under the curve at 0.7857, MM-CrossNet secured top ranks in overall Accuracy and F1 classification balance. Formal hypothesis testing using McNemar’s paired test on the test set confirmed that the paired prediction differences between MM-CrossNet and each baseline, including XGBoost and LightGBM, achieved statistical significance at a significance threshold of p less than 0.05.

Single-modality comparisons and component ablation experiments systematically isolated the individual predictive contributions of specific data streams and architectural components. When trained on individual modalities alone, model performance deteriorated markedly across all configurations; logistics records proved to be the most informative individual stream with an area under the curve of 0.7124, whereas geographic variables alone yielded the lowest predictive utility at 0.6318. The 0.0705 difference between the logistics-only variant and the complete multimodal architecture confirms that cross-modal fusion captures vital complementary information. In ablation tests, removing the textual module caused the sharpest decrease in performance, reducing the area under the curve by 0.0679, followed closely by the removal of the logistics module, which induced a drop of 0.0511. Furthermore, comparing fusion strategies on identical encoders demonstrated that cross-attention surpassed standard feature concatenation, average pooling, and fully connected multilayer fusion by margins up to 0.0186 in area under the curve, confirming the specific value of dynamic attention weights.

Interpretability analysis utilizing LightGBM Split Gain revealed that logistics timeliness and product disclosure depth served as the most prominent predictive indicators within the fused representation. Order approval duration emerged as the single most influential variable, followed by product description length, customer freight ratios, and target-encoded product categories. Segmenting the empirical records revealed an inverse relationship between fulfillment durations and customer satisfaction: orders delivered within three days recorded an average satisfaction rate of 87.2 percent, whereas orders requiring more than 30 days dropped precipitously to 24.2 percent. Sensitivity across merchandise categories showed marked variation, with telephony and bedding supplies experiencing sharp satisfaction collapses during shipping delays, whereas sports equipment maintained relatively stable ratings. Computational profiling established that MM-CrossNet required 52.3 seconds per training epoch and exhibited an inference latency of 74.2 microseconds per sample on standard server hardware, confirming its practical viability for offline batch processing.

To connect predictive scores with commercial workflows, the study outlined an illustrative three-tier operational customer relationship management framework based on predicted satisfaction probabilities. Under this non-calibrated operational mapping, orders scoring below 0.40 trigger immediate automated outreach, priority compensation, and merchant escalation to mitigate severe service failures. Orders scoring between 0.40 and 0.65 receive targeted retention measures, such as promotional discount vouchers, loyalty incentives, and merchant description audits. Orders scoring 0.65 or higher transition toward customer advocacy activation, including automated invitations for product reviews, referral bonuses, and cross-selling campaigns. Translating model probabilities into concrete customer relationship tiers demonstrates how multimodal algorithmic scoring can bridge the gap between abstract predictive metrics and actionable retail operational policies, directly supporting organizational customer retention initiatives.

The authors identified several methodological boundaries that constrain immediate cross-market generalization. First, because the empirical analysis relied exclusively on the Olist Brazilian marketplace, observed logistics sensitivities may reflect infrastructure constraints unique to South American transport networks rather than universal e-commerce dynamics, necessitating future validation on cross-platform repositories such as Amazon Reviews, JD.com, and Shopee. Second, relying on observational records precludes causal claims; observed satisfaction drops represent statistical associations that may incorporate unmeasured confounders, such as seller reputation or inventory quality. Third, the textual module utilized statistical lexical proxies rather than modern multilingual pretrained language models due to computational scope and historical data constraints. Finally, because the Olist dataset consists predominantly of single-order transactions with an approximate repeat purchase rate of only 0.3 percent, direct longitudinal validation against multi-order churn labels remains an objective for future investigation.

Subject of Research: Multimodal machine learning for e-commerce customer satisfaction prediction

Article Title: An ensemble multi-modal cross-attention model for e-commerce customer satisfaction prediction

Article References: Xu, Y. (2026). An ensemble multi-modal cross-attention model for e-commerce customer satisfaction prediction. Discover Artificial Intelligence, 6(1), Article 1697. https://doi.org/10.1007/s44163-026-02317-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02317-6

Keywords: cross-attention mechanism, multimodal machine learning, customer satisfaction prediction, e-commerce analytics, Olist dataset, heterogeneous ensemble, LightGBM, XGBoost, logistics performance, text sentiment analysis, SentiLex-PT, retrospective order analysis, McNemar test, feature importance, class imbalance modeling, digital marketing capability