1 Introduction

With the rapid development of computers, more and more people shop online, and the e-commerce industry is growing rapidly [1]. However, as computer technology advances, the amount of information has increased dramatically, and users often become confused by excessive information when shopping, making it difficult to choose truly suitable products [2]. In this context, accurately pushing information to users has become a major research problem. User profiling is an analytical method that deeply examines basic user attributes and purchase behavior, helping the model formulate more precise personalized push strategies [3]. Stacking is a widely used ensemble learning technique that enhances the accuracy of information recommendation by combining multiple distinct base models [4]. Extreme Gradient Boosting (XGBoost) is a tree-based machine learning algorithm known for its flexibility, high accuracy, and computational efficiency, and is commonly applied in data mining and economic forecasting [5]. Logistic Regression (LR) is a common classification algorithm that performs analysis, prediction, and classification, with the advantages of a simple structure and computational efficiency [6]. Random Forest (RF) is a robust machine learning algorithm that improves prediction performance by integrating the results of multiple decision trees [7]. Multiple Linear Regression (MLR) is a widely used statistical model that analyzes multivariable problems and is applied in finance, biostatistics, and other fields [8]. Therefore, this study uses RF, LR, and XGBoost as base learners in the Stacking ensemble learning framework, with MLR as the meta-learner, and combines user feature information to construct an e-commerce information push model. The innovation of the study lies in improving the Stacking framework by adding weights and analyzing user information using the K-means clustering algorithm (K-means).

2 Related Work

With the rapid growth in the number of Internet users, accurately delivering information to users has become a major research problem, and many scholars at home and abroad have conducted in-depth research on this. Chen and Yang addressed the low push efficiency in online teaching systems and proposed a resource information push model based on Latent Dirichlet Allocation, combining it with temporal interest information to design an interest topic module and to calculate topic similarity and student interest weights to achieve personalized information push [9]. To address information overload in personalized agricultural knowledge services, Wu et al. proposed an agricultural knowledge service push framework based on Text Convolutional Neural Network and Long Short-Term Memory Neural Network, using Generative Adversarial Networks to depict user profiles and applying data obfuscation techniques to protect user privacy [10]. Huang et al. focused on the problem of manufacturing cloud services being unable to correctly perceive user preferences in information push and proposed a model based on user preference features and deep learning algorithms, extracting user preference features via Latent Dirichlet Allocation and Convolutional Neural Network models and performing clustering using a label clustering method [11].

Stacking demonstrates excellent predictive performance and shows great potential in various fields such as traffic forecasting and fault diagnosis, and research on Stacking continues to deepen [12]. Muhammad et al. addressed the weak generalization ability in facial presentation attack detection and proposed a Stacking-based ensemble learning model that generates different synthetic data subsets using the Alpha fusion algorithm and trains the meta-model on the outputs of base models over these synthetic data subsets to improve prediction accuracy [13]. To address missing data, Li et al. proposed a stacking ensemble learning model based on a data-generating neural network, integrating data via a neural generative adversarial network, optimizing hyperparameters with a tree-structured Parzen estimator, and predicting results via ensemble learning [14]. Thabet et al. studied the issue of excessive data in financial system performance evaluation and proposed a banking efficiency evaluation model based on Stacking ensemble learning, using K-Nearest Neighbor, Random Forest, and other machine learning models as base models, and applying cross-validation to train the base models to prevent overfitting [15].

The base learners in stacking ensemble learning often employ machine learning algorithms. For a long time, machine learning algorithms have been widely applied in various fields due to their powerful data processing capabilities. To achieve more accurate river classification, Ahmed et al. proposed a river classification model based on multiple machine learning algorithms. They balanced the data using the minority oversampling technique, classified the rivers using linear kernel support vector machines, RF, and multi-layer perceptrons, and compared the results [16]. Ding et al. addressed the problem of ground vibration prediction during mine explosions and proposed a model based on XGBoost and the imperialist competitive algorithm. They combined XGBoost with the particle swarm optimization algorithm, an artificial neural network, and support vector regression to achieve more efficient prediction [17]. To improve the mechanical properties of recycled aggregate concrete, Duan et al. proposed a model based on meta-heuristic search algorithms. They combined the imperialist competitive algorithm with XGBoost to estimate the 28-day compressive strength of recycled aggregate concrete [18]. Kumar and Swathi addressed the problem of dynamic rainfall prediction and proposed a rainfall prediction model based on adaptive boosting trees. They extracted weather features such as temperature, dew point, and precipitable water vapor, and conducted rainfall data analysis using R tools and adaptive boosting trees to achieve more accurate rainfall prediction [19].

In conclusion, the current e-commerce information push methods have, to some extent, alleviated information overload, but they still lack diversity and model accuracy. Therefore, it is urgent to develop a new method to optimize the accuracy and generalization of information push technology. The current research on the Stacking ensemble learning model and machine learning algorithms indicates that the Stacking framework can combine multiple algorithms to improve detection accuracy. Commonly used machine learning algorithms, such as XGBoost and RF, have strong data analysis capabilities and are suitable for classification and predictive analysis, with good generalization. Therefore, an e-commerce information push model based on Stacking ensemble learning and machine learning algorithms was proposed. By combining RF, LR, and XGBoost as base learners and using MLR as the meta-learner, an improved Stacking ensemble framework was constructed. This framework combines multiple machine learning models for prediction, achieving higher predictive performance than a single algorithm. It is expected that this model can perform e-commerce information push more accurately and efficiently, promoting the development of the e-commerce industry.

3 Research Design

3.1 User Profiling Construction Based on User Features and K-Means Clustering

To achieve a precise e-commerce information push, the primary task is to analyze user feature information. User profiling is a labeled user information analysis model that extracts, analyzes, and integrates user information from the database to deeply understand user needs and achieve precise push [16, 17]. In the e-commerce field, constructing user profiles allows analysis of users’ basic attributes, purchase behavior, and other information, which helps formulate more precise personalized push strategies [18]. User profiling primarily involves data processing, analysis, and modeling. By mining and analyzing data, user information is labeled to promote precise e-commerce information push. The application mechanism of user profiling in e-commerce information push is shown in Fig. 1.

Fig. 1
Fig. 1
Full size image

Mechanism of user profiling in e-commerce information push

As shown in Fig. 1, the application of user profiling in e-commerce information push is primarily achieved through data processing and analysis. First, user purchase history, basic attributes, browsing records, and other information are collected, followed by data cleaning and reduction to meet information needs, removing unnecessary data. Then, labeling is performed. A label system is designed based on basic attribute data, user information, and interest preference information. Finally, data analysis is conducted. Clustering analysis classifies categories, constructs user profiles, and outputs user behavior features to perform predictive analysis for e-commerce information push. In the user profiling model, user behavior prediction is achieved by calculating user similarity and applying score weighting [20]. The formula for score weighting is presented in Eq. (1).

$$R_{u,p} = \frac{{\mathop \sum \nolimits_{s \in S} (w_{u,s} \cdot R_{s,p} )}}{{\mathop \sum \nolimits_{s \in S} w_{u,s} }}$$
(1)

In Eq. (1), wu,s represents the correlation between two users u and s. Clustering analysis is the key step in constructing user profiles. By calculating similarity, user feature information is clustered and divided, which helps the model achieve precise targeting [21]. The K-Means algorithm is a common unsupervised learning algorithm that iteratively optimizes centroid positions to cluster data, thereby increasing inter-class differences. It has the advantages of simple and efficient computation and strong flexibility [22]. Therefore, this study uses K-means clustering for user profiling. The data analysis process of user profiling based on K-Means clustering is shown in Fig. 2.

Fig. 2
Fig. 2
Full size image

Data analysis process of user profiling based on K-Means clustering

As shown in Fig. 2, the K-Means clustering process in the user profiling model first determines the number of clusters, then performs random initialization on the processed user data. The same number of data objects as clusters is randomly selected as initial cluster centers. Then, the distance between each user data sample and the centroid is calculated, and the sample is assigned to the nearest cluster. The distance between samples and centroids is updated based on the mean of the samples in each cluster. Iterative optimization continues until convergence, and clustering results are output. In the K-Means algorithm, calculating the distance from samples to centroids is the most critical step, as shown in Eq. (2).

$$[d(x,\mu )]^{2} = \mathop \sum \limits_{i = 1}^{n} (x_{i} - \mu_{i} )^{2}$$
(2)

In Eq. (2), \(d(x,\mu )\) represents the distance from sample x in a cluster to centroid µ. The sum of squared distances is calculated as shown in Eq. (3).

$$CSS = \mathop \sum \limits_{j = 0}^{m} \mathop \sum \limits_{i = 0}^{n} (x_{i} - \mu_{i} )^{2}$$
(3)

In Eq. (3), j is the sample index, and m represents the number of samples in a cluster. The clustering effect within a cluster is evaluated based on the sum of squared distances. The smaller the sum of squared distances, the better the clustering effect [23].

3.2 Design of Stacking Base Learners Based on Multiple Machine Learning Algorithms

Based on user profiling, using recommendation algorithms to push e-commerce information allows formulation of more precise push strategies. Ensemble learning models are a highly practical machine learning technique that enhances overall model performance by combining the predictions of multiple models. Common ensemble learning methods fall into three main types: Bagging, Boosting, and Stacking. Among them, Bagging is a parallel ensemble learning method with strong robustness. However, this method aggregates results via voting or averaging, which cannot dynamically adjust weights to adapt to different data distributions, and it also lacks generalization ability. Boosting is a serial ensemble learning method that predicts results by sequentially training multiple weak learners. Its computational cost increases significantly with the number of models. The Stacking framework is a parallel ensemble learning method that effectively integrates heterogeneous models’ predictions via meta-learning. By leveraging the strengths of different models, it enhances both accuracy and generalization capabilities. Compared to Bagging and Boosting models, it demonstrates superior performance [24]. Therefore, this study adopts the Stacking framework as the recommendation algorithm for e-commerce information push prediction. Based on user feature information, predictive methods that accurately forecast information push are selected as the base learners for the Stacking framework to provide preliminary predictions of product information. Given that user behavior data typically contains diverse features such as browsing history, purchase records, and basic user information, the data processing requirements are quite demanding. To address this, the study adopted three algorithms as base learners: RF for parallel processing of large-scale feature data, XGBoost to prevent overfitting, and the LR algorithm for its high computational efficiency. These algorithms were employed to process user information and perform initial information recommendation, thereby achieving more accurate predictions for e-commerce information push.

Among them, the XGBoost algorithm is an efficient gradient-boosting tree-based predictor. It continuously optimizes decision tree parameters to achieve more precise predictions. Therefore, XGBoost is used as a base learner in the Stacking framework for preliminary e-commerce information push prediction. The XGBoost-based prediction mechanism is shown in Fig. 3.

Fig. 3
Fig. 3
Full size image

Prediction mechanism of e-commerce information push based on XGBoost (Icon source from: https://iconpark.oceanengine.com/official)

As shown in Fig. 3, XGBoost predicts by constructing gradient-boosted decision trees. First, user profiling features and user behavior features are merged, followed by secondary data processing. Processed user data is initialized and iteratively optimized. In each iteration, the objective function and regularization term are calculated, residuals are computed, a new decision tree is constructed, leaf node weights are calculated, and new predictions are generated. Regularization is used to prevent overfitting, and the predictions are fed into the meta-learner. In XGBoost, the model prediction is computed using an additive model, as shown in Eq. (4).

$$y_{i}^{(t)^{\prime}} = \mathop \sum \limits_{k = 1}^{t} f_{k} (x_{i} ) = y_{i}^{(t + 1)^{\prime}} + f_{t} (x_{i} )$$
(4)

In Eq. (4), \({\text{f}}_{\text{k}}\text{(}{\text{x}}_{\text{i}}\text{)}\) represents the prediction value of the decision tree in the t-th iteration. To prevent overfitting caused by too many leaf nodes, XGBoost optimizes model complexity using the objective function [25]. The calculation equation of the objective function is shown in Eq. (5).

$$Obj = \mathop \sum \limits_{i = 1}^{n} l(y_{i} ,y_{i} ^{\prime}) + \mathop \sum \limits_{i = 1}^{K} \Omega (f_{i} )$$
(5)

In Eq. (5), \(\mathop \sum \limits_{i = 1}^{K} \Omega (f_{i} )\) is the regularization term, \(f_{i}\) represents the tree model to be trained, and \(\mathop \sum \limits_{i = 1}^{n} l(y_{i} ,y_{i} ^{\prime})\) is the loss function. XGBoost processes e-commerce user information to generate ranked information push results, but it still has limitations in generalization and comprehensive prediction. LR maps results to probability values through linear combinations of features. Classification is performed according to probability values. LR is commonly used for binary classification tasks and has high classification performance [26]. Therefore, LR is also adopted as a base learner in the Stacking framework for preliminary e-commerce information push prediction. The mechanism of e-commerce information push based on LR is shown in Fig. 4.

Fig. 4
Fig. 4
Full size image

Prediction mechanism of e-commerce information push based on LR (Icon source from: https://iconpark.oceanengine.com/official)

As shown in Fig. 4, LR primarily uses the Sigmoid function. First, user behavior features and label data are standardized and concatenated to perform binary classification. Then, during model training, it calculates weights and biases, adds a regularization term, and performs linear regression. The output of linear regression is mapped to the range 0–1 using the Sigmoid function to generate probability predictions. The predictions are then input into the meta-learner for further prediction. Mapping predicted values using the Sigmoid function is the core step of LR, shown in Eq. (6).

$$P(y = 1\left| x \right.) = \sigma (\omega^{T} x + b) = \frac{1}{{1 + e - (\omega^{T} x + b)}}$$
(6)

In Eq. (6), \(P(y = 1\left| x \right.)\) represents the posterior probability of class 1, \(\sigma\) is the Sigmoid function, \(\omega\) is the weight, and b is the bias. Based on the probability values, the loss function is calculated to determine the probability of the correct class, as shown in Eq. (7).

$$L(\omega ) = - \frac{1}{N}\mathop \sum \limits_{i = 1}^{N} [y_{i} \log (p_{i} ) + (1 - y_{i} )\log (1 - p_{i} )]$$
(7)

In Eq. (7), \(y_{i}\) and \(p_{i}\) represent the true value and predicted value of sample i, respectively. The study uses LR and XGBoost as base learners for preliminary push prediction, improving accuracy and flexibility, but both methods have limitations in processing complex data. RF can perform classification and regression tasks by parallelly training multiple decision trees, handling complex data with many features, and achieving high accuracy and stability. Therefore, RF is adopted as the third base learner in the Stacking framework. Its outputs are input into the meta-learner for e-commerce information push prediction. The RF prediction mechanism is shown in Fig. 5.

Fig. 5
Fig. 5
Full size image

Prediction mechanism of e-commerce information push based on RF (Icon source from: https://iconpark.oceanengine.com/official)

As shown in Fig. 5, RF predicts by constructing multiple decision trees. First, user information data is processed, followed by random sampling to select samples and features. Then, attributes for node splitting are chosen based on the Gini index, gradually dividing elements into branches to construct the decision tree. This process is repeated to build multiple decision trees. The ensemble of predictions from all trees is used as the final prediction and input into the meta-learner. The Gini index equation is shown in Eq. (8).

$$Gini(D) = 1 - \mathop \sum \limits_{k = 1}^{K} \left( {\frac{{C_{k} }}{D}} \right)^{2}$$
(8)

In Eq. (8), \({\text{D}}\) represents the set of user samples, \({\text{C}}_{\text{k}}\) represents the subset of class \({\text{k}}\) in set \({\text{D}}\), and \({\text{K}}\) is the total number of classes. RF computes the final prediction by integrating results from multiple decision trees. The equation for the final prediction is shown in Eq. (9).

$$y^{\prime} = \arg \max i\sum j = i = I_{m} I(f_{j} (x) = i)$$
(9)

In Eq. (9), \({\text{I}}\) is the indicator function, and \({\text{f}}_{\text{j}}\text{(}{\text{x}}\text{)}\) is the prediction value. The study employs the Stacking framework to combine predictions from random forests, logistic regression, and XGBoost models. A novel meta-learning model is trained on these predictions to achieve more accurate forecasting.

3.3 Information Push Model Construction

To prevent overfitting caused by complex architectures, a simple meta-learning model is urgently needed to train base learning models. The Multi-Layer Regression (MLR) algorithm, a common machine learning method, excels at analyzing multivariate problems due to its structural simplicity and computational efficiency. Accordingly, the study adopts MLR as the meta-learning model to train prediction results from base learning models, including random forest, logistic regression, and XGBoost. This establishes an e-commerce information push model based on user feature characteristics and an improved Stacking ensemble learning framework. The final e-commerce information push prediction mechanism (RLXM), which integrates user feature characteristics and the enhanced Stacking ensemble learning framework, is illustrated in Fig. 6.

Fig. 6
Fig. 6
Full size image

Prediction mechanism of e-commerce information push in the UI-Stacking model (Icon source from: https://iconpark.oceanengine.com/official)

As shown in Fig. 6, the e-commerce information push prediction mechanism based on user information characteristics and the improved Stacking RLXM model mainly applies through three basic models and one meta-learning model. Firstly, user information data is cleaned and augmented, and the user information features are split into a training set and a test set. The training set is fed into the basic model, and the model outputs predictions for 9 feature dimensions. Then, the prediction results are trained using five-fold cross-validation, and the resulting model is input into the MLR layer for further training. Subsequently, the test set data is input into the base learner for prediction again, the prediction results are output, weights are added, and the added-weighted prediction results are input into the trained MLR model to complete the final prediction. Finally, the recommendation list is generated based on the prediction results, and the TOP-N products are selected using a scoring function for e-commerce information push. To improve the model’s stability, the study uses five-fold cross-validation to train the base learner model, thereby reducing the impact of different data partitions on the prediction results. Five-fold cross-validation divides the e-commerce user information feature dataset into five parts, including four training sets and one test set. Based on the test set data, the prediction matrix is generated using Eq. (10).

$$P_{i} = [P_{i1} ,P_{i2} ,P_{i3} ,P_{i4} ,P_{i5} ]^{T} (i = 1,2,3)$$
(10)

In Eq. (10), i represents different machine learning algorithms used in the study, T is the transpose symbol, and \(P_{i1}\) represents the first predicted label. The output values of base learners typically include predicted labels, prediction probabilities, and confidence scores. Existing meta-feature learning methods rely solely on the final predicted labels from base learners for ensemble learning, overlooking the rich intermediate information generated during prediction. Moreover, the feature dimensionality of Stacking ensemble learning models is determined by the number of base learners. The study uses Random Forest (RF), Linear Regression (LR), and XGBoost as base models, resulting in a relatively small overall dimensionality that makes it difficult to better distinguish between feature quality. Furthermore, the Multi-Linear Regression (MLR) method transforms nonlinear outputs of basic learners into manageable linear feature combinations through confidence-weighted summation and weight allocation mechanisms. With its inherent advantages of simplicity, interpretability, and computational efficiency, it effectively maintains model stability and generalization capabilities. To address this, the study incorporates the prediction probabilities and confidence scores from the three base learners into the Multi-Layer Regression (MLR) algorithm, tripling the dimensionality to enhance the model’s predictive performance. The prediction confidence level quantitatively assesses a model’s reliability in generating specific predictions, serving as a probabilistic metric that indicates how closely the prediction aligns with the actual value. Its primary function is to quantify the model’s prediction uncertainty, thereby providing a basis for dynamic weight allocation in ensemble learning frameworks. This confidence level is primarily determined by the base learner’s prediction error rate, as calculated by Eq. (11).

$$\varepsilon_{i} = \frac{1}{N}\mathop \sum \limits_{j = 1}^{N} [(P_{ij} \ne y_{j} )],\;i = 1,2,3$$
(11)

In Eq. (11), \(P_{ij}\) is the prediction result of the j-th base learner, and \(y_{j}\) is the true label of the j-th sample. Based on prediction error, prediction confidence is generated, calculated as shown in Eq. (12).

$$a_{i} = \frac{1}{{\varepsilon_{i} }} - 1$$
(12)

To help MLR better distinguish input features, feature weights are computed based on confidence, as shown in Eq. (13).

$$\omega_{i} = \frac{{a_{i} }}{{\mathop \sum \nolimits_{k = 1}^{3} a_{k} }},i = 1,2,3$$
(13)

In Eq. (13), \(\omega_{i}\) is the weight of the base learner \(C_{i}\). By improving and adding feature dimensions and weights, the Stacking framework is enhanced, increasing final prediction accuracy.

4 Results and analysis

4.1 Dataset Information and Preprocessing

The study selected the user behavior from the Taobao on the Aliyun Tianchi platform as the dataset (source: https://www.heywhale.com/mw/dataset/603e359c63d3c30015bda7f/file). This data set recorded the behavior information of 106,042 users in Taobao platform in 9 days, and included a total of about 3,150,000 interaction records. Each record contains multiple fields, including the user ID, product ID, product category, user behavior type, and behavior timestamp. The behavior type mainly includes clicks, purchases, add-to-cart actions, and favorites. All user data in the dataset has been anonymized to protect privacy. The user age range is 18 to 65 years old, with an average of 25.75 product interactions per user. The dataset was cleaned by deduplication and denoising to remove outliers and fill in missing data. Then, data augmentation was performed to generate synthetic interaction data, with random behavior insertion and deletion to improve the model’s generalization ability. The preprocessed dataset was split into a training and validation set using stratified sampling with a 8:2 ratio.

4.2 Model evaluation methods

4.2.1 Mean Squared Error

Mean Squared Error (MSE) is a key metric for evaluating model accuracy, measuring the average squared difference between predicted and actual values. The definition of the MSE is given by Eq. (14).

$$MSE = \frac{1}{n}\sum\limits_{i = 1}^{n} {(y_{i} - \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y}_{i} )^{2} }$$
(14)

In Eq. (14), \(y_{i}\) denotes the true value, \(\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y}_{i}\) represents the predicted value, and n indicates the sample size.

4.2.2 Root Mean Squared Error

Root Mean Square Error (RMSE), derived from the square root of the MSE, shares the same units as the original data, providing a more intuitive representation of the model’s error magnitude. The definition of RMSE is given by Eq. (15).

$$RMSE = \sqrt {\frac{{\sum\limits_{i = 1}^{n} {(y_{i} - \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y}_{i} )^{2} } }}{n}}$$
(15)

4.2.3 Mean Absolute Error

Mean Absolute Error (MAE) is the average absolute value of the difference between predicted and actual values, reflecting the absolute magnitude of the error. The definition formula of MAE is shown in Eq. (16).

$$MAE = \frac{1}{n}\sum\limits_{i = 1}^{n} {\left| {y_{i} - \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y}_{i} } \right|}$$
(16)

4.2.4 Receiver Operating Characteristic

The Receiver Operating Characteristic (ROC) curve is a key metric for evaluating classification model performance, which plots the classifier’s performance across different thresholds using True Positive Rate (TPR) and False Positive Rate (FPR) as the x-axis and y-axis, respectively. The Area Under the Curve (AUC) value represents the area under the ROC curve, which measures model performance. A higher AUC value (closer to 1) indicates better model performance, while a lower AUC value (closer to 0) suggests poorer performance. Here, TPR and FPR respectively indicate the model’s ability to correctly identify positive samples, while FPR represents the proportion of negative samples incorrectly predicted as positive. The specific formula is defined in Eq. (17).

$$\left\{ \begin{gathered} TPR = \frac{TP}{{TP + FN}} \hfill \\ FPR = \frac{FP}{{FP + TN}} \hfill \\ \end{gathered} \right.$$
(17)

In Eq. (17), \(TP\) is the number of correctly identified positive cases, \(FN\) indicates the number of negative cases incorrectly identified as positive, \(FP\) is the number of samples misclassified as positive, and \(TN\) shows the number of correctly identified negative cases.

4.2.5 Normalized Discounted Cumulative Gain

The Normalized Discounted Cumulative Gain (NDCG) is a metric for evaluating ranking quality. By integrating both result relevance and ranking position, it assigns higher weights to top-ranked items, thereby assessing whether the model’s rankings align with user expectations. The mathematical expression of NDCG is given in Eq. (18).

$$NDCG = \frac{1}{N}\sum\limits_{u = 1}^{N} {\frac{DCG@k(u)}{{IDCG@k(u)}}}$$
(18)

In Eq. (18), \(DCG@k(u)\) denotes the cumulative loss gain of user u's recommendation list, while \(IDCG@k(u)\) represents the maximum ideal value. The NDCG value ranges from 0 to 1, with higher values indicating better ranking quality.

4.3 User Profiling Analysis Based on User Features and K-Means

To verify the analysis capability of the user profiling module based on user feature information and K-Means and its ability to construct user profiles, the study conducted simulation experiments to evaluate the module’s performance. The experiments ran on Windows 10 with an Intel Xeon Gold 6116R CPU at 2.90 GHz, 32 GB memory, using Python as the backend language. IBM SPSS Statistics 26 was used for data processing, and a word cloud analysis tool was used for visualization. The experiment uses a grid search to optimize the parameters of each basic model and identify the optimal parameter combination. The specific experimental parameter setting is shown in Table 1.

Table 1 Experimental parameter configuration
Full size table

Based on the above parameters, build a model for experimentation. To verify the feasibility of the user profiling module, the study compared the model with Naive Bayesian Classifier-Expectation–Maximization (NBC-EM), Word2Vec-Bidirectional Encoder Representation from Transformers (Word2Vec-BERT), and K-Means-Radial Basis Function Neural Network (K-Means-RBF). The hyperparameters of all models are tuned via grid search to ensure they are in an optimal state. Receiver Operating Characteristic (ROC), user feature coverage, and user profile accuracy were used to compare prediction performance. The results are shown in Fig. 7.

Fig. 7
Fig. 7
Full size image

Comparison of user profile prediction performance of different models

As shown in Fig. 7a, the ROC curve of the proposed user profiling module had an Area Under the Curve (AUC) of 0.92, higher than those of the comparison models, and was closest to the top-left corner, indicating better prediction performance. Figure 7b showed that the proposed module achieved user feature coverage of 0.91 and profile accuracy of 0.97, both of which were higher than those of the comparison models, indicating that the module covered more user features and achieved higher prediction accuracy. Overall, the user profiling module performed well in both accuracy and coverage, demonstrating its feasibility. Based on this verification, the study analyzed the basic information of platform users, as shown in Fig. 8.

Fig. 8
Fig. 8
Full size image

Analysis of basic user information on JD.com

Figure 8a showed that users aged 25–29 accounted for the largest proportion (35.10%) and that users aged 25–34 accounted for over 50%, representing the main users of JD.com from May to November 2023. This was attributed to this age group being in the career development stage. Figure 8b indicated that the main users were female, accounting for 73.88%, suggesting that the platform should focus on female shopping needs to design better personalized push strategies. Figure 8c showed that users aged 25–29 and 30–34 had the highest activity, with overall activity increasing in June and November, likely due to promotional events. To further analyze user behavior, users were categorized into browsing experts, collection experts, and purchase experts, as shown in Fig. 9.

Fig. 9
Fig. 9
Full size image

Analysis of user behavior on JD.com

Figure 9a showed that female users accounted for 78.12% of collection experts and were the primary users across all behavior categories. Figure 9b indicated that users aged 25–29 had the highest proportion in different behavior categories, 36.34%, 34.12%, and 32.01%, higher than other age groups. Overall, female users aged 25–29 had the largest proportion across different behavior categories, indicating the strongest purchasing power. These results demonstrated that the model effectively analyzed user information and achieved high prediction accuracy, thereby supporting the development of a personalized push strategy.

4.4 Analysis of an Improved Stacking Framework Based on Multiple Machine Learning

To evaluate the predictive performance of the improved Stacking framework, the study compared the model with Graph Neural Networks-Stacked Long Short-Term Memory (GNN-Stacked LSTM) and BERT-K-Nearest Neighbor (BERT-KNN). The hyperparameters of all models are searched by grid search to ensure that all models are in an optimal state. The experiment used the same dataset and experimental environment as mentioned above, with 100 iterations and a learning rate of 0.01. To validate the effectiveness of the confidence-weighted strategy employed in the study, an ablation experiment was conducted to compare the prediction accuracy of Standard Stacking (A), Stacking with probabilities (B), and Stacking with probabilities + confidence (C). The comparison results are shown in Fig. 10.

Fig. 10
Fig. 10
Full size image

Comparison of ablation experiments. Note: ** indicates p < 0.01

As shown in Fig. 10a, the prediction accuracy of Stacking with probabilities + confidence ranges from 88.00% to 93.25% in the training set, demonstrating high accuracy. Moreover, the median prediction accuracy of Stacking with probabilities + confidence is 91.35%, significantly higher than Standard Stacking and Stacking with probabilities (p < 0.01), indicating that the ensemble learning model using the weighted strategy of prediction confidence and prediction probability outperforms Standard Stacking and the stacking method that only applies prediction probability. Figure 10a shows that in the test set, the median prediction accuracy of Stacking with probabilities + confidence is 94.13%, significantly higher than Standard Stacking and Stacking with probabilities (p < 0.01), further validating its effectiveness. To validate the predictive ability of machine learning algorithms in the Stacking framework, the fitting performance of each algorithm was analyzed, as shown in Fig. 11.

Fig. 11
Fig. 11
Full size image

Fitting analysis of four machine learning algorithms

Figure 11a showed that XGBoost had the best fit, closely matching the real values. Figure 11b showed that RF differed slightly from actual data at a sales value of 150, but overall matched well. Figure 11c and d showed that LR and MLR predictions did not differ significantly from the real values, indicating good performance. Overall, all four algorithms closely matched the real data, demonstrating good predictive performance and supporting their integration into the Stacking framework. To further validate the performance of the improved Stacking framework, the study employed XGBoost, RF, LF, and MLR as baseline models, along with Adaptive Boost (AdaBoost) and Collaborative Filtering (CF), to conduct comprehensive performance testing. The results are presented in Table 2, which compares the improved Stacking framework with the contrast models.

Table 2 Comparison of prediction results of different models
Full size table

As shown in Table 2, the RMSE values for the training set were all below 20 for the RF, LF, and XGBoost algorithms, indicating that these three algorithms have strong predictive performance and can help the Sacking model achieve more accurate predictions. The RMSE value of the improved Stacking model was 6.82 ± 1.21, and the MAE value was 3.34 ± 0.77, which were significantly lower than those of the comparison algorithms (p < 0.05), indicating that this model has good predictive performance. This is because the improved Stacking model integrates the predictive effects of multiple models. To further evaluate the model, the coefficient of determination (R2) was used to compare the improved Stacking model with other models over 10 experiments, as shown in Fig. 12.

Fig. 12
Fig. 12
Full size image

Comparison of predictive performance of different models

Figure 12a shows that, across 10 experiments on the training set, the maximum R2 of the improved Stacking model reached 0.99, significantly higher than that of other models, indicating better fitting due to the integration of multiple algorithms. Figure 12b showed that in the test set, the maximum R2 of the improved Stacking model reached 0.98, still significantly higher than other models, demonstrating good stability. This stability improvement resulted from using five-fold cross-validation during training. Overall, the improved Stacking model outperformed the comparison models in both stability and fit, demonstrating superior predictive performance for user information.

4.5 Application Performance Analysis of the Improved UI-Stacking Model

Based on the performance evaluation of the improved Stacking framework, the study further analyzed the practical application of the UI-Stacking model. The study compared the UI-Stacking model with Recency-Frequency-Monetary-Collaborative Filtering (RFM-CF), Particle Swarm Optimization-K-Means (PSO-K-Means), and Heterogeneous Graph Neural Networks (HGNN) to validate model performance. The Taobao dataset, which contained abundant product information and complete user data, was selected for simulation experiments to evaluate the practical recommendation performance of the UI-Stacking model. The dataset spanned 15 days and included over 100,000 adult users aged 18–25, 4,758 products, and 855,651 product interactions. Data cleaning and augmentation were performed, yielding approximately 400,000 user-product interaction records. The experiments were conducted on Windows 10 with an Intel Xeon Gold 6142 CPU and an RTX A4000 GPU, 16.9 GB of memory, using Python as the backend language and PyTorch as the deep learning framework. The maximum behavior sequence length was set to 50, and the learning rate was set to 0.1. To evaluate the model’s recommendation performance, the study compared user push success rates, prediction accuracy, and recall rates across different recommendation list lengths. The results are shown in Fig. 13.

Fig. 13
Fig. 13
Full size image

Recommendation performance under different list lengths

As shown in Fig. 13a, when the recommendation list length was 5, the UI-Stacking model achieved a recall rate of 0.81 and a prediction accuracy of 0.83, significantly higher than the comparison models, demonstrating better predictive performance. Figure 13b showed that when the list length was 10, the UI-Stacking model performed best, achieving a push success rate of 0.64 and a recall rate of 0.82, indicating optimal performance at this list length. Figure 13c showed that when the list length was 20, the recall rate was 0.80 and the push success rate was 0.62, both still higher than those of the comparison models, demonstrating the model’s greater stability. The Normalized Discounted Cumulative Gain (NDCG) directly measures how effectively a model prioritizes highly relevant information, thereby reflecting its impact on information validity, user satisfaction, and task completion efficiency. To further analyze the model’s recommendation performance, the study compared NDCG and F1 score metrics across different models, with the results presented in Fig. 14.

Fig. 14
Fig. 14
Full size image

Comparison of recommendation performance among models

Figure 14a shows that with 10 recommended items, the RLXM model proposed by H Research achieves an average NDCG score of 0.167, significantly outperforming the benchmark models (0.157, 0.147, and 0.139) in recommendation quality, owing to its in-depth user profiling. Figure 14b further demonstrates that the RLXM model maintains a consistent F1 score of 0.15 across 10 experiments, surpassing the benchmark models in both recommendation performance and stability. Overall, the results demonstrated that the UI-Stacking model outperformed the comparison models in both stability and prediction accuracy, indicating that the model effectively analyzed user information and delivered accurate push notifications, thereby promoting the development of the e-commerce field.

5 Discussion and Conclusion

5.1 Discussion

The study proposed a user profiling module based on user information features and K-Means clustering, which could deeply analyze user data. According to the simulation experiment results, the study model constructed different categories of user profiles. Among them, the first typical user category consisted of users aged 25–29. This group showed a browsing rate of 93.12%, which was significantly higher than their purchase and favorite rates, and was classified as a browsing expert group. This indicated that the proposed model could deeply analyze user behavior and basic information, and construct different categories of user profiles based on this analysis. The results were consistent with those of Zhang et al. in 2024, although the proposed model provided a more in-depth analysis of user data [27]. Furthermore, the research proposes an enhanced Stacking framework based on multiple machine learning algorithms, demonstrating significant advantages in predicting user behavior information. In comparative experiments, the improved Stacking model achieved an RMSE of 6.82 ± 1.21 and MAE of 3.34 ± 0.77 on the training set, and 7.21 and 3.51 on the test set, respectively—both substantially lower than competing algorithms, showcasing high stability and prediction accuracy. This is attributed to the model’s adoption of three well-performing base models (Random Forest, Logistic Regression, and XGBoost) for prediction and training, combined with a Meta-Learner (MLR) algorithm. Through nonlinear weighted fusion, the Meta-Learner dynamically adjusts weights between base model predictions and intermediate information, significantly enhancing stability. In the fitting analysis, the improved Stacking model achieved a maximum R2 of 0.99 in the training set across 10 experiments and 0.98 in the test set across 10 experiments, both significantly higher than those of the comparison models, demonstrating superior stability and fitting performance. These results were similar to those of Labidi and Sakhrawi in 2023, but the proposed improved Stacking model showed better performance in prediction accuracy and fitting [28].

The UI-Stacking model, based on user information features and improved Stacking, also demonstrated outstanding performance in e-commerce applications. In terms of prediction accuracy, stability, and push performance, the proposed UI-Stacking model outperformed RFM-CF, PSO-K-Means, and HGNN. In the application effect experiments, the UI-Stacking model achieved optimal performance with a recommendation list length of 10, achieving a prediction accuracy of 0.83 and a push success rate of 0.64, both of which were higher than those of the comparison models, indicating better recommendation performance. When the recommendation list lengths were 5 and 20, the UI-Stacking model achieved recall rates of 0.81 and 0.80, respectively, which were still significantly higher than those of the comparison models, indicating higher prediction accuracy. This improvement is primarily attributed to the enhanced performance of the foundational models employed in the upgraded Stacking framework. Specifically, the Random Forest (RF) algorithm demonstrates strong generalization capabilities, the Logistic Regression (LR) algorithm effectively processes low-dimensional data and models explicit relationships between features, while the XGBoost algorithm exhibits superior learning capacity. The study utilizes these three base learners for predictions, incorporating prediction probabilities and confidence metrics. By employing the Multivariate Linear Regression (MLR) algorithm as a meta-learner to process outputs from the base learners, the framework achieves enhanced model stability and prediction accuracy. Compared with the conclusions proposed by Chen et al. in 2023, this approach demonstrates superior predictive performance [29]. Furthermore, the UI-Stacking model also showed excellent potential in recommendation effectiveness. Across 10 simulation experiments, the model achieved an average F1 score of 0.79, clearly higher than the comparison models, indicating stronger push effects and greater stability. When the recommendation list length was 10, the average NDCG was 0.167, significantly higher than that of the comparison models, demonstrating a better recommendation effect. This was attributed to the in-depth analysis of user information and the strong predictive performance of the Stacking framework. These results were similar to those of Bammou et al., who improved model prediction performance using ensemble algorithms. However, the proposed model exhibited advantages in both stability and fitting performance [30].

The research’s contributions are mainly reflected in the following two aspects. First, the RF, LR, and XGBoost algorithms are used as base learners, and the MLR algorithm is used as the meta-learner to construct an ensemble learning framework. The base learners’ prediction results and confidence levels are input to the meta-learner, enabling the ensemble learning framework to learn more abundant feature information. It has broken through the quality bottleneck of traditional integration technology. Second, the K-means algorithm was adopted to construct user portraits, and an e-commerce information push model was constructed in combination with the ensemble learning framework. Remarkable results were achieved in terms of push performance and recommendation effect, providing more scientific theoretical guidance for the precise push technology of e-commerce information and promoting the practical application of ensemble learning methods.

5.2 Conclusions and Recommendations

The study addressed the limitations of traditional recommendation algorithms in terms of accuracy and stability by proposing an e-commerce information push model based on user features and an improved Stacking framework. The study used RF, LR, and XGBoost as base learners in a Stacking ensemble framework, employed the MLR algorithm as the meta-learner, and combined user information features to construct an e-commerce information push model. Experimental results indicated that, in terms of prediction accuracy, stability, and recommendation performance, the proposed model outperformed the comparison models. The study achieved improvements in prediction accuracy and push effectiveness, but further enhancements were still needed. In the future, the model can be applied more broadly, adapted to different e-commerce scenarios, and further refined to improve its precision and practical application performance.