From Consumer Insights to Action: Reshaping Digital Sales and Shopping Experience Strategies for FMCG

1. Introduction

The FMCG industry faces challenges in the digital transformation context, including fragmented consumer demand, complex sales pathways, and delayed behavioral responses. Traditional sales strategies based on static profiles struggle to support precise conversion and efficient outreach. Addressing the optimization challenge of "transitioning from consumer insights to action," this study constructs a digital sales experience reconstruction mechanism centered on behavioral modeling, strategy generation, and multi-source data fusion. By employing sequence modeling and generative adversarial networks, it achieves personalized sales path prediction and strategy optimization, focusing on resolving core issues like untimely strategy responses and low recommendation relevance. The aim is to enhance the consistency and agility of sales efficiency and user experience.

2. Challenges in FMCG Digital Sales

In FMCG digital sales scenarios, data silos and disjointed data chains are particularly pronounced. Cross-platform consumer behavior trajectories are difficult to model uniformly, directly impacting the accuracy and timeliness of user profiling. The lack of unified coding interfaces for conversion data between mobile and offline channels results in data delays of up to 2.3 seconds between CRM and CDP systems, reducing user identification accuracy by approximately 18%1 . Furthermore, delays exceeding 1.5 seconds in data feedback between promotional strategies and personalized recommendations often cause missed golden conversion windows, creating a structural barrier of "high-frequency interference—low-frequency conversion." Regarding multi-platform collaboration and heterogeneous data integration, existing systems lack unified feature extraction standards, hindering iterative optimization of subsequent sales models. To address these issues, a more consistent data modeling and consumer behavior recognition mechanism must be established to provide high-quality inputs for strategy generation.

3. Digital-Based FMCG Sales Experience Reconstruction Model

3.1. Data Preprocessing and Feature Extraction

To achieve intelligent reconstruction of FMCG sales experiences, multi-source consumer behavior data must first undergo unified preprocessing and structured encoding2 . During data cleansing, events such as browsing, clicking, and adding to cart are normalized using timestamps to eliminate redundant behavior sequences and abnormal jump records. In the feature extraction phase, a sliding-window-based user behavior aggregation strategy is introduced to construct a multi-scale sequence feature matrix. To enhance the temporal correlation of behavioral representations, a feature weighting function is defined:

$$f_i={\displaystyle \frac{1}{1+e^{-\beta (t_i-t_0)}}}$$

(1)

where $f_i$represents the weight of feature at time $t_i$ , $t_0$denotes the center point of the behavior sequence, and $\beta$is the temporal decay coefficient controlling the amplification effect of recent behaviors. Embedding encoding is concurrently applied to perform low-dimensional mapping on discrete features such as user ID, product category, and terminal device, uniformly outputting a dimension of 32. The processed feature vector matrix serves as the foundational input for subsequent GAN strategy optimization modules and personalized sales path models, ensuring data consistency and structural adaptability in sales experience reconstruction algorithms.

3.2. Consumer Behavior Digital Model Construction

Following data preprocessing and vectorization encoding, the consumer behavior digital model abstracts multidimensional interactions through sequence modeling and fusion mechanisms. First, the user behavior state matrix $X\in {\mathrm{IR}}^{T\times d}$ is constructed, where $T$ denotes the time window length, and $d$ represents the single-step behavioral feature dimension 3 . A multi-head attention mechanism models the temporal patterns of user clicks, favorites, and adds-to-cart actions, generating context-dependent representations ${\mathrm{H}}_t$calculated as follows:

$${\mathrm{H}}_t=\mathrm{Softmax}\left({\displaystyle \frac{{\mathrm{Q}}_t\cdot K_t^{\mathrm{T}}}{\sqrt{d_k}}}\right)\cdot {\mathrm{V}}_t$$

(2)

where ${\mathrm{Q}}_t={\mathrm{X}}_t\cdot W_Q,{\mathrm{K}}_t={\mathrm{X}}_t\cdot W_K,{\mathrm{V}}_t={\mathrm{X}}_t\cdot W_V$ and $d_k$ denote the key vector dimensions. To enhance user decision path modeling, a behavior scoring function incorporating preference factors is introduced:

$$s_{u,i}={\displaystyle \sum _{j=1}^n\frac{w_j\cdot f_j(u,i)}{{\displaystyle {\sum }_{j=1}^n{w}_j}}}$$

(3)

Where$s_{u,i}$ represents user $u$ 's preference score for product $i$ ,$f_j(u,i)$ denotes the $j$ th behavioral feature function, and $w_j$ is its weight. A feature interaction module is further introduced to perform multi-order feature combinations on behavioral labels, page dwell time, purchase history, etc., constructing the final embedded representation for prediction and strategy generation. As shown in Figure 1, the model structure comprises an input layer, feature fusion layer, attention representation layer, and decision output layer, forming a closed-loop path learning mechanism. The distribution of relevant features and model dimension configuration are detailed in Table 1. Upon completion, this model will serve as the foundational input structure for subsequent strategy optimization modules and support the dynamic feedback mechanism of generative algorithms.

3.3. Sales Strategy Optimization Based on Generative Adversarial Networks

To dynamically generate sales strategies tailored to diverse consumer behavior paths, a strategy optimization model based on Generative Adversarial Networks (GANs) is constructed. The generator ($G(\cdot )$ ) generates candidate strategy sequences, while the discriminator ($D(\cdot )$ ) evaluates their consistency with the distribution of actual sales paths. During the strategy generation phase, the input noise vector ($z~N(0,I)$ ) is transformed into latent behavioral trajectories ($\widehat{s}=G(z)$ ) via embedding mapping. The generator weights are optimized through the following objective function:

$$L_G=-{\mathrm{IE}}_{\mathrm{z}\text{~}N(0,I)}\left[\log D(G(z))\right]$$

(4)

Among these, $L_G$ represents the generator loss function, while $D(G(z))$ denotes the discriminator's probability prediction for samples generated as "real paths." The discriminator employs a multi-layer fully connected structure with residual modules, whose loss function is defined as follows:

$$L_D=-{\mathrm{IE}}_{\mathrm{s}\text{~}{p}_{\mathrm{real}}}\left[\log D(s)\right]-{\mathrm{IE}}_{\mathrm{z}\text{~}N(0,I)}\left[\log (1-D(F(z)))\right]$$

(5)

where $s$ represents the true user path data, and $p_{\mathrm{real}}$ denotes the true policy distribution. During training, the generated policy dynamically adjusts its weighted modeling for conversion rate, page click weight, and temporal continuity. The configuration of its policy control factors is shown in Table 2. After multiple training rounds, the generator progressively learns nested patterns of high-frequency purchasing behavior Error! Reference source not found. , as illustrated in Figure 2. Early training stages exhibit scattered distributions across broad areas with limited strategy coverage, while later stages concentrate generated trajectories within high-conversion intervals. The optimized GAN module will serve as the foundational training basis for subsequent parameter tuning phases.

3.4. Model Training and Parameter Tuning

After constructing the generative adversarial model, we combine the alternating training mechanism of the discriminator and generator with an adaptive gradient adjustment strategy to ensure stable convergence within the non-convex space. Initial training employs the Adam optimizer for joint training of both modules, with learning rate $\eta =2\times {10}^{-4}$ , batch size 64, discriminator update frequency $nD=5$ , and parameter normalization triggered after each generator iteration4 . To ensure model convergence under non-convex optimization, we incorporate a smoothing loss function with a gradient regularization term:

$$L_{\mathrm{smooth}}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N{({\widehat{y}}_i-y_i)}^2}+\lambda \cdot {‖{\nabla }_{\theta }{\widehat{y}}_i‖}_2^2$$

(6)

where ${\widehat{y}}_i$ represents the model's predicted path strategy score, $y_i$ denotes the actual behavior data feedback score, $\theta$ is the network parameter, and $\lambda$ is the regularization coefficient controlling the smoothness of gradient changes. The training progresses through a warm-up phase and a convergence phase, as shown in Figure 3. During the initial 30 epochs, generator loss fluctuates while the discriminator accuracy steadily improves. After epoch 40, convergence stabilizes, and the training error approaches saturation, reflecting the benefit of regularization in maintaining training robustness.

4. Experiments and Results Analysis

4.1. Experimental Dataset and Evaluation Metrics

The experiment utilizes 18 months of transaction and behavioral data from an e-commerce platform in a province in East China. The dataset comprises 489,732 records covering user IDs, behavioral sequences, product attributes, and order histories, with 27 data dimensions updated daily. Evaluation metrics include AUC, Precision@10, Recall, and NDCG@10, assessing both recommendation accuracy and ranking capabilities. (1) Behavioral sequence data (2) Product metadata (3) User basic information (4) Historical conversion paths. This dataset serves as input for the parameter tuning phase, supporting subsequent experimental setup and comparative validation5 .

4.2. Model Hyperparameter Optimization

To optimize the convergence performance and ranking capability of the policy generation model under large-scale behavioral data, grid search was employed to adjust key parameters such as learning rate, embedding dimension, and batch size while keeping the network architecture constant. All combinations were evaluated on the validation set using AUC, NDCG@10, and HitRate@10 to measure the balance between preference learning and target click sequence fitting capabilities. Parameter configurations and corresponding metric results are shown in Table 3.

As shown in Table 3, increasing the embedding dimension from 64 to 128 yields a significant improvement in ranking accuracy (with NDCG@10 rising from 0.6614 to 0.6798); however, further expansion to 256 leads to diminishing returns. A learning rate of 0.0002 achieves better convergence stability, with AUC fluctuations remaining below 0.015. While enlarging the batch size improves HitRate, it provides limited gains in overall ranking performance. The tuning configuration above will serve as the baseline input parameters for the comparative experiments in the following section6 .

4.3. Comparative Results of Sales Strategy Experiments

After hyperparameter optimization, to validate the sales path modeling capabilities of strategy generation models under different architectures, three mainstream models were selected: the standard GRU sequence model, the Transformer-based behavioral modeling model, and the aforementioned GAN-based strategy generation framework. Their performance across dimensions including click-through rate prediction, conversion rate estimation, and ranking accuracy was tested under unified training and validation set divisions7 . The evaluation results for each metric are shown in Table 4, with heatmaps (Figure 4) visually illustrating the multidimensional performance differences in strategy response quality.

As shown in Table 4, the Transformer model with attention mechanism outperforms the GRU model by +0.0129 on the ranking metric (NDCG@10). Meanwhile, the GAN model demonstrates improved performance across all five metrics. Specifically, the average strategy path length increased from 4.92 to 5.34, which was statistically significant under a two-tailed paired t-test (t = 4.51, p < 0.001). This result reflects the model's capacity to more effectively capture diverse behavioral preferences through extended recommendation sequences.

4.4. Model Robustness Analysis

To systematically evaluate model robustness under real-world scenarios involving abnormal inputs, user cold starts, and behavior loss, three typical perturbation scenarios were defined10. The magnitude of changes in key metrics before and after perturbations was compared ( ), with the GAN policy model as the evaluation subject. Under a unified test set structure, three performance metrics—AUC, NDCG@10, and HitRate@10—were recorded along with their deviations from the baseline state12. The experimental results are shown in Table 5, with data variations covering common proportions of behavior missing (30%), cold-start users (20%), and access delay interference (5-second window) in real-world scenarios.

Table 5 shows that behavioral missing data has the most significant impact on NDCG@10, with a deviation of -0.0356, indicating high ranking sensitivity. Meanwhile, AUC exhibits smaller fluctuations under cold-start and delay conditions, with the maximum decline not exceeding 0.0261, suggesting the backbone structure possesses a certain buffering capacity against sample imbalance. HitRate remains generally stable, with average deviations from single-point perturbations not exceeding 0.02, indicating high output consistency in strategy recommendations despite user behavior fluctuations. Subsequent work will further explore the model's deployment adaptability in dynamic scenarios by integrating these perturbation behaviors8 .

5. Conclusions

For FMCG digital sales scenarios, we constructed a multi-level modeling framework spanning behavioral data preprocessing to strategy generation, integrating generative adversarial mechanisms to achieve personalized sales path optimization. The proposed model demonstrates superior ranking capabilities and path coverage efficiency across multi-dimensional metrics, exhibiting robust performance and strong generalization. The innovation lies in introducing behavioral cross-modeling and GAN-based feedback generation, effectively mitigating data chain disconnects and recommendation failures. However, the model still faces challenges such as insufficient understanding of cold-start users and limitations in cross-platform data integration. Future work could incorporate large language models for deep semantic decoding of behaviors and explore real-time update mechanisms to enhance strategy responsiveness and effectiveness.