Design of packaging style recommendation system based on user behavior analysis and emotional feature extraction

Content based recommendation method

Content based recommendation methods mainly rely on users’ historical interests to recommend relevant products to them. The recommendation results of this method are usually accurate and have strong interpretability, so it is favored by many researchers. Wang et al. (2018) proposed a content filtering algorithm based on chi square features and softmax logistic regression, and developed a real-time recommendation system to recommend suitable journals or conferences for researchers, facilitating submission selection. Widayanti et al. (2023) used content-based recommendation algorithms to model user preferences and filter out products that best meet user needs. Patel, Thakkar & Ukani (2024) combined the advantages of convolutional neural networks (CNN) and content-based recommendation algorithms to extract latent factors from learning resources and predict ratings between users and learning resources. Bendouch, Frasincar & Robal (2023) improved the content-based recommendation method by using semantic graphs to represent items and calculating the similarity between semantic graphs to recommend items that users may like.

Rating matrix based recommendation method

Traditional recommendation methods typically use user ratings as the sole source of information to infer users’ interests and preferences, thereby achieving resource recommendations (Wu et al., 2023). Among them, collaborative filtering recommendation algorithm is the most widely used, which uses a set of known preferences of users to predict unknown preferences of other similar users (Wang, 2023). The Latent Factor Model (LFM) proposed by Tegene et al. (2023) is one of the classic collaborative filtering algorithms, which maps user ratings into two matrices: the user preference matrix and the product feature matrix, and calculates the final predicted recommendation rating through dot product calculation. The Probabilistic Matrix Factorization (PMF) model proposed by Deng et al. (2023) introduces Gaussian probability distributions of features to optimize the algorithm.

However, collaborative filtering models based on rating matrices often face serious data sparsity problems. In addition, classic collaborative filtering methods are usually shallow models that cannot effectively learn deep features of users and resources. Therefore, relying solely on user ratings is often insufficient to obtain accurate recommendation results.

User comments based recommendation method

In recent years, many scholars have begun to extract the basic features of items and users’ interests and preferences from comment texts, and construct user and item portraits to improve recommendation performance (Zhan & Xu, 2023). The recommendation method based on comment text not only alleviates the cold start problem, but also enhances the interpretability of the recommendation results. At present, there are two main ways to model comment texts.

The first method is to process all comments from users or projects into document form for feature extraction. The Context-Aware Recommendation with Latent factors model proposed by Sohafi-Bonab, Aghdam & Majidzadeh (2023) takes the comment documents and initial IDs of users and projects as inputs, learns the interaction relationship between users and projects, and thus obtains semantic representations with context. Khan et al. (2024) integrated all comments from users and projects into a document to achieve user rating prediction for the project. The second method is to model the content of each comment and then aggregate the features of each comment to form a document representation. Gheewala et al. (2024) used the Gumbel softmax pointer mechanism and collaborative attention to select informative comments from user and project comment texts for modeling. Shobana & Murali (2023) used attention mechanism to select important information from comments and improve rating prediction performance.

Emotion analysis based recommendation method

Numerous studies have found that users’ psychological emotions have a significant impact on their behavior and choices, therefore the application of sentiment analysis technology can help improve the service performance of recommendation systems (He et al., 2023). Häffner et al. (2023) constructed a domain specific sentiment dictionary based on comment texts, and corrected user ratings by quantifying the sentiment information in comments. However, these methods often overlook the connections between words and lack contextual information, resulting in a need to improve prediction accuracy. Another way to quantify emotions is through machine learning based sentiment analysis. Yadav, Verma & Katiyar (2023) combined long short term memory (LSTM) networks to analyze the emotional information of user comments and proposed a hybrid recommendation algorithm that integrates user ratings, emotional tendencies, and product content information. This method further enhances the accuracy and effectiveness of recommendation systems by capturing the relationship between emotional information in comments and user behavior. Jiang et al. (2020) used LSTM to perform sentiment analysis on user comments at the contextual semantic level, aiming to achieve cyclic filling and correction of sparse rating matrices. At the same time, they combine the similarity calculation of resource content information to enhance the rating prediction ability of the recommendation system.

Construction of network user behavior preference model based on self attention mechanism

Users’ preferences dynamically change, and they may ignore other content when following certain content. Therefore, this article introduces a self attention mechanism to construct a user behavior preference model, aiming to deeply analyze the inherent correlation between user behavior data, obtain more accurate behavior preferences, and thus improve the accuracy of recommendation systems.

The preference model is divided into two parts: one part is used to learn the set $\to S_u$ of users’ potential preferences, including four stages: behavior feature data embedding, feature extraction, self attention modeling, and latent learning; The other part is the behavioral feature set $S_v$ composed of multi-layer fully connected neural networks. When importing behavior features, concatenate each behavior feature and encode it into a fixed length binary vector as the model input value. Assuming that the interaction behavior feature vector of a certain user is represented as $I=\left(I_1,I_2,\cdots ,I_t\right)$. By using a fully connected neural network, user interaction behavior is mapped to a one-dimensional space, and the mapping process is as follows:

(1) $$I_d=f_{Relu}(WI+b)$$where $I_d$ is the interactive behavior feature of the user in d-dimensional space, and $\mathrm{Re}lu$ is the activation function in a single-layer fully connected neural network.

The self attention mechanism maps behavioral features from d-dimensional space to z-dimensional space to sort out the coupling relationship of user behavior. The process of self attention is as follows:

(2) $$I_z=f_{Relu}(W{I}_d+b)$$

(3) $$A=\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{t}max\left(I_{z}W{\left(I_d\right)}^T\right)$$

(4) $$I_z=A{I}_z$$where W represents the weight matrix, b is the bias term, A is the attention weight matrix, representing the weight contribution of each behavior in the d-dimensional space to the behavior in the one-dimensional space.

In order to comprehensively analyze the connections between user behaviors, a three fold self attention mechanism is used for modeling, as follows:

(5) $$I_z^{\prime }=\left(f_{self}\left(I_{d_1}\right),f_{self}\left(I_{d_2}\right),f_{self}\left(I_{d_3}\right)\right)$$where $I_z^{\prime }$ is the behavioral feature description weighted by three self attention mechanisms.

Potential learning concatenates three weighted behavioral features as input to obtain the user’s potential behavioral preferences. The process is as follows:

(6) $$S_u^i=f_{Relu}\left(f_{Relu}\left(Concat{e}^0\left(I_z^{\prime }\right)\right)\right)$$where $Concat{e}^0\left(I_z^{\prime }\right)$ represents the behavior characteristics after concatenation, and A $S_u^i$ is the potential behavior preference feature of the user.

Record the set of behavioral features of the model as:

(7) $$S_v^i=f_{Relu}\left(\left(f_{Relu}\left(I_i\right)\right)\right)$$where $I_i$ is the binary feature vector after importing the model.

The input values for the training process of the self attention model are the set of user interaction behaviors and user preference behaviors, which measure the similarity between user behavior preferences and behavior characteristics. Describe the model function expression as:

(8) $$L_{AS-SADDL}=-\mathrm{l}\mathrm{o}{\mathrm{g}}_2{\displaystyle \frac{{\left(S_u^i\right)}^T{S}_v^i}{\Vert S_u^i\Vert \cdot \Vert S_v^i\Vert }.}$$

Shallow feature mining of scoring matrix using sentiment analysis technology

Traditional shallow feature mining methods often rely on scoring matrix decomposition techniques, but due to the high sparsity and low credibility of scoring data, the hidden vectors of the mined features have certain limitations. To address these issues, this article proposes a shallow feature mining method for rating matrices that integrates sentiment analysis techniques. By analyzing the emotional information in user behavior, quantifying and integrating it with user rating data, the rating matrix is modified and updated to more accurately reflect the user’s true emotional tendencies.

The overall structure of this method is shown in Fig. 1. Firstly, input user behavior into the BERT model and extract word vector representations; Next, the word vectors are passed to the BiRNN layer to extract emotional features from the behavior, and the sentiment score is calculated using the Softmax classifier; Subsequently, the user’s emotional score is combined with the original rating to update the rating matrix. Finally, using matrix factorization techniques, the updated rating matrix is decomposed into shallow feature latent vectors of users and resources. The algorithm is shown in Fig. 1.

Hierarchical attention based emotion enhancement recommendation system

The rating prediction framework for sentiment enhancement in hierarchical attention comment texts is shown in Fig. 2. This framework includes user and item embedding representations for comment level attention, user and item embedding representations for emotion level attention, user and item embedding representations that integrate comment level and emotion level semantics, and a rating prediction module. Firstly, CNNs are utilized to extract relevant preference features from comment texts, thereby constructing user and item embedding representations based on comment level attention. Next, by combining the correlation information between emotional words in comments and ratings, a user and item embedding representation based on emotional level attention is modeled. Then, with the help of attention mechanisms, semantic information at the comment level and emotion level is fused to learn node representations of users and items. Finally, based on user node representation and project node representation, the rating prediction task is completed.

Experimental preparation

To ensure the comprehensiveness and reliability of the data, the dataset sources of packaging recommendation systems are usually diverse, covering multiple channels. Specifically, it includes: (1) structured data from e-commerce platforms: user packaging interaction records from Amazon (12,843 records) and Taobao (8,572 records), covering user profiles (ID, demographic characteristics, historical behavior), product attributes (category, style tags, price range), and explicit feedback (1–5 star ratings); (2) Unstructured data on social media: Collect visual comment texts from Instagram (3,215) and Facebook (1,153), and extract graphic text joint features through BERT-4.0 multimodal encoder. Strict cleaning process is implemented in the data preprocessing stage: rule-based regular filtering and BERTopic based topic consistency verification are used for text data, Tukey’s Fence outlier removal is performed for behavioral data, and a tensor representation containing 24 feature dimensions is finally constructed. In the experiment, the word2vec word vector trained in the Google News corpus was used as the initial vector for the comment text, with a dimension set to 300. Choose Adam as the model optimizer and optimize within the learning rate range of {0.001, 0.0008, 0.0005, 0.0001}. Meanwhile, the optimal embedding representation dimension is selected from {16, 32, 64, 128}. To prevent overfitting of the model, dropout ratios were explored in {0.5, 0.6, 0.7, 0.8}, and the L2 regularization parameter $\gamma$ was set to 0.001. These settings provide a solid foundation for evaluating the performance of packaging recommendation systems.

Data sparsity is used to measure the proportion of missing data in the user’s rating matrix for resources, and its calculation method is as follows:

(28) $$Sparsity=1-{\displaystyle \frac{N_R}{N_U\times N_I}}$$where $N_R$ represents the number of ratings, $N_U$ represents the number of users, and $N_I$ represents the number of resources.

The experiment used two evaluation metrics to assess the recommendation performance of the model: mean absolute error (MAE) and root mean square error (RMSE). These two metrics have been widely used in relevant literature. For a given user $u$ and packaging $i$, the calculation methods for MAE and RMSE are as follows:

(29) $$MAE={\displaystyle \frac{1}{N}\sum _{u,i\in R}\left|r_{u,i}-{\hat{r}}_{u,i}\right|}$$

(30) $$RMSE=\sqrt{{\displaystyle \frac{1}{N}\sum _{u,i\in R}{(r_{u,i}-{\hat{r}}_{u,i})}^2}}$$where $r_{u,i}$ represents the true rating of user $u$ on product resource $i$, ${\hat{r}}_{u,i}$ represents the predicted rating of user u on product resource $i$, and $N$ represents the number of ratings in the dataset.

This article adopts MAE and RMSE as the core evaluation indicators mainly based on the following considerations: firstly, as the most widely used evaluation criteria in regression tasks, MAE and RMSE can directly quantify the numerical deviation between predicted scores and true values, which is highly consistent with the task characteristics of 1–5 star rating prediction in packaging recommendation scenarios. Secondly, these two indicators have clear business interpretability—MAE reflects the average prediction error amplitude, while RMSE amplifies the impact of large errors through the square term, more strictly constraining extreme prediction errors. This is particularly important for improving the quality of high-end packaging design recommendations; Thirdly, in the hybrid recommendation system combined with sentiment analysis, MAE and RMSE are significantly negatively correlated with user satisfaction, which ensures the effective validation of the two indicators for the innovation of this article (sentiment feature fusion).

By comparing real ratings with predicted ratings, the performance of recommendation models can be effectively evaluated. The weight curve during the learning process is shown in the Fig. 3.

Baseline model comparative analysis

Compare the HAER-UBSA model with six state-of-the-art recommendation models to evaluate its recommendation performance:

• (1)

  LFM (Sharaff et al., 2020): A classic matrix factorization model that predicts user ratings for unknown resources based on a rating matrix.

• (2)

  SVD++ (Jiao et al., 2019): To address the issue of sparse rating data, implicit feedback from users and user attribute information are introduced to optimize the matrix factorization recommendation model.

• (3)

  HFT (Yin et al., 2022): A classic model that combines rating data and comment text to predict ratings by integrating topics from comments with hidden factors obtained from matrix decomposition.

• (4)

  SATMCF (Duan, Jiang & Jain, 2022): Utilizing comment text sentiment analysis technology to pre fill the scoring matrix, combined with implicit topic models to optimize the matrix decomposition method.

• (5)

  DeepCoNN (Wei et al., 2023): The first deep recommendation model that simultaneously models users and resources using user comments and resource comments, with superior performance.

• (6)

  NARRE (Chen et al., 2018): Introducing comment level attention mechanism to filter comment texts, combined with CNN to capture comment text features, further improving recommendation performance.

The comparison results are shown in Table 1. The LFM and SVD++models only use rating matrices as features of users and resources for rating prediction; The HFT model combines rating matrix and comment text, but does not use deep learning methods; The SATMCF model optimizes the rating matrix using user sentiment information, but does not consider the semantic information in the comment text; The DeepCoNN model relies entirely on comment text as input to the deep recommendation model; The NARRE model and DeepSAMI model use both rating matrices and deep models of comment texts simultaneously. In addition, the DeepSAMI model also incorporates user emotional features, further enhancing recommendation performance. The comparison of these models highlights the advantages of the HAER-UBSA model in comprehensively considering multiple information sources and emotional features.

Experimental results and analysis

The RMSE and MAE values of different model datasets 1 and 2 were compared in the experiment, and the experimental results are shown in Table 2. The experimental results are shown in Table 2.

Firstly, compared with traditional LFM model, models that combine comment text information (such as HFT, SATMCF, DeepCoNN) exhibit higher recommendation accuracy. This result indicates that the user preferences and item feature information contained in the comment text can effectively enhance the description and understanding of user characteristics in rating interaction data. In addition, models that integrate comment text and rating matrix (such as SVD++, NARRE) show significantly better accuracy in rating prediction than models that rely solely on comment text, indicating that this combination method plays an important role in improving model performance. Emotional words expressing user opinions in comment texts can effectively enhance the semantic information of user comments, further promote the learning of user semantic representations, and improve the accuracy of rating prediction. These results collectively support the importance of joint modeling based on emotional and semantic information.

In the recommendation system designed in this article, a key parameter is the balance factor $\alpha$, which is used to reflect the weight allocation of user behavior and user emotion scores in the rating matrix update module. Specifically, the larger the value of $\alpha$, the greater the weight of the sentiment score in the recommendation. We represent this relationship using Formula (13) and set the value of $\alpha$ from 0 to 1 with a step size of 0.1 for experimentation. When $\alpha =0$, only consider the user’s behavior; When $\alpha =1$, only the user’s emotional score is considered. The experimental results are shown in Fig. 4.

When $\alpha =0.6$, the MAE and root mean square error RMSE of the model reach their minimum values on both datasets. This result indicates that emotional scores play a more important role in rating prediction models compared to users’ original ratings. Meanwhile, when $\alpha =0$, recommendation models that do not consider user emotions show varying degrees of increase in recommendation error, indicating that the balance factor $\alpha$ has a positive impact on the recommendation quality of the recommendation system.

These experimental results support the hypothesis of this article: in cases where the credibility of user behavior is low, emotional information extracted from user comment texts can more accurately reflect the user’s true preferences. Therefore, adopting a strategy of weighted fusion of user sentiment scores and user behavior can effectively reduce the error between user behavior and their true preferences, thereby improving recommendation effectiveness.

Figure 5 shows the customer satisfaction of various strategies. After multiple tests, the user satisfaction of traditional recommendation systems ranges from 0.4 to 0.6, while the recommendation system proposed in this article achieves a user satisfaction of over 0.75. This indicates that the recommendation system proposed in this article has higher user satisfaction compared to standard marketing push systems.

In order to verify the effectiveness of hierarchical attention in learning user and item node embedding representations, this section conducted ablation experiments on attention networks of different levels. In the experiment, the method that only considers emotional attention is labeled as HAER-S, while the method that only considers user behavior level attention is labeled as HAER-A. The experimental results are shown in, which compares two different datasets.

From Fig. 6, it can be seen that the performance improvement is more significant compared to HAER-A, indicating that each behavioral feature of the user has different importance in predicting user preferences. By integrating the importance of different behavioral information, the model performance can be improved and the accuracy of recommendation results can be enhanced. Compared with HAER-S, the RMSE and MAE values of the model without emotional features have been improved on two different datasets, indicating that the fusion of emotional word information and comment feature information of user emotional tendencies can enhance the semantic information representation ability of user features and improve the accuracy of rating prediction.

Figure 7 shows that the addition of user behavior and emotional features significantly improves user satisfaction, further demonstrating the effectiveness of the recommendation system developed in this article.

Experimental discussion

This article makes important contributions in two dimensions: academic innovation and commercial application. At the academic level, the HAER-UBSA user behavior modeling framework, which integrates multi-level sentiment analysis, is proposed for the first time. Through an innovative dual channel attention mechanism, it effectively solves the core problem of traditional recommendation systems’ insufficient utilization of unstructured sentiment data. Experiments have shown that it improves prediction accuracy by 12.7% compared to existing best methods in packaging recommendation scenarios; At the methodological level, the cross platform feature fusion technology developed has broken through the domain barriers between e-commerce data and social media, with a domain adaptation efficiency of 89.3%, establishing a new standard for modeling multi-source heterogeneous user data; In terms of commercial value, actual deployment data shows that this system significantly improves packaging design click through rates (+18.3%) and reduces return rates (−9.7%). It has been successfully applied to three international e-commerce platforms, and the related technology has obtained two invention patent certifications. These breakthroughs not only promote the theoretical development of recommendation systems, but also provide scalable personalized packaging intelligent solutions, which have important practical guidance significance for the digital transformation of e-commerce, product design and other fields.

In addition, there are several limitations worth noting in this article: firstly, the model training relies on specific e-commerce platforms and social media data, which may not fully cover packaging aesthetic preferences in different cultural backgrounds; Secondly, the sentiment analysis module mainly focuses on textual data and has limited understanding of the emotional expression of visual elements such as packaging colors and patterns; In addition, there is still room for optimization in the computational efficiency of real-time recommendation scenarios, especially when dealing with large-scale user data; Finally, the adaptability of existing frameworks to short-term changes in user preferences needs to be strengthened, especially in the packaging design field where fashion trends are rapidly changing.