2.2.1 Aesthetic features of post images and post popularity.

Composition and visual balance guide viewers’ attention and reduce cognitive load, making image information easier to process [23,24]. Foreground-background contrast increases subject salience and supports efficient visual communication, consistent with Gestalt principles [25,26]. Color attributes such as hue, saturation, brightness, clarity, and contrast further affect attention and affective response: warm colors and high saturation can enhance attention and emotional arousal, while low clarity or excessive contrast may increase cognitive load [27–29].

2.2.2 Content features of post images and post popularity.

Image content features such as faces and brand logos can also affect engagement. Faces can shape consumers’ affective and informational responses to online visual content [30]. In social media settings, consumer-selfie images featuring faces have also been associated with higher levels of likes and comments [7]. However, they may also distract attention from product or brand information [30]. Similarly, brand logos serve as visual signals that support recognition and emotional connection [31,32], but overly prominent logos may heighten persuasion awareness and reduce engagement [31].

2.3.1 Fundamental textual features of posts and post popularity.

Text length serves as a fundamental metric for assessing the informational richness of posts and has a significant impact on user engagement and post popularity. Studies have shown that longer texts often contain more information, effectively capturing user attention and fostering interaction. However, excessively long texts may increase cognitive load, thereby reducing users’ willingness to engage and diminishing post popularity [33,34].

2.3.2 Interactive textual features of posts and post popularity.

Interactive features within posts play a crucial role in driving user engagement and enhancing post popularity. Studies have shown that the use of interactive elements (such as hashtags “#” and mention tags “@”) significantly increases post visibility, thereby attracting greater user attention and engagement [6]. Furthermore, the strategic use of interactive components not only expands the reach of information, enabling sustained exposure across a broader audience, but also effectively stimulates user interactions (such as likes, comments, and reposts), thereby positively influencing post popularity [35].

2.3.3 Content text features of post and post popularity.

Emotional tone, brand mentions, and other linguistic features in post text play a crucial role in shaping user engagement and enhancing post popularity. Studies have shown that posts containing positive emotional elements or explicit brand cues can evoke users’ emotional resonance and cognitive responses, thereby significantly increasing user engagement [12]. These textual features not only enhance the appeal of information but also effectively drive user interactions, providing strong support for the formation of post popularity [36].

3.2.1 Measurement of post popularity.

Post popularity is operationalized using three observable user engagement indicators: likes, comments, and reposts [14,42,43]. We also construct total engagement as the unweighted sum of these indicators to capture the overall volume of interactions received by a post. Because likes are more frequent than comments and reposts, this aggregate measure represents total interaction volume rather than an equally balanced multidimensional engagement construct. Therefore, likes, comments, and reposts are also analyzed separately.

3.2.2 Measurement of multimodal content and contextual features.

Aesthetic features of images are measured across three dimensions: visual balance, subject-background separation, and color characteristics. Color Visual Balance is computed as the average Euclidean distance of RGB values between mirrored pixels (left–right). Higher score indicates greater color imbalance. Subject-background separation includes size and color differences, calculated by the area ratio and the Euclidean distance of RGB values between the subject and background. Color characteristics include hue, saturation, value, definition, and contrast. Hue, saturation, and value represent the average color properties of the image, while definition is measured using the Laplacian operator, and contrast is calculated by the mean squared deviation of grayscale values from neighboring pixels.

Content features include the presence of faces and brand logos in the image, detected using the Baidu AI Cloud APIs. Textual features are divided into three categories: fundamental, interactive, and content features. Fundamental features are based on word count, interactive features include the number of hashtags (#) and mentions (@), and content features include brand name detection and sentiment orientation, determined via Python scripts and sentiment analysis (Baidu AI Cloud API). Additionally, post timestamps and follower count are incorporated as temporal and social influence indicators.

3.3.1 Regression model design.

To analyze the relationships between content features and post popularity, an OLS regression model was constructed, as represented by the following functional form:

(1)

(1.1)

(1.2)

(1.3)

In this model, post popularity is quantified as the sum of likes, comments, and reposts, reflecting the overall engagement level of a post. This aggregate metric has been widely used in prior literature as an indicator of post popularity on social media platforms. The independent variables include vectors for image aesthetic features (Img_aes), image content features (Img_cont), textual features (Text_feat), and control variables. Because comments and reposts include zero values, all engagement outcomes were transformed using , where x denotes the original number of likes, comments, reposts, or total engagement. This transformation retains zero-valued observations while reducing the influence of the highly right-skewed engagement distributions.

However, the dataset does not contain post-level measures of algorithmic exposure, paid amplification, or campaign support. Because these unobserved factors may be correlated with both content characteristics and engagement outcomes, the estimated coefficients should be interpreted as conditional associations rather than unbiased causal effects.

3.3.2 Prediction model design.

To predict post popularity, we develop a multimodal deep learning model that combines two complementary sources of information. One branch extracts latent visual representations from post images using VGG19, while the other processes interpretable visual, textual, temporal, and account-level features. The two representations are fused to support two prediction tasks: continuous engagement estimation and popularity-level classification. Fig 3 provides an overview of the model, while detailed architecture and training settings are reported in Section 5.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Architecture of the multimodal post popularity prediction model.

https://doi.org/10.1371/journal.pone.0356612.g003

4.1.1 Post popularity on social media platforms.

Table 2 reports the descriptive statistics for post popularity and related visual, textual, and contextual features. Engagement metrics show substantial variation, especially comments and reposts, which follow long-tail distributions with extreme values. This pattern is consistent with social media content dynamics, where only a small share of posts becomes highly popular while most receive limited engagement. Therefore, the skewed distribution supports our four-level popularity classification based on engagement percentiles.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive statistics of variables.

https://doi.org/10.1371/journal.pone.0356612.t002

4.1.2 Descriptive patterns of visual and textual features.

Visual inspection of the sample images reveals several patterns. Most images show dispersed color balance and high brightness (value), often skewed toward cooler hues. Foreground-background separation varies, with significant size and color contrast in many cases. Notably, 39% of the images contain human faces, while 29% contain identifiable brand logos.

Textually, the average post contains about 86 words, consistent with the platform’s short-form style. However, interactive elements are underutilized: the average number of hashtags (#) is 0.64, and mentions (@) appear in just 22% of posts. Approximately half include brand names. Sentiment analysis shows an overall positive tone, though negative and neutral posts still make up a notable minority.

These findings suggest that while Weibo brand content shows some visual and textual consistency, there is significant variability in content design, interactivity, and emotional framing. This variation provides a basis for examining the associations between multimodal features and post popularity in the following sections.

4.2.1 Associations of image aesthetic features with post popularity.

Regression analysis shows that the color visual balance metric is significantly negatively correlated with likes (β = −0.03, p = 0.00) and reposts (β = −0.05, p = 0.01). Because higher values indicate greater color imbalance, images with relatively more balanced color distributions tend to receive more likes and reposts [23,24].

In subject-background separation, size and color differences are not significantly associated with the engagement metrics, indicating that these separation measures provide limited explanatory information for engagement in the present sample.

Regarding color, hue is negatively correlated with likes (β = −0.03, p = 0.00), suggesting warmer tones may be more emotionally stimulating than cooler ones [44]. Saturation is positively correlated with comments (β = 0.10, p = 0.00), reposts (β = 0.14, p = 0.00), and overall engagement (β = 0.11, p = 0.00), indicating that posts with more saturated images tend to receive higher levels of engagement [45]. Value is negatively associated with likes (β = −0.09, p = 0.00) and comments (β = −0.05, p = 0.06), suggesting overly bright images reduce contrast and clarity, dampening user interest [29]. Definition is positively associated with reposts (β = 0.04, p = 0.04), likely because clear images are perceived as more professional and shareable [28]. Contrast is not significantly associated with post popularity.

4.2.2 Associations of image content features with post popularity.

In image content features, the presence of a face is negatively associated with likes (β = −0.05, p = 0.01) and comments (β = −0.07, p = 0.00), but positively correlated with reposts (β = 0.06, p = 0.09). Faces may divert attention from brand elements or reduce visual clarity, leading to fewer likes and comments [30]. The positive association with reposts may be consistent with greater perceived relatability or expressiveness [7].

In contrast, the presence of a brand logo is positively correlated only with likes (β = 0.05, p = 0.02), suggesting that its predictive relationship may be stronger for lightweight engagement than for more effortful forms of interaction.

4.2.3 Associations of textual features with post popularity.

Regression analysis shows that word count is negatively correlated with likes (β = −0.03, p = 0.00) but positively correlated with comments (β = 0.05, p = 0.00), reposts (β = 0.21, p = 0.00), and overall engagement (β = 0.11, p = 0.00). Shorter posts are more likely to get likes, while longer posts provide more information, encouraging comments and shares.

Hashtag count is positively correlated with likes (β = 0.13, p = 0.00), comments (β = 0.39, p = 0.00), reposts (β = 0.60, p = 0.00), and overall engagement (β = 0.46, p = 0.00), suggesting that hashtags improve discoverability and engagement [35]. Mentions (@) are positively correlated with comments (β = 0.10, p = 0.00), reposts (β = 0.10, p = 0.00), and engagement (β = 0.10, p = 0.00), indicating that posts containing more mentions tend to receive more comments and reposts [46].

The presence of a brand name is positively correlated with comments (β = 0.14, p = 0.00) and engagement (β = 0.08, p = 0.01), suggesting that mentioning the brand boosts user attention and discussion [12]. In contrast, positive sentiment is negatively correlated with comments (β = −0.07, p = 0.00) and reposts (β = −0.05, p = 0.04), indicating that emotionally charged or slightly negative content prompts stronger reactions than overly positive posts, which may seem bland or less urgent [11,38].

Days Since Posting is negatively associated with likes (β=−0.73, p = 0.0) and shows smaller negative associations with comments, reposts, and total engagement. This result indicates that more recently published posts tended to receive higher engagement in the observed sample, despite having less nominal time to accumulate interactions. This pattern may reflect changes over time in account reach, platform activity, or user behavior, as social media engagement is often concentrated shortly after publication. Therefore, the coefficient should be interpreted as a broad temporal association rather than as the isolated effect of engagement accumulation time. Although Days Since Posting is an important temporal predictor, several visual and textual features remain significantly associated with engagement after this temporal control is included, indicating that content features provide additional predictive information beyond the observed temporal context.

4.2.4 Robustness.

Breusch–Pagan tests indicate heteroskedasticity across the OLS specifications (all p < 0.001); therefore, we report HC3-robust standard errors. VIF diagnostics show high collinearity between the two subject–background separation metrics, so we conduct a robustness check by replacing them with a standardized composite index, SB_index = z(size difference) + z(color difference). After aggregation, all VIFs are below 3 (max = 2.80), suggesting no severe multicollinearity. Additional robustness checks using negative binomial models, GLM–negative binomial models, and quantile regressions provide broad support for the main findings, while also revealing some model-specific and tail-specific heterogeneity. Detailed results are reported in S1 Appendix.

5.2.1 Continuous prediction model.

The continuous prediction model adopts a two-branch architecture. The image branch uses VGG19 pretrained on ImageNet. The original top classification layers are removed, and global average pooling is applied to obtain deep visual representations. The structured-feature branch receives manually extracted image features, text features, posting-time variables, and account-level indicators. These features are standardized and then processed through fully connected layers. The two branches are concatenated, followed by dense layers with L2 regularization and Dropout. The output layer uses a linear activation function to predict log-transformed total engagement.

During model construction, the VGG19 backbone is initialized with ImageNet weights. Its convolutional layers are used for visual feature extraction, and the final layers can be fine-tuned to adapt the visual representation to the Weibo image context. The structured branch includes variables such as color balance, hue, saturation, brightness, clarity, contrast, face presence, logo presence, text length, sentiment, hashtag count, mention count, posting time, days since posting, and follower count.

The outputs from both branches are concatenated and processed through a fully connected network. To prevent overfitting, L2 regularization and Dropout are applied. The final output layer generates a continuous value for post popularity, and the model is trained using the Adam optimizer with mean squared error (MSE) as the loss function.

5.2.2 Classification prediction model.

For the classification task, posts are divided into four popularity levels using the 10th, 90th, and 99th percentiles of total engagement. Posts below the 10th percentile are classified as low-popularity posts; posts between the 10th and 90th percentiles are classified as moderate-popularity posts; posts between the 90th and 99th percentiles are classified as high-popularity posts; and posts above the 99th percentile are classified as viral posts. Accordingly, the four popularity levels should be interpreted as categories of aggregate interaction volume rather than balanced profiles of likes, comments, and reposts.

This percentile-based scheme is designed to reflect the long-tailed distribution of social media engagement. The moderate category represents routine engagement outcomes, whereas the high and viral categories distinguish strongly performing posts from the small number of exceptionally popular posts in the upper tail. Thus, the four-level scheme is a decision-oriented operationalization rather than a claim that post popularity naturally consists of four objectively discrete groups.

To reduce dependence on the specific thresholds, sensitivity checks were conducted using alternative cutoffs, including the 5th, 85th, and 99th percentiles and a stricter 99.5th-percentile viral threshold. The results remained qualitatively consistent: macro-F1 changed by less than 0.02, and viral PR-AUC changed by less than 0.05. Detailed results are reported in S1 Appendix.

Importantly, the classification model complements rather than replaces the continuous prediction model. The continuous model retains the full variation in engagement outcomes, whereas the four-level classification provides more interpretable categories for content screening and managerial decision-making.

The classification model follows the same two-branch fusion structure as the continuous model. After the VGG19-based image representations and structured features are combined, the fused representation is processed through fully connected layers with ReLU activation, batch normalization, L2 regularization, and Dropout. The final softmax layer outputs the probability of each popularity class.

5.3.1 Continuous prediction model.

For the continuous prediction task, the dependent variable was transformed as , where total engagement was calculated as the sum of likes, comments, and reposts. The dataset was randomly divided into training and validation sets using an 80/20 split with a fixed random seed. Structured features were standardized using parameters estimated from the training set and then applied to the validation set. This procedure was used to reduce scale differences among variables and avoid information leakage.

All images were converted to RGB format and resized to 224 × 224 pixels before being input into VGG19. Image augmentation, including random rotation, width-height shifting, and horizontal flipping, was applied only to the training set. The model was trained using the Adam optimizer with a learning rate of 0.0001, mean squared error as the loss function, and mean absolute error as an additional metric. The batch size was set to 32, and early stopping monitored validation loss to restore the best model weights.

On the validation set, the continuous model achieved a log-scale MSE of 1.7421 and MAE of 1.0086. After reversing the log transformation, the MAE on the original engagement scale was 1175.80. These results suggest that the model captures general engagement patterns, but exact count prediction remains difficult because social media engagement is highly skewed and affected by external factors. Therefore, we further developed a categorical prediction model to classify posts into actionable popularity levels.

5.3.2 Classification prediction model.

Because the original distribution of the four popularity classes was highly imbalanced, we adopted a class-balanced sampling strategy for the classification task. At each fold, non-overlapping training, validation, and test subsets were constructed. Approximately equal numbers of posts were randomly sampled from each popularity class for model training, preventing the moderate-popularity category from dominating the learning process. A separate class-balanced test set was constructed from observations that did not overlap with the training data, with equal numbers of posts sampled from each class. The validation set consisted of previously unused observations and was used only for early stopping and model selection, rather than for model fitting. This sampling and evaluation procedure was repeated across 10 folds.

Feature standardization was performed using parameters estimated from the training data and then applied unchanged to the validation and test data. This procedure prevented information leakage between the training and evaluation stages.

The classification model was trained using the Adam optimizer and sparse categorical cross-entropy loss. The initial learning rate was set to 0.0005, the batch size was 32, and the maximum number of epochs was set to 100. Early stopping monitored validation loss with a patience of 10 epochs and restored the best weights. A ReduceLROnPlateau scheduler reduced the learning rate by a factor of 0.5 when validation loss did not improve for 5 epochs, with a minimum learning rate of 1e-6.

Fig 8 reports the training and validation curves for the classification model. The loss and accuracy curves stabilized during training, and the gap between training and validation performance remained limited, suggesting that Dropout, L2 regularization, batch normalization, early stopping, and learning-rate scheduling helped reduce overfitting.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Training and validation performance of the classification prediction model.

https://doi.org/10.1371/journal.pone.0356612.g008

Across the 10 folds, the proposed model achieved an average accuracy of 0.81, macro-precision of 0.79, macro-recall of 0.81, macro-F1-score of 0.80, and one-vs-rest multiclass AUC of 0.80. In practical terms, an average accuracy of 0.81 means that approximately 81% of posts in the class-balanced test folds were assigned to the correct popularity level. The model is therefore intended as a preliminary screening and prioritization tool rather than a deterministic decision rule. Because the four classes were equally represented in the test set, a naive classifier that always predicted a single category would achieve an accuracy of 0.25 rather than 0.80. Therefore, the observed accuracy cannot be attributed to the dominance of the moderate-popularity category. Table 4 further reports precision, recall, and F1-score for each popularity class.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Class-specific performance of the proposed classification model.

https://doi.org/10.1371/journal.pone.0356612.t004

As shown in Table 4, the model achieved meaningful precision and recall across all four popularity levels rather than relying primarily on the moderate category. The model performed best for moderate-popularity posts, followed by high-, low-, and viral-popularity posts. The relatively lower performance for viral posts indicates that exceptionally popular content remains more difficult to identify, possibly because the original pool of viral posts is relatively small and their engagement outcomes are more sensitive to external amplification factors.