Predictive Modeling of Engagement on Instagram: Machine Learning Applied to the Gaming Sector

Beatriz Shida Diliberti; Adriana Camargo de Brito

DOI: 10.22167/2675-6528-202602719

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

Social media are strategic channels for brands, especially in the electronic games industry, where Instagram engagement is a crucial metric. The predictive capacity of structural, temporal, textual, and visual features on the engagement of electronic game posts on Instagram was investigated, using supervised machine learning techniques in a progressive and multimodal approach. The research used administrative data from 2,125 posts from two official game profiles, focused on the Brazilian and Latin American audiences. The engagement rate was operationalized as a binary classification based on the median. The algorithms Logistic Regression, Random Forest, and XGBoost were evaluated in three progressive classifications, incorporating structural data, perceptual visual attributes via OpenCV, and semantic representations from the CLIP ViT-B/32 model. In the classification with structural data, Random Forest achieved an AUC-ROC of 0.84 and an F1-score of 0.78, indicating limited predictive capacity. With the inclusion of perceptual visual and semantic attributes, the AUC-ROC rose to 0.88 and the F1-score to 0.81. The results demonstrated that visual content plays a central role in predicting engagement and that the integration of multiple modalities is fundamental to more accurately modeling audience response in electronic game contexts.

Keywords: Multimodal classification; Instagram; machine learning; electronic games marketing; engagement prediction.

1. Introduction

Social media have transformed the way brands and consumers interact, becoming strategic channels for communication and digital marketing. In Latin America, Instagram is one of the most relevant platforms. According to the Digital 2024: Brazil (2024) report, Brazil has over 110 million active users, positioning itself as the network’s second-largest market in the world. In the regional scenario, Instagram acts as a central content discovery channel, directly influencing consumer behavior and brand loyalty. This relevance is even greater in the electronic games sector, where publishers and developers use the platform to promote trailers, updates, and promotional campaigns.

In this competitive environment, the evaluation of publication performance becomes fundamental, and engagement assumes a central role in digital performance analysis. Engagement on social media can be understood as the audience’s behavioral response to a content stimulus, manifested through observable actions such as likes, comments, shares, and saves. Although the concept has cognitive and emotional dimensions, literature reviews indicate that, in the empirical context of digital platforms, engagement is predominantly operationalized through these measurable interactions (Trunfio and Rossi, 2021). Furthermore, the phenomenon is multifactorial, being influenced by factors such as the exposure obtained by the publication, visual and textual characteristics of the content, posting time, and algorithmic distribution mechanisms. Studies on content diffusion demonstrate that digital popularity metrics tend to exhibit a highly asymmetric distribution and cumulative temporal dynamics, in which initial interactions can influence the future growth trajectory of the content (Cheng et al., 2014). Thus, engagement should not be interpreted as an isolated metric, but as a result of the interaction between multiple determinants that vary over time and the platform’s competitive context. The basis for performance calculation has also evolved, with the effectiveness of a campaign being measured more reliably by reach, which accounts for the number of unique users impacted (Kwai Mui and Tong Ming, 2020), and not just by the total number of followers. This normalization is crucial for comparing performance across different profiles. Engagement, however, has undergone a redefinition over the years, with more recent studies suggesting that actions of greater depth, such as sharing and saving publications, are fundamental indicators of consumer connection with the brand, going beyond traditional metrics focused solely on likes and comments (Alvarez-Milán et al., 2018; Kotler and Keller, 2013).

To deal with this complexity, machine learning techniques have been applied to the prediction of engagement on social networks, allowing the identification of patterns in large volumes of data and the capture of non-linear relationships between variables. Studies show that the integration of textual and contextual data is essential for identifying popularity patterns (Moro et al., 2016; Xu et al., 2025). More recently, multimodal approaches that combine textual and visual information have demonstrated significant gains in predictive performance compared to models based exclusively on structural metadata (Philp et al., 2022; Meghawat et al., 2018). As an aesthetics-oriented network, Instagram motivates the use of multimodal approaches (Baltrušaitis et al., 2017), integrating text and image. Advanced visual representation models, such as Contrastive Language-Image Pretraining (CLIP), developed by OpenAI (Radford et al., 2021), allow capturing aspects of composition and style that impact engagement, especially in the video game sector, where the visual quality of content is a central element of communication with the audience (Kruk et al., 2019). Despite the growing academic interest in predicting engagement on social networks, studies that adopt a progressive and multimodal approach, evaluating the incremental gain of each data modality, are still scarce in the literature, especially in the video game sector in the Latin American context (Philp et al., 2022; Moro et al., 2016).

Given this context, the scarcity of studies exploring the predictive capacity of multiple data modalities in a progressive approach within the electronic games sector and the Latin American scenario justifies the relevance of this research. The objective of this study is to investigate the predictive capacity of structural, temporal, textual, and visual characteristics on the engagement of electronic games posts on Instagram, through supervised machine learning techniques in a progressive and multimodal approach. The scope of the analysis focuses on content related to battle royale genre games for PC and Console platforms, with a focus on the Brazilian and Latin American audience.

2. Material and Methods

This research was characterized as a study of an applied nature, with a quantitative approach and descriptive and explanatory objectives, aiming to analyze the performance of electronic game posts on Instagram and evaluate the predictive potential of structural, temporal, textual, and visual variables on audience engagement (Gil, 2008). The study was carried out in Python language (version 3.x), in the Google Colab environment, using libraries such as Pandas (McKinney, 2010) and NumPy (Harris et al., 2020) for data manipulation, and Scikit-learn (Pedregosa et al., 2011) for implementing models and metrics.

The study was structured into two complementary predictive modeling approaches: continuous engagement rate prediction and binary engagement classification, using multimodal data derived from the posts. Initially, an exploratory continuous regression step was conducted to directly estimate the engagement rate from structural, temporal, and textual variables. Subsequently, the problem was reformulated as a binary classification task, distinguishing high and low engagement posts based on the median of the training set, due to the asymmetric distribution of the variable in digital environments (Cheng et al., 2014; Moro et al., 2016).

The classification stage was developed in a progressive and multimodal approach, incorporating different types of data incrementally. First, structural, temporal, and textual variables were used. Next, perceptual visual attributes extracted via OpenCV were incorporated. Finally, semantic representations obtained through the CLIP model were included. This strategy allowed for the evaluation of the incremental gain of each data modality in the predictive capacity of the models (Philp et al., 2022; Meghawat et al., 2018; Baltrušaitis et al., 2017).

The data were collected from two official “battle royale” electronic game profiles, belonging to the same publisher, with one profile aimed at the Latin American public and the other at the Brazilian public. The collection was based on primary administrative data extracted from the Instagram platform, through reports available in the “Meta Business Suite”. To ensure correspondence between metadata and media files, each post was identified by a unique code (post_id).

The data extraction was conducted on two complementary fronts. The post metadata, containing structural, temporal, and textual information, was exported in JSON format. The post performance reports, including reach and interaction metrics, were obtained in CSV format. The JSON files were processed using Python’s zipfile and json libraries, with consolidation via Pandas (McKinney, 2010), while the CSVs were loaded directly.

To integrate the two data sources, the “matching” procedure was applied, a correspondence technique between records from different databases (Christen, 2012; Rahm and Do, 2000). The correspondence was established through the publication date (day_key) and a textual identifier derived from the first characters of the caption of each post (caption_key). The captions underwent character encoding correction to ensure text integrity (Rahm and Do, 2000).

After filtering inconsistencies and removing duplicates, an analytical base composed of 2,125 posts was obtained for classification with structural data. An initial filtering process removed 795 posts without reach data. Subsequently, 1,655 posts were removed in the “matching” and inconsistency filtering process, resulting in the final base of 2,125 posts.

From the raw data collected, the variables used in the study were extracted and derived. From the publication date, the time (hour) and the day of the week (day_of_week) were obtained. From the caption, the text size (caption_size), the number of hashtags (num_hashtags), and the presence of a CTA (cta_presence) were calculated. From the reach and interactions, the engagement rate (eng_rate) was calculated, defined as the ratio between the total interactions (likes and comments) and the publication’s reach. From the number of slides, the num_slides variable was obtained. The variable referring to the post format (image, carousel, or video) was collected but excluded from the predictive models due to inconsistencies in temporal distribution (Bergmeir and Benítez, 2012).

From the static images, available in 1,416 out of 2,125 posts, perceptual visual attributes were extracted via OpenCV (Bradski, 2000) and semantic representations via CLIP (Radford et al., 2021). Static images were selected for perceptual visual and semantic classifications, excluding videos and posts without an associated image file. This filter resulted in the reduction of the dataset to 1,416 posts with valid static images.

Perceptual visual attributes extracted via OpenCV included brightness (average grayscale pixel intensity), contrast (image tone variation, pixel standard deviation), entropy (visual complexity via Shannon Entropy, 1948), edges (edge density via Canny detector, 1986), and sharpness (sharpness level via Laplacian operator variance, Marr and Hildreth, 1980). These metrics were calculated on the grayscale version of each image.

Semantic representations were extracted by the CLIP ViT-B/32 model (Radford et al., 2021). Each image was resized to 224×224 pixels and segmented into 32×32 pixel “patches”, processed by the model’s visual “encoder” to generate a 512-dimensional “embedding” vector (Dosovitskiy et al., 2021). These vectors were reduced to 10 components using Principal Component Analysis [PCA] (Jolliffe, 2002), resulting in the variables `clip_pca_0` to `clip_pca_9`.

For continuous regression predictive modeling, the Ridge, Random Forest Regressor, and XGBoost Regressor models were evaluated. Performance was assessed by the Mean Absolute Error [MAE], Root Mean Squared Error [RMSE], and Coefficient of Determination [R²] metrics (Chicco et al., 2021). Validation was temporal, with 80% of the data for training and 20% for testing (Bergmeir and Benítez, 2012).

In binary classification, posts were classified as high engagement (value 1) if the engagement rate surpassed the median of the training set, and low engagement (value 0) in other cases. Three supervised classification algorithms were evaluated: Logistic Regression, Random Forest Classifier, and XGBoost Classifier (Breiman, 2001; Chen and Guestrin, 2016; Fávero and Belfiore, 2024).

To mitigate the risk of imbalance between high and low engagement groups, in Logistic Regression and Random Forest, the parameter “class_weight=balanced” was used (Pedregosa et al., 2011). In XGBoost, the equivalent adjustment was performed by the parameter “scale_pos_weight” (Chen and Guestrin, 2016). The input variables of Logistic Regression were standardized by “z-score” before training (James et al., 2013).

The strategy of splitting between training and testing varied according to the classification. For classification with structural data (2,125 posts), temporal validation was adopted (80% training, 20% testing). In the visual perceptual and semantic classifications (1,416 posts with static images), stratified “split” was used in an 80/20 ratio, ensuring that posts with high and low engagement were represented in the same proportion (James et al., 2013; Pedregosa et al., 2011).

To measure the performance of binary classification models, metrics such as Area under the ROC curve [AUC-ROC], F1-score, precision, recall, and confusion matrix were adopted (Chicco et al., 2021; Willmott and Matsuura, 2005). AUC-ROC was the primary metric, due to its robustness to group imbalance and its ability to measure the model’s discriminative capacity independently of the decision threshold (Fávero and Belfiore, 2024).

To verify that the obtained results reflect genuine associations between the analyzed variables and post engagement, a permutation test was conducted on the subset of visual perceptual and semantic classifications. The procedure consisted of randomly shuffling the high and low engagement labels of the posts and repeating the training of the Random Forest model in 20 independent iterations (Ojala and Garriga, 2010; Molnar, 2022).

3. Results and Discussion

The research results are presented in a sequence of four steps, following the progressive and multimodal approach outlined in the methodology. Initially, continuous regression was explored, followed by binary classification of engagement with structural data. Subsequently, perceptual visual attributes were incorporated, and finally, semantic representations. This structure allows for a detailed analysis of the incremental gain in the predictive capability of the models as each new data modality is integrated, offering an in-depth understanding of the factors influencing engagement in electronic gaming posts on Instagram.

Engagement rate analysis

Initially, the engagement rate, the study’s dependent variable, was presented. Based on 2,125 analyzed posts, the average engagement was 2.71%, while the median registered 1.18%. This significant difference between the values indicates that a few posts concentrate the majority of interactions, while most exhibit lower engagement levels. Such a pattern is consistent with the literature on social networks, which describes an unequal distribution of interactions among content, a phenomenon known as attention inequality (Bakhshi et al., 2014; Lerman and Hogg, 2010; Zhu and Lerman, 2016). The engagement rate distribution revealed a positive skew, with the median and mean far apart, confirming that engagement is not uniformly distributed among publications.

Continuous regression

In the exploratory stage of continuous regression, the evaluated models — Ridge, Random Forest Regressor, and XGBoost Regressor — demonstrated low predictive performance. In all cases, the coefficient of determination (R²) was negative, indicating that the models failed to predict engagement more effectively than an approach based simply on the training set’s average. For instance, the Ridge model achieved a Mean Absolute Error (MAE) of 0.0221 and a Root Mean Squared Error (RMSE) of 0.0284, while the Random Forest Regressor and XGBoost Regressor presented MAE of 0.0249 and 0.0246, and RMSE of 0.0409 and 0.0499, respectively. These results suggest that the isolated structural variables were insufficient to explain engagement in the test set.

This limited performance does not represent a methodological flaw, but rather a characteristic of phenomena with high variability and noise, as observed in previous studies (Cheng et al., 2014; Moro et al., 2016). Social media engagement is influenced by contextual and latent factors that are not captured by simple structural variables, such as publication time or caption length. Given this limitation, the problem was reformulated into a binary classification task, distinguishing high and low engagement posts based on the median of the training set, a more suitable approach for predicting popularity on social media (Philp et al., 2022). It was observed that even non-linear models, such as Random Forest and XGBoost, did not outperform the linear Ridge model in this initial phase, reinforcing the need to incorporate new data dimensions.

Classification with structural data

With the reformulation of the problem for binary classification, the models were re-evaluated using exclusively structural, temporal, and textual data. The Random Forest algorithm achieved the best overall performance, registering an Area Under the ROC Curve (AUC-ROC) of 0.84 and an F1-score of 0.78. XGBoost presented very close results, with AUC-ROC of 0.84 and F1-score of 0.77. Logistic Regression, although with slightly lower performance, achieved an AUC-ROC of 0.80 and F1-score of 0.76, indicating that even a simpler model can identify patterns between structural variables and engagement.

The analysis of the Random Forest model’s confusion matrix, considering a test set of 425 posts, revealed that 174 were correctly identified as high engagement and 157 as low engagement. 94 incorrect classifications were observed, with 56 false positives (low engagement posts classified as high) and 38 false negatives (high engagement posts classified as low). This distribution suggests a tendency for the model to overestimate high engagement posts. Although moderate, this performance represents a significant advancement compared to continuous regression, where models failed to outperform a prediction based on the average.

The AUC-ROC of 0.84, obtained with exclusively structural, temporal, and textual variables, demonstrates that operational metadata of posts, such as publication time, caption length, and presence of Call to Action (CTA), possess real, but limited, predictive capacity. This finding is in line with previous studies indicating the restricted explanatory power of isolated structural metadata in predicting engagement, reinforcing the need to integrate additional content dimensions to enhance the predictive performance of models (Moro et al., 2016; Philp et al., 2022).

Classification with perceptual visual data

The incorporation of perceptual visual variables, extracted via OpenCV, to the structural, temporal, and textual variables resulted in an expressive gain in model performance. Random Forest remained the best-performing algorithm, achieving an AUC-ROC of 0.88 and an F1-score of 0.81. This result represents an increase of 0.04 points in AUC-ROC compared to the previous classification with structural data. XGBoost also showed improvement, reaching an AUC-ROC of 0.87 and an F1-score of 0.80. Logistic Regression, in turn, achieved an AUC-ROC of 0.83 and an F1-score of 0.76.

Low-level visual features, such as brightness, contrast, entropy, and edge density, extracted directly from image pixels without semantic interpretation, contributed significantly to engagement prediction. The Random Forest confusion matrix for this phase indicated that the model correctly classified 114 low-engagement posts and 116 high-engagement posts. 28 false positives and 26 false negatives were recorded, showing a reduction in the volume of errors compared to classification based solely on structural data. This informational gain corroborates studies that associate luminosity and visual complexity properties with the performance of social media publications (Kruk et al., 2019; Meghawat et al., 2018).

Classification with semantic data

In semantic classification, perceptual visual variables were replaced by semantic representations, obtained by the CLIP ViT-B/32 model and reduced to ten components via Principal Component Analysis (PCA). The results demonstrated equivalent performance to perceptual visual classification for the Random Forest and XGBoost algorithms. Both achieved an AUC-ROC of 0.88 and an F1-score of 0.81, confirming the robustness of including visual information. Logistic Regression, however, presented the greatest incremental gain between stages, achieving an AUC-ROC of 0.86 and an F1-score of 0.79, suggesting that high-dimensional semantic representations capture more linearly separable patterns than simple perceptual attributes.

The confusion matrix of the Random Forest for semantic classification showed 119 low-engagement posts and 112 high-engagement posts classified correctly. 23 false positives and 30 false negatives were identified. A slight inversion in the error pattern was observed compared to visual perceptual classification, with a slightly higher volume of false negatives. This indicates that CLIP’s semantic representations more consistently capture patterns associated with low engagement, although they may not fully capture subjective elements that determine high engagement in electronic gaming contexts (Radford et al., 2021; Ohme et al., 2024).

The results obtained confirm that the incorporation of visual information, whether perceptual or semantic, produces a consistent gain in predictive performance compared to the exclusive use of structural metadata. The Random Forest algorithm stood out as the most robust across the three classifications, maintaining an AUC-ROC of 0.84 with structural data and increasing it to 0.88 in the perceptual and semantic visual classifications. XGBoost also demonstrated competitive performance, especially in semantic classification, while Logistic Regression, although initially inferior, significantly leveraged CLIP’s semantic representations, evidencing its adaptability. These findings are consistent with the literature on social media engagement prediction, which points to significant gains when incorporating visual information into predictive models (Philp et al., 2022; Baltrušaitis et al., 2017).

To ensure the validity of the findings, a permutation test was conducted on the subset of visual perceptual and semantic classifications. The Random Forest model was trained 20 times with randomly shuffled engagement labels, resulting in an average AUC-ROC of 0.50, a value close to chance. This result strongly contrasts with the AUC-ROC of 0.88 obtained by models trained with the actual labels, confirming that the observed performance is not a product of overfitting or spurious patterns in the data. The expressive difference between the scenarios validates that the identified associations between the predictor variables and engagement are genuine in the analyzed context (Ojala and Garriga, 2010).

In summary, the present study analyzed the predictive capacity of structural, temporal, textual, and visual characteristics on the engagement of electronic games posts on Instagram, using a progressive and multimodal approach with supervised machine learning techniques. The results demonstrated that the incorporation of visual information into the set of predictor variables produced consistent performance gains, directly addressing the proposed objective. Isolated structural metadata showed real, but limited, predictive capacity, with an AUC-ROC of 0.84 in classification with structural data. The inclusion of perceptual visual attributes via OpenCV raised this value to 0.88, evidencing the relevance of properties such as brightness, contrast, complexity, and image sharpness for engagement on visual platforms like Instagram.

The semantic representations extracted by the CLIP ViT-B/32 model maintained performance for the Random Forest and XGBoost models, while providing the largest incremental gain for Logistic Regression, suggesting that semantic embeddings capture patterns complementary to perceptual attributes. Permutation testing confirmed that the results reflect real associations between the predictor variables and engagement, and are not artifacts of the modeling process. However, it is important to consider the study’s limitations, such as the analysis being restricted to posts from only two official profiles of a single genre of electronic games, which may limit the generalization of the findings. Furthermore, the constant changes in Instagram’s content distribution algorithms may alter the relationship between post characteristics and observed engagement over time.

4. Conclusion

This study investigated the predictive capacity of structural, temporal, textual, and visual features on the engagement of electronic gaming posts on Instagram, employing supervised machine learning techniques in a progressive and multimodal approach. It was found that the initial modeling, based exclusively on structural, temporal, and textual data, presented a real, but limited, predictive capacity, with an AUC-ROC of 0.84 and an F1-score of 0.78 for the Random Forest algorithm. However, the incorporation of perceptual visual attributes, extracted via OpenCV, consistently improved model performance, achieving an AUC-ROC of 0.88 and an F1-score of 0.81. It was observed that the semantic representations, obtained by the CLIP ViT-B/32 model, maintained this performance level for the Random Forest and XGBoost algorithms, and provided the greatest incremental gain for Logistic Regression, suggesting that semantic embeddings capture patterns complementary to perceptual attributes. A permutation test confirmed that the results reflect genuine associations between the predictor variables and engagement, and are not artifacts of the modeling process.

The main contribution of this work lies in demonstrating that visual content plays a central role in predicting engagement in electronic games posts on Instagram, and that the integration of multiple data modalities is fundamental to more accurately modeling audience response. This finding offers valuable insights for digital marketing strategies in the gaming sector, allowing for content creation optimization to maximize interaction. However, the study has limitations, as the analysis was restricted to posts from only two official profiles of a single genre of electronic games, which may limit the generalization of findings to other contexts and genres. Additionally, the constant changes in Instagram’s content distribution algorithms may alter the relationship between post characteristics and observed engagement over time. For future studies, it is suggested to expand the analysis to other electronic games profiles and genres, incorporate sentiment analysis of comments, and apply model interpretability techniques, such as SHAP and LIME, to deepen the understanding of the identified predictive patterns.

Bibliographic References

Alvarez-Milán, A.; Felix, R.; Rauschnabel, P. A.; e Hinsch, C. 2018. Strategic customer engagement marketing: A decision making framework. Journal of Business Research, 92, 61-70.

Cheng, J., Adamic, L. A., Dow, P. A., Kleinberg, J., e Leskovec, J. (2014). Can cascades be predicted? WWW 2014 – Proceedings of the 23rd International Conference on World Wide Web, 925-935.

Digital 2024: Brazil. 2024. Digital 2024: Brazil. Disponível em: https://datareportal.com/reports/digital-2024-brazil. Acesso em: 19 set 2025.

Kotler, P., e Keller, K. L. 2013

Kwai Mui, C.; Tong Ming, L. 2023. A critical review on impression rate and pattern on social media sites. Journal of Digital Marketing Research, 5(1), 1–15.

Trunfio, M., e Rossi, S. 2021. Conceptualising and measuring social media engagement: A systematic literature review. In Italian Journal of Marketing (Vol. 2021, Number 3, pp. 267–292). Springer International Publishing.

Article originating from the Final Course Project of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform