A Hybrid RNN-LSTM Framework for Predicting Stakeholder Engagement on Social Media Platforms

Social media has become an essential tool for businesses to communicate with their customers, partners, employees, and the world at large in the digital revolution age. Through channels such as Facebook, Twitter, LinkedIn, and Instagram, companies are able to post information about themselves or their professional services while interacting with their audience in a two-way format. Active social media engagement between stakeholders can enhance brand exposure and foster trust, loyalty, and long-term relationships between organizations and their stakeholders [1, 2]. The huge quantity of content generated by users on these platforms offers organizations a unique opportunity to analyze patterns of interaction, expressions of sentiment, and levels of engagement that can generate data-driven communication strategies focused on different stages in the audience engagement funnel [3].

Predicting stakeholder engagement proves to be a difficult task even when using big social media data. The dynamic nature of engagement behaviour changes as a function of time, including temporal posting patterns, content attributes, sentiment polarity, and socio-demographic feature differences [4]. The management of user engagement and popularity has been substantially investigated with traditional ML methods; however, they are restricted to a static basis of features and fail to take into account the sequential patterns inherent within social media interactions [5]. Therefore, such methods might not capture how posting behavior changes in the past and short-term context impacts future engagement.

In recent years, deep learning models (e.g., Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) networks) have proven to be very effective in modeling sequential data such that temporal dependencies across time-ordered observations can be learned [6]. RNN architectures are good at capturing short-term dependencies, while the LSTM networks solve the problem of vanishing gradients, and long-term dependencies can also be modeled. These features would render sequential DL models particularly applicable for social media analytics as they engage with posts in an immediate context and over time.

In this work, we present an effective mixed RNN–LSTM model to forecast stakeholder engagement on social media channels. The developed method makes use of a thorough text processing, feature extraction based on the usage of term frequency inverse document frequency (TF-IDF), and modeling for temporal sequences to classify viral and non-viral posts. We experiment on a balanced multi-platform dataset of social media posts and show the strong performance of the predictions. This work does not highlight architectural novelty. Instead, we emphasise practical effectiveness, modelling dynamicity over time, and deployability in real-world social media analytics systems. The results provide evidence for the feasibility of lightweight sequential modeling as an aid for engagement prediction in organizational scenes and support the task of content optimization and recommendation.

The proposed system focuses on predicting social media users’ engagement by means of a hybrid RNN–LSTM architecture. The entire framework consists of five major steps: data gathering and organizing, data pre-processing and feature extraction, constructing temporal sequences, designing model architecture, and evaluating performance. This pipeline integrates the system within a structured flow to allow reproducibility, robustness, and adequacy for real-world environments.

3.1 Dataset acquisition

The study is based on a publicly available dataset Social Media Engagement Report available on Kaggle [23]. The dataset consists of 100k social media posts from various platforms, making for balanced posting categories, platforms, and time spans. Each post is accompanied by textual content, engagement scores (like count, comment count, and share count), sentiment labels, demographical attributes, and temporal metadata, i.e., time and day of the week when it was posted.

To mitigate class imbalance and potential bias, the dataset was created by ensuring a roughly equal balance of high- and low-engagement posts in the final dataset. Demographic audience categories that are commonly represented in the content include older adult users (45+ years) and also mature adults, as well as adolescents. The area where engagement tended to be higher was the geographical approach, in countries like Croatia and Malawi. Temporal patterns: Posts published during the evening and night hours of weekdays receive higher engagement. We argue that these properties make the dataset amenable to investigations into engagement dynamics, without as much biased learning behavior.

3.2 Data preprocessing and feature engineering

A comprehensive multi-step preprocessing pipeline was applied to ensure data quality and compatibility with sequential deep learning models.

3.2.1 Cleaning and filtering

All text inputs were converted to lowercase and cleaned from URLs, user mentions, hashtags, punctuation signs and numerical digits via regular-expression-based filtering. Intending to avoid repetition and biased learning, duplicate records were eliminated. Due to the data sparsity of sentiment values for missing sentiment, the default neutral (Mixed Sentiment) category was created to make the dataset balanced in order to avoid bias.

3.2.2 Tokenization and stopword removal

Tokenization was performed using the Natural Language Toolkit (NLTK), adding each post as a sequence of separate tokens. Stopwords such as “is,” “the,” and “in” were eliminated to keep semantically meaningful words for engagement prediction. The step filters noise and enhances the discriminative capability of resultant features.

3.2.3 Vectorization

Textual data were transformed into numerical representations using TF-IDF vectorization. The vocabulary size was restricted to the top 1,000 most informative terms, balancing representational richness with computational efficiency. TF-IDF was selected over contextual embeddings due to its interpretability, lower computational overhead, and suitability for real-time deployment scenarios.

3.2.4 Feature scaling

Numerical engagement metrics, including likes, comments, and shares, were normalized to a common scale to ensure uniform feature contribution during model training. This prevents dominance of features with larger numeric ranges and stabilizes gradient updates.

3.3 Proposed architecture

Stakeholder engagement is highly time-dependent, mixed with the relevancy of posts, the posts were first ordered chronologically per-account or campaign. Fixed-length sliding windows were then made as temporal sequences, where each sequence represents consecutive posts to the target engagement outcome. Crucially, dataset splitting is done at the sequence level and not the post level in order to prevent information leakage across training, validation, and testing sets.

short-term (contextual) patterns and long-term dependencies in the engagement behavior. After preprocessing the data split into training, validation and testing sets, maintaining class balance (70% training, 10%validation and 20% test). The input layer receives concatenated TF-IDF textual features and normalized metadata features. To model the near immediate sequential changes between consecutive posts, we use a lightweight RNN layer as a temporal smoothing component. These are then sequentially followed by two stacked LSTM layers, with 128 and 64 units in them respectively, to capture longer-term dependencies and evolving engagement trends. The dropout is used as regularization to avoid overfitting. Then a dense layer with ReLU as activation fuses the temporal information, followed by a sigmoid-based output layer for binary classification between high- and low-engagement.

Model evaluation is conducted using confusion matrices and standard performance metrics, and the outputs support downstream tasks such as engagement prediction and content recommendation.

Pseudocode of the proposed methodology:

Input: Social media posts with textual content, timestamps, hashtags, platform identifiers and audience segments; engagement metrics likes, comments, shares; binary engagement labels; preprocessing resources; and predefined model hyperparameters.

Output: Predicted engagement labels with probability scores, performance metrics, learning curves, and engagement optimization insights.

Process: The data set is loaded and validated and then text cleaning, tokenization, stop word removal and TF-IDF are performed for feature extraction. Features are scaled and posts are sorted in chronological order to create temporal sequences. The dataset is divided into training, validation and test splits at the sequence level. As shown in Figure 1, we develop a combined RNN–LSTM model, which we train with the Adam algorithm and early stopping before reporting classification metrics and error analysis. Lastly, the trained model and preprocessing pipeline are saved for testing in deployment scenarios or future inference.

Figure 1. Architecture of the proposed RNN–LSTM engagement prediction framework

3.4 Model training

The framework was implemented using Python 3.10 using TensorFlow, Keras, Pandas and Matplotlib environment on a Windows 11 machine equipped with an Intel i7 (3.10 GHz) processor, NVIDIA GeForce RTX 3050 GPU and 64 GB RAM. Training was performed with Adam optimizer (with learning rate of 0.0001) and binary cross-entropy loss. Overfitting was managed using dropout regularization and early stopping. The training parameters are summarized in Table 2.

Table 2. Training parameters for implementation of proposed model

3.5 Performance evaluation

The performance of the model was evaluated using Accuracy, Precision, Recall, F1-score and AUC-ROC. True positive, false positive, true negative and false negative were quantified using a confusion matrix. It is undeniable that the overall model discriminating across varying decision thresholds as a whole was evaluated by the ROC curve to form an integrated view on predictive robustness.

A novel hybrid RNN–LSTM framework for anticipating stakeholder engagement on social network systems with rich content (text), interaction features (engagement metrics) and associated metadata. Building on a balanced dataset of 100 k which includes posts from five sources an aspect of this approach combines traditional preprocessing, including noise removal, tokenization and feature extraction using TF-IDF based representation, also for the temporal sequence modeling in order to learn short-term contextual effects and longer term engagement patterns. Experimental evaluation demonstrated strong predictive performance with 99.60% accuracy, 99.61% precision, 99.60% recall, an F1-score of 0.9961 and an AUC-ROC of 1 under a binary classification scenario. Critically, these results were achieved with a computationally light architecture, showing the merit of a lightweight temporal modeling for engagement prediction without the need for resource-demanding multimodal or graph-based frameworks.

In general, the findings indicate that it is practical to incorporate the proposed model as a useful feature for real-time engagement prediction in social media analytics applications and thus assist organizations in content strategy and stakeholder communication optimisation. Precision and computational efficiency are traded off while still allowing for deployment in real-time decision environments, crucial for both timely decision-making and scalability. There are a few limitations concerning the above findings. The proposed model is text based and uses numeric engagement indicators while excluding the visual, audiovisual content which plays a critical role in most of the social networking action models. Moreover, TF-IDF representation is both efficient and interpretable without the richer deeper contextual semantics contained in recent pre-trained language models. Furthermore, the analysis was performed on the static history data, and its generalization with dynamic platform changes and user behavior trends is an open issue.

In the future, we will investigate a number of such extensions for improving the robustness and applicability of our method across varying experimental conditions. Second, combining multimodal inputs like images and videos is believed to enhance prediction power, especially for visually-based platforms. Second, transformer-based structures and attention mechanisms can be introduced to model more complex contextual dependencies while preserving computation efficiency. Third, the development of adaptive or continual learning pipelines will enable the model to perform parameter updates with the arrival of new data, facilitating sustained performance in an evolving social media context. Lastly, expanding the model to multi-level engagement classification and cross-lingual analysis will increase its generalizability for global organizational contexts.