Post Engagement Prediction (PEP)

• Post Engagement Prediction (PEP) is a framework that forecasts future user interactions with social media posts using early signals and diverse features.
• It employs methodologies such as classification, regression, ranking, and trajectory forecasting to address varied engagement objectives.
• PEP research leverages multimodal signals—including content, behavioral, and relational data—to enhance prediction accuracy while addressing issues like class imbalance and temporal leakage.

Post Engagement Prediction (PEP) denotes the modeling of future interaction with a post, or with a user–post candidate exposure, from signals available before or shortly after publication. In published work, the target has been instantiated as binary high/low Facebook post impact over comments, shares, and grouped reactions (Rosas-Quezada et al., 2019), zero-versus-nonzero tweet engagement defined by likes plus retweets (Arazzi et al., 2023), per-user ranking of tweets by retweets and favorites (Zamani et al., 2015), multivariate forecasting of likes, shares, comments, and emojis from irregular early histories (Tian et al., 7 Feb 2025), and user-level future engagement continuity from behavioral sequences (Semenov et al., 2022). The literature therefore treats PEP not as a single task but as a family of formulations spanning classification, regression, ranking, trajectory forecasting, and propensity modeling. A separate usage of the same acronym appears in social-media language-model pretraining, where “Post Engagement Prediction” refers to predicting structural relations in propagation trees rather than directly forecasting likes or comments (Cui et al., 10 Aug 2025).

1. Conceptual scope

The broadest framing treats engagement as future participation under an observation window and a prediction window, with no universal algorithm or feature set working reasonably well across all social media types (Sadeque et al., 2019). Within that broader participation view, post-level PEP focuses on interactions with a single content item or a candidate exposure event. Depending on the platform, engagement may mean likes, comments, replies, shares, quotes, favorites, reactions, or composite event vectors.

A key divide in the literature is between behavioral engagement and latent or subjective engagement. Most social-media PEP papers use observed platform actions as labels. Adjacent work on interactive media instead predicts self-reported engagement from interaction logs or multimodal behavior, using instruments such as the User Engagement Scale or post-session questionnaires (Carlton et al., 2021, Ma et al., 2024). This suggests that PEP can target either externally visible interaction or an inferred internal state, but the two should not be conflated.

Another conceptual distinction is the unit of prediction. Some work predicts outcomes for individual posts; some predicts whether a specific user will engage with a specific tweet; some predicts whether a user will remain active in the near future; and some aggregates posts into opinions, narratives, or topics and predicts collective trajectories (Jeromela, 2023, Semenov et al., 2022, Tian et al., 7 Feb 2025). Topic-level work also shows that engagement can be modeled as a normalized latent quantity rather than a raw count. In the cryptocurrency study, the engagement coefficient $\alpha_c$ scales expected interactions by follower count and interaction type, so that very low and very high normalized engagement carry different interpretations (Qureshi et al., 2022).

2. Problem formulations

Published PEP work spans several recurrent formulations.

Formulation	Target definition	Representative work
Binary post classification	High/low impact, any engagement, or thresholded future activity	(Rosas-Quezada et al., 2019, Arazzi et al., 2023, Ozegovic et al., 11 Feb 2025)
Regression / forecasting	Future counts or trajectories over fixed horizons	(Pham et al., 2016, Tian et al., 7 Feb 2025)
Learning to rank	Order posts by expected engagement within a user	(Zamani et al., 2015)
User-level continuation	More than a threshold number of future actions	(Semenov et al., 2022)
Topic-level normalized engagement	Estimate latent engagement coefficient from interactions	(Qureshi et al., 2022)

The Facebook brand-post study is a clear example of binary reformulation. It defines six classification tasks over $|C|$, $|S|$, $|R|$, $|R+|$, $|R-|$, and $|R\odot|$, with labels obtained by comparing each post’s value to the global average for that metric, so that

$$F(x)=y,\qquad Y=\{\text{high-impact},\text{low-impact}\}$$

and

$$y_{i,k}= \begin{cases} \text{high-impact}, & \text{if } p_{i,k}>\bar{x}_k\ \text{low-impact}, & \text{otherwise}. \end{cases}$$

This is explicitly classification-only rather than regression (Rosas-Quezada et al., 2019).

The Twitter graph paper uses a different binary target: $$eng(t_i)=Favorite_{t_i}+Retweet_{t_i}, \qquad l_{t_i}= \begin{cases} 0,& eng(t_i)=0\ 1,& eng(t_i)\ge 1. \end{cases}$$ Its prediction unit is the tweet node in a graph of semantically related tweets (Arazzi et al., 2023).

Ranking-oriented PEP appears in the MovieTweeting study, where the goal is not merely to predict engagement counts but to rank each user’s rating tweets by retweets plus favorites. That work argues that when the downstream problem is ordering posts, learning-to-rank can be better aligned than count regression, and it evaluates ranking quality with mean NDCG@10 over users (Zamani et al., 2015).

Trajectory forecasting appears most clearly in IC-Mamba, which takes a post tuple $|C|$0 with interval-censored history

$|C|$1

and predicts future multivariate engagement vectors

$|C|$2

This is not a scalar popularity score but a multi-horizon, multi-metric forecasting problem (Tian et al., 7 Feb 2025).

3. Signals and representations

Feature design is a central axis of PEP. The Facebook study organizes post signals into four families: content-based, style-based, behavioral, and metadata-based features. Content is a bag-of-words over the 10,000 most frequent tokens; style includes post length, uppercase, lowercase, numerals, and symbols; behavioral features count emojis, hashtags, mentions, and links; metadata includes media-link types and unusual time features based on percentages of posts at the same hour, day, month, and year (Rosas-Quezada et al., 2019). Its central empirical claim is that what companies write is the strongest signal, while how they write provides complementary gains.

Context-only Twitter modeling pushes feature engineering in a different direction. Without tweet text, the strongest features are users’ prior engagement history and the popularity of hashtags and links in the tweet, along with viewer–author interaction history, current trends over windows from 0.5h to 48h, and language and tweet-type indicators (Jeromela, 2023). This places historical behavior and interaction affinity at the center of context-based PEP.

Several papers enlarge the signal space beyond conventional metadata. TweetGage introduces inter-post relations by building a graph

$|C|$3

where tweets are connected if they share at least one hashtag within $|C|$4 minutes, with edge weight

$|C|$5

Its node features combine user statistics, tweet metadata, and a pre-trained BERT text embedding (Arazzi et al., 2023). IC-Mamba instead emphasizes irregular early engagement histories and interval structure. It encodes temporal context with relative and absolute time embeddings and modulates them by engagement magnitude through

$|C|$6

so that elapsed time and current count magnitude are jointly represented (Tian et al., 7 Feb 2025).

Domain-specific semantic features can also matter. In the Reddit food study, estimated calorie, protein, carbohydrate, and fat densities per 100g are derived from titles via nearest-neighbor matching to USDA food items, and nutritional density improves the control-only baseline, especially for predicting top-1% “resonant” posts (Ozegovic et al., 11 Feb 2025). The cryptocurrency paper similarly shows that engagement can be normalized by audience opportunity: expected interaction counts satisfy

$|C|$7

and the resulting $|C|$8 is a follower-normalized engagement coefficient rather than a raw count (Qureshi et al., 2022).

Interaction logs create yet another feature regime. In interactive media, session length, time to completion, and inter-event pause bins of $|C|$9 s, $|S|$0 s, $|S|$1 s, and $|S|$2 s were more informative than many raw event identities (Carlton et al., 2021). This suggests that temporal structure, pacing, and skip-like behavior are first-class PEP features when fine-grained consumption traces are available.

4. Modeling paradigms

The methodological spectrum runs from classical supervised learning to graph models, state space models, and prompt-based multimodal reasoning. Classical PEP work includes Naïve Bayes, Decision Trees, SVM, and k-NN for Facebook brand-post classification; Linear Regression and Regression Tree after PCA or sparse autoencoder preprocessing for blog-comment prediction; and Logistic Regression, Random Forests, Gradient Boosting, Naive Bayes, SVC, and MLP for context-only tweet engagement (Rosas-Quezada et al., 2019, Pham et al., 2016, Jeromela, 2023). These systems typically rely on engineered tabular features and remain important baselines.

Ranking-based methods form a distinct paradigm. The MovieTweeting work compares regression against ListNet, RankingSVM, AdaRank, RankNet, LambdaRank, and ListMLE, then aggregates rankings via supervised Kemeny aggregation: $|S|$3 Its main claim is that direct ranking objectives can outperform regression when the application is post ordering (Zamani et al., 2015).

Graph-based PEP appears in several forms. TweetGage uses GraphSAGE-style aggregation over a tweet graph, with neighborhood message

$|S|$4

and update

$|S|$5

The misinformation-engagement paper instead predicts a binary user propensity on user–user graphs and introduces EgoCL, which replays ego-graphs rather than isolated nodes in continual learning (Arazzi et al., 2023, Bo et al., 2022).

Sequence models and continuous-time forecasting dominate early-trajectory work. IC-Mamba modifies selective state space models so that the hidden-state transition depends explicitly on elapsed time,

$|S|$6

which is the core censoring-aware adaptation for irregular engagement snapshots (Tian et al., 7 Feb 2025). Sequence-aware DNN and LSTM models also appear in volunteer-engagement continuation, where recent inter-event gaps and cumulative behavior are combined to predict whether future activity exceeds a threshold $|S|$7 (Semenov et al., 2022).

A different line treats future engagement outputs as a profiling representation rather than the final endpoint. The game-analytics paper predicts three survival curves—days after first login, level reached, and accumulated playtime—plus lifetime value, then segments users into short, medium, long, and loyal classes or low, medium, and high spender classes (Río et al., 2019). This suggests a multi-trajectory view of engagement rather than a single scalar score.

Finally, one paper uses “Post Engagement Prediction” to name a self-supervised continued-pretraining task. There, PEP predicts whether post pairs in a conversation tree are in root, branch, or parent relations, with

$|S|$8

This is not direct engagement forecasting but a pretraining objective intended to encode interaction structure into LLMs (Cui et al., 10 Aug 2025).

5. Evaluation regimes and empirical regularities

Evaluation protocols vary substantially. Reported metrics include F-score for Facebook brand-post classification (Rosas-Quezada et al., 2019), Accuracy / Precision / Recall / AUC$|S|$9 / AUC$|R|$0 / F1 for TweetGage (Arazzi et al., 2023), RMSE / MAPE / $|R|$1 for IC-Mamba (Tian et al., 7 Feb 2025), NDCG@10 for ranking (Zamani et al., 2015), PRAUC and RCE for Twitter RecSys context-only prediction (Jeromela, 2023), AUC for interaction-log engagement inference (Carlton et al., 2021), and ROC-AUC for Reddit nutrition-based prediction (Ozegovic et al., 11 Feb 2025). Class imbalance is pervasive, and several papers highlight that binary high/low formulations can be strongly skewed toward the negative or low-impact class.

Some empirical patterns recur across otherwise different settings. On Facebook brand posts, content features consistently improve prediction for all six tasks; style is the second-best single non-content feature family; Decision Trees perform best without content; and Naïve Bayes or SVM dominate once text is included (Rosas-Quezada et al., 2019). On Twitter, relational context is useful but not sufficient alone: TweetGage reaches 0.89 Accuracy, 0.95 AUC$|R|$2, 0.94 AUC$|R|$3, and 0.89 F1, versus 0.72, 0.80, 0.80, and 0.72 for the strongest non-graph baseline, yet its ablation study shows that graph structure alone is around chance level and works best when combined with post, user, and text features (Arazzi et al., 2023).

Early forecasting under irregular observations has recently become a major theme. IC-Mamba reports RMSE $|R|$4 in the crucial first $|R|$5 minutes and claims a 4.72% improvement over state of the art across engagement metrics, while maintaining strong performance out to 28-day opinion-level forecasting (Tian et al., 7 Feb 2025). This contrasts with older regression setups such as blog-comment prediction, where unsupervised representation learning already showed that the representation–predictor pairing matters: sparse autoencoder reduced test RMSE for Linear Regression by 42% over baseline, while PCA gave the best Regression Tree result and a reported 15% RMSE improvement over baseline (Pham et al., 2016).

When the goal is ordering rather than count prediction, ranking objectives can be advantageous. In MovieTweeting, all learning-to-rank methods outperform all regression models, and aggregating LTRs with regressors via supervised Kemeny improves NDCG@10 from 0.8242044953 for LTRs-only aggregation to 0.8261454943 for LTRs+REGs (Zamani et al., 2015). By contrast, context-only Twitter engagement models show that strong contextual discrimination does not guarantee calibration: the thesis obtains challenge-like PRAUC values such as 0.7273 for likes and 0.7623 for the auxiliary react target, but all best challenge-like RCE values remain negative (Jeromela, 2023).

Adjacent engagement-inference work also reinforces the importance of temporal behavior. In interactive media, session length and time to completion were the strongest correlates of self-reported engagement, and Logistic Regression achieved $|R|$6, precision $|R|$7, recall $|R|$8, and $|R|$9 on the final held-out test (Carlton et al., 2021). Domain-specific content features can matter most for extreme outcomes rather than mild ones: on Reddit food posts, adding nutrition raises ROC-AUC from 0.669 to 0.709 for resonance prediction when added to controls, and calorie density is a top SHAP-ranked feature (Ozegovic et al., 11 Feb 2025).

6. Limitations, terminology, and research directions

The literature repeatedly exposes methodological weaknesses. Several influential datasets use random splits or pooled cross-validation rather than chronology-aware evaluation. The Facebook brand-post benchmark uses random 10-fold cross-validation with labels defined from the whole dataset average, creating a realistic possibility of temporal leakage and threshold contamination (Rosas-Quezada et al., 2019). TweetGage does not report explicit train/validation/test split ratios (Arazzi et al., 2023). The Reddit nutrition paper uses random 80/20 splitting with no temporal holdout (Ozegovic et al., 11 Feb 2025). Older challenge-style ranking work also contains leakage-prone features such as tweet age and possibly globally computed aggregates (Zamani et al., 2015).

A second recurrent limitation is incomplete observability. Many studies lack impressions, true exposure, or recommendation-position information, so they predict realized engagement conditional on data availability rather than conditional on actual opportunity to engage. The cryptocurrency work partly compensates by normalizing interaction by follower count, but even there follower count is only a proxy for exposure (Qureshi et al., 2022). Context-only Twitter modeling shows the opposite extreme: many sophisticated history features can be built without content, yet the resulting models still underperform substantially on probability-quality metrics (Jeromela, 2023).

The third issue is terminology. In most of the literature, PEP denotes forecasting future engagement outcomes. In (Cui et al., 10 Aug 2025), however, PEP denotes a self-supervised pretraining task over propagation trees. The underlying intuition is still about post interaction structure, but the objective is Root Prediction, Branch Prediction, and Parent Prediction rather than direct engagement forecasting. This suggests that “PEP” now names both an application family and a representation-learning strategy.

Several forward directions are already visible in the cited work. Chronological splits and leakage-resistant evaluation appear in IC-Mamba, which uses earliest 70% train, next 15% validation, latest 15% test (Tian et al., 7 Feb 2025). Relational modeling beyond isolated posts appears in tweet graphs, ego-graph replay, and propagation-structure pretraining (Arazzi et al., 2023, Bo et al., 2022, Cui et al., 10 Aug 2025). Richer domain semantics appear in nutrition estimation and in the survey’s emphasis that no universal feature set works across all networks (Ozegovic et al., 11 Feb 2025, Sadeque et al., 2019). Multimodal fusion by textual serialization into an LLM-readable “multimodal transcript” suggests another direction for low-label regimes, especially when the target is broad valence or engagement bands rather than exact counts (Ma et al., 2024).

A plausible synthesis is that mature PEP systems will combine several layers that earlier papers often studied separately: lexical or multimodal content, user and creator history, interaction affinity, relational context among posts or users, and time-aware forecasting under irregular observation. The surveyed evidence does not support a metadata-only view of engagement, nor a content-only one. It instead supports a composite view in which engagement is simultaneously semantic, behavioral, relational, and temporal.