This study proposes an integrated framework for predicting users' emotional reactions to films by leveraging a heterogeneous graph neural network (HGNN) that explicitly models three semantically distinct node types-Users, Movies, and Emotions-and four relation types (viewed, rated, dominant emotion, preferred emotion). The pipeline includes rigorous data cleansing, construction of a 68 knode/130 k-edge knowledge graph, initialization of multimodal node features, and training a twolayer relational graph convolutional network with class-balanced loss. On an 80 / 10 / 10 split the model attains Accuracy = 73.8 %, Macro F1 = 71.3 %, surpassing logistic regression and Random Forest baselines by 30.3 p.p. and 19.1 p.p. in accuracy, respectively. Recommendation-oriented metrics further confirm its effectiveness (Hit Rate@10 = 0.84, NDCG@10 = 0.79). Ablation reveals that incorporating emotion embeddings from EmoBank boosts Macro F1 by 4.2 p.p., while classweighting mitigates a 14 % imbalance-induced drop. Limitations include performance degradation for rare emotions (F1 < 0.60), “cold-start” sensitivity (−12 p.p. accuracy), and computational overhead when scaling beyond one million edges. Future work will explore dynamic HGNNs for temporal preference drift, multimodal feature fusion, few-shot adaptation for new items, and fairness-aware training to reduce detected gender bias to ≤ 3 %. These directions aim to enable lowlatency (≤ 120 ms) and ethically robust deployment in real-world recommender systems.
0000-0003-2389-3668 (Y. Halias); 0000-0002-2441-6292 (K. Lipianina-Honcharenko); 0000-0001-6541-0359 (M. Komar); 0009-0002-4293-7515 (M. Telka); 0009-0009-8829-0316 (V. Lukianchuk)
© 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
emotional response. Emotions are a key driver of user behavior, determining satisfaction with
a viewing experience and the desire to continue interacting with a platform [
Recent years have witnessed the rapid development of Graph Neural Networks (GNNs) [
Despite considerable progress in classical recommendation systems, the task of predicting
emotional reactions to content remains underexplored [
This paper is devoted to the development and evaluation of the Emotional GNN model, which focuses on predicting users’ emotional reactions in movie recommendation systems. A heterogeneous graph is proposed, where nodes represent users, movies, and emotional states, and edges represent their relationships. This structure allows the model to uncover deep interaction patterns and predict the likelihood of a specific emotional response from the user after watching a particular movie.
The aim of this study is to evaluate the effectiveness of the Emotional GNN approach in the task of emotion prediction and to compare its performance with traditional classification methods on tabular data. It is expected that leveraging the structural information of the graph and focusing on the emotional component will improve the quality of recommendations and open new avenues for the development of personalized recommendation systems.
The paper is structured as follows: Section 2 reviews related work and current approaches in the field of recommendation systems and emotion prediction. Section 3 describes the integrated approach and architecture of the proposed system. Section 4 presents the implementation of the Emotional GNN model and the data preparation process. Section 5 demonstrates the results of experimental evaluation and a comparative analysis of model performance. The study concludes in Section 6, which summarizes the findings and outlines directions for future research.
Over the past decades, recommender systems have undergone significant transformation, evolving from simple filtering approaches to complex hybrid and deep learning models. The most traditional and widely used methods include collaborative filtering, which relies on similarities between users or items [10, 11], as well as content-based approaches that analyze the characteristics of the movies themselves [12]. These methods form the foundation of modern recommender systems; however, they are limited in their ability to account for complex behavioral patterns and do not personalize recommendations at the level of the user’s emotional response.
In the field of emotion analysis, a significant number of studies focus on emotion
classification based on text, video, or audio signals [13–16]. Such systems are widely used in
review analysis, chatbots, and automated support systems; however, their application in the
context of recommendations remains limited. Some studies explore the use of emotions as
auxiliary features to improve recommendation quality [17], but they do not focus on directly
predicting the user’s emotional state [
At the same time, Graph Neural Networks (GNNs) have proven to be a powerful tool for
modeling relationships in complex structures such as social networks or recommendation
graphs [18–20]. Their use in recommender systems enables consideration of not only direct
interactions between users and movies, but also indirect connections through shared
preferences, genres, or other attributes [
Nevertheless, the idea of using graph neural networks for directly predicting the emotional
response to a movie remains underexplored [
Notably, the study in [24] proposed the UCCA-GAT model, which demonstrates competitive results in emotion classification tasks. On the SemEval-2018 Task-1C dataset, this model achieved an accuracy of 61.2%, and on GoEmotions — 71.2%, confirming the effectiveness of incorporating semantic text structure via graph representations.
Another study [25] introduced the Emotion-specific Transformer, which accounts for emotional characteristics in representations, though it is limited to textual sources and does not include structural information about user interactions. Nonetheless, the model achieved an accuracy of 61.9% and a macro-F1 score of 52.0% in the WASSA-2022 task, indicating improvements over baseline transformer architectures due to the inclusion of emotion-specific features.
Thus, both [24] and [25] consider the emotional component in the context of text classification, but do not model the structured interaction between users, movies, and emotions. In contrast, the heterogeneous graph neural network (HGNN) proposed in our work models the user–movie–emotion triad, enabling context-aware prediction of emotions in recommender systems. This approach not only achieves higher accuracy (Accuracy = 73.8%, Macro F1 = 71.3%) but also preserves the structural integrity of the emotional interaction environment. Therefore, the model represents a new class of solutions at the intersection of graph learning and affective computing, suggesting a significant degree of novelty compared to existing counterparts [26].
The process of building a system for predicting users' emotional reactions to films is implemented as an integrated pipeline that combines graph-based representation of interactions, deep learning using a Graph Neural Network (GNN), and evaluation of prediction quality in the context of personalized recommendations. A heterogeneous Graph Neural Network (HGNN) is chosen as the model, as it takes into account the types of nodes and relationships in the graph, enabling the formation of context-dependent entity representations. The method is implemented through sequential steps, which are described below and illustrated schematically in Figure 1.
The first stage involves aggregating and cleaning the data required for graph construction. The data sources include: - User interactions with films — views, ratings, likes; - Film metadata — genres, release year, description, keywords; - Emotional annotations — emotion labels obtained using NLP classifiers based on reviews (e.g., IMDb, TMDb) or manually annotated.
All data undergoes duplicate filtering, and records with missing key fields are removed. Names are normalized, transliterated, tags are cleaned, and genres are standardized into a unified format.
Based on the collected data, a heterogeneous graph G=(V,E) is formed, where the nodes represent three types of entities: - User — unique ID and, where possible, demographic attributes; - Movie — described by genre or textual features; - Emotion — from a predefined set (e.g., joy, sadness, anger, fear, excitement). Graph edges describe the following types of interactions: - User → Movie — rating or view (optionally weighted by rating); - Movie → Emotion — the emotion most frequently associated with the film; - User → Emotion (optional) — if user preferences for certain emotions are known.
A key feature is typification: each node and edge has a type (defined by functions ϕ(v), ψ(e)), allowing the model to account for context during training.
To train the graph model effectively, vector representations (features) must be assigned to each node: - For movies — one-hot vectors based on genre, or embeddings of descriptions; - For users — age, country, activity level (or random vectors if unknown); - For emotions — fixed embeddings can be used from resources like EmoBank.
Since some emotions are significantly more frequent (e.g., "joy" or "trust"), class balancing is applied: - oversampling / undersampling; - weighted loss functions; - regularization of certain edge types.
The proposed model (Fig. 2) is a heterogeneous knowledge graph that integrates three semantically distinct types of nodes — users, movies, and emotions — along with typed edges between them. This model serves as the foundation for building emotion prediction systems within the domain of personalized recommendations. For each node , a type function is defined as
ϕ ( v ) ϵ {User , Movie , Emotion } For each edge e, connecting nodes vi→vj, the edge type is defined as
ψ ( e ) ϵ {viewed , rated , dominant emotion , preferred emotion} (2) Thanks to the use of typed edges and semantically rich attributes, this structure effectively supports heterogeneous graph neural networks (R-GCN, HGT), enabling precise modeling of complex relationships between users, movies, and emotions. This ensures improved accuracy of emotion prediction, especially in multi-class classification tasks or multi-label recommendation scenarios. (1)
The model is trained using: - Loss functions: categorical cross-entropy or weighted binary cross-entropy; - Optimization: Adam optimizer, learning rate 0.001, Dropout = 0.3, L2 regularization; - Graph splitting: train/val/test — 80/10/10%.
To enhance generalization, early stopping is implemented — training stops if the macro-F1 metric on the validation set does not improve for 5 consecutive epochs.
The quality of emotion prediction is evaluated using the structured Table 1, presenting metrics according to the classification type [27] and usage context.
NDCG@k C ∑ T Pi Accuraty= i=1 N P = i R =
i F 1i=
T Pi T Pi+ F Pi
T Pi T Pi+ F N i 2 Pi Ri
Pi+ Ri Pμ+ Rμ 1 k
∑ rel j k j=1
k ∑ rel j Rec @ k = j=1 |R| MRR= 1 ∑N 1 N j=1 rank j
predicted emotions among examples (where is the number of classes)
for each class. , , are counts of true positives, false positives, and false negatives respectively
matrix that counts the number of cases when the real emotion i, and is predicted by j
all classes; sensitive to frequent emotions relj is an indicator of whether the j-th predicted emotion is relevant;
k rel 1 C Macro − F 1= ∑ F M1iacro-F1 gives equal weight
CC i=n1 to all classes; Weighted-F1 Weighted − F 1=∑ i F w1eights by class size i=1 N i Cij=|{x∨ y =i , ^y = j }|
HL= 1 ∑N ∑L ‖[ y jl ≠ ^y jl vp]erecdtoicrtse(dwehleemreentiss innubminbaerry NL j=1 l=1 of emotions)
HR @ k = N1 ∑j=N1 ‖[ Rel j ∩ Poprnreeeds etja,nkr≠tgi∅ent te]hmeottoipon-is predictions
The developed recommendation system prototype implements a complete pipeline for predicting users' emotional reactions — from constructing a domain-oriented graph “user ↔ movie ↔ emotion” to producing multi-label predictions. The graph representation enabled the integration of heterogeneous relationships (ratings, viewing events, genre affiliations) into a unified context, enhancing the informativeness of the surrounding structure for each node. After preprocessing, the graph was fed into a graph neural network trained using mini-batch stochastic gradient descent; early stopping based on validation macro-F1 was applied to control overfitting.
The software implementation was done in Python 3.10 using PyTorch as the core deep learning framework, supplemented by PyTorch Geometric for handling graph structures. The pandas and NumPy libraries provided efficient tabular data transformations, while scikit-learn was used to compute key metrics (precision, recall, macro-F1, ROC-AUC) and to build baseline comparison models. NetworkX was employed for preliminary graph construction and visualization, and Matplotlib and Seaborn facilitated graphical presentation of training curves and confusion matrices, simplifying result interpretation. The entire experimental logic is orchestrated within reproducible Jupyter notebooks with fixed random seeds to ensure replicability.
Input data are organized into four interrelated CSV files. The fileusers.csv contains unique user identifiers and demographic attributes (age, gender, country, etc.); movies.csv stores movie metadata (movie_id, genre tags, brief descriptions). The interactions.csv file aggregates “user_id × movie_id” interactions with rating or viewing event fields, serving as graph edges. Finally, emotions.csv associates each movie with one or multiple emotional labels, which serve as target classes. This modular structure facilitates system scalability and integration of additional sources, such as temporal or social features, to further improve emotional prediction accuracy.
Based on the data, a graph in the HeteroData format from the PyG library was constructed, where each node type had a separate feature matrix.
Key statistics of the constructed graph and the training parameters of the graph neural network are summarized in Table 2.
≈ 12 000 ≈ 6 000 > 80 000 ≈ 50 000 2 64 0.3 Adam 0.001 1024 50 prevent overfitting, an early stopping mechanism was implemented: training was halted if the macro F1 score on the validation set did not improve for 5 consecutive epochs. This allowed retaining the most generalized version of the model without losing performance due to overfitting on the training data.
In the experiment, 80% of the graph was used for training, 10% for validation, and 10% for testing, providing an independent assessment of the model’s generalization capability in the multi-class classification of eight dominant emotions in user–movie pairs. The convergence dynamics are shown in Fig. 1, which presents the training and validation loss and accuracy curves; these confirm stable training without overfitting after approximately 35 epochs.
Table 3 summarizes the subsystem metrics on the test set. The proposed GNN achieves an
Recall scores of 71.9% and 70.7%, respectively. The difference between Macro and Weighted F1 scores (71.3% vs. 72.6%) indicates a moderate class imbalance, which does not lead to significant bias of the model toward frequent emotions.
Comparison with two baseline approaches (logistic regression and Random Forest with a “flat” feature representation) is presented in Table 4. The GNN demonstrates an accuracy gain of 30.3 percentage points over logistic regression and 19.1 percentage points over Random Forest; similarly, Macro F1 increases by 33.1 and 21.4 percentage points, respectively. The most common classification errors occur between closely valenced pairs “excitement ↔ joy” and “sadness ↔ disappointment.” For rare classes, precision is lower; however, the use of class weights during training prevents the model from “ignoring” them, thus preserving generalization capability in an imbalanced environment.
The proposed integrated approach for predicting users' emotional reactions is based on a heterogeneous graph neural network that models the triad "user – movie – emotion," explicitly accounting for node and edge types [28]. On the test dataset, the system achieved an Accuracy of 73.8% and a Macro F1 score of 71.3%, outperforming logistic regression by 30.3 percentage points and Random Forest by 19.1 percentage points. The relevance of recommendations was confirmed by the metrics Hit Rate@10 = 0.84 and NDCG@10 = 0.79, while maintaining balanced sensitivity across classes was evidenced by a Balanced Accuracy of 71.6%, despite a moderate imbalance in emotional categories.
At the same time, experimental results revealed several limitations. The largest classification errors occurred within clusters of emotions close in valence ("joy ↔ admiration," "sadness ↔ disappointment"), where the F1 scores of some rare classes dropped below 0.60. The "cold start" effect for new users and movies reduces accuracy by approximately 12 percentage points, whereas increasing the graph size to 1 million edges raises the time per epoch by four times, demonstrating the need for distributed computing. Additionally, a bias towards a young male audience (~4 percentage points in Macro F1) was detected, indicating a risk of unfair recommendations.
Future research should focus on dynamic GNNs (DyHGT, TGN) to model temporal evolution of preferences and multimodal features (CLIP embeddings, audio features), which are expected to improve the Macro F1 score by 3–5 percentage points. Adaptive few-/zeroshot learning mechanisms can reduce accuracy losses during the "cold start" to ≤ 5 percentage points, while explainable GNNs (GNN-Explainer) and personalized weight regularization aim to narrow the fairness gap to ≤ 3%. Running A/B tests in real environments and online retraining will help maintain latency ≤ 120 ms and increase the User Satisfaction Score to ≥ 4.3/5.
In conclusion, the results confirm that the heterogeneous graph model provides a significant improvement in the accuracy and relevance of emotional recommendations compared to existing approaches [24, 25]. However, for industrial deployment, it is necessary to address issues of scalability, real-time embedding updates, and ensuring ethical and transparent system operation, thereby forming a roadmap for future research and enhancements.
The authors used GPT-4 and DeepL to prepare this paper: Grammar and Spelling Checker. After using these tools, the authors reviewed and edited the content as necessary and are solely responsible for the content of the publication. [10] Kipf, T. N., & Welling, M. (2017). Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Representations (ICLR). DOI: 10.48550/arXiv.1609.02907 [11] Schlichtkrull, M., Kipf, T. N., Bloem, P., van den Berg, R., Titov, I., & Welling, M. (2018).
Modeling Relational Data with Graph Convolutional Networks. In European Semantic Web Conference (pp. 593–607). Springer. DOI: 10.48550/arXiv.1703.06103 [12] Buechel, S., & Hahn, U. (2017). EmoBank: Studying the Impact of Annotation Perspective and Representation Format on Dimensional Emotion Analysis. In ACL. DOI: 10.48550/arXiv.2205.01996 [13] Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., & Bengio, Y. (2018). Graph Attention Networks. In International Conference on Learning Representations (ICLR).
DOI: 10.48550/arXiv.1710.10903 [14] Zhang, S., Yao, L., Sun, A., & Tay, Y. (2019). Deep Learning based Recommender System: A Survey and New Perspectives. ACM Computing Surveys, 52(1), 1–38. DOI: 10.1145/3285029 [15] Kusal, S., Patil, S., Choudrie, J., Kotecha, K., Vora, D., & Pappas, I. (2022). A Review on Text-Based Emotion Detection – Techniques, Applications, Datasets, and Future Directions. arXiv preprint arXiv:2205.03235. DOI: 10.48550/arXiv.2205.03235 [16] Soleymani, M., Garcia, D., Jou, B., Schuller, B., Chang, S. F., & Pantic, M. (2017). A survey of multimodal sentiment analysis. Image and Vision Computing, 65, 3-14. 10.1016/j.imavis.2017.08.003 [17] Cambria, E., Poria, S., Hazarika, D., &Kwok, K. (2020). SenticNet 6: Ensemble Application of Symbolic and Subsymbolic AI for Sentiment Analysis. In CIKM. DOI: 10.1145/3340531.3412003 [18] Wang, X., Ji, H., Shi, C., Wang, B., Ye, Y., Cui, P., & Yu, P. S. (2019). Heterogeneous Graph Attention Network. In The World Wide Web Conference (pp. 2022–2032). DOI: 10.48550/arXiv.1903.07293 [19] Zhang, X., Zhao, J., & LeCun, Y. (2015). Character-level Convolutional Networks for Text Classification. In Advances in Neural Information Processing Systems (NeurIPS). DOI:: 10.48550/arXiv.1509.01626 [20] Huang, W., Zhang, T., Rong, Y., & Huang, J. (2020). Graph Learning based Recommender
Systems: A Review. arXiv preprint arXiv:2012.06995. DOI: 10.48550/arXiv.2105.06339 [21] Sharma, N., Ahmad, M. A. F., Jajoo, D., & Sandhya, A. (2024). Emotion Based Movie
Recommendation System Using CNN and GNN. Studies in Science of Science, 42(11). [22] Ziaee, S. S., Rahmani, H., & Nazari, M. (2024). MoRGH: Movie Recommender System using GNNs on Heterogeneous Graphs. arXiv preprint arXiv:2401.03808. DOI: 10.21203/rs.3.rs-3860094/v1 [23] Bhattacharyya, S., Yang, S., & Wang, J. Z. (2025). A Heterogeneous Multimodal Graph Learning Framework for Recognizing User Emotions in Social Networks. arXiv preprint arXiv:2501.07746. DOI: 10.48550/arXiv.2501.07746 [24] Xia, L., et al. Emotion Classification in Texts Over GNNs: A UCCA-GAT Approach. In
SemEval, 2021. [25] Lee, J., et al. Leveraging Emotion Features with Emotion-Specific Transformers for Text
Classification. In ACL, 2022. [26] Zou, L., Xia, L., Ding, Z., Huang, J., & Hua, X. S. (2019). Reinforcement learning for userintent-driven conversational recommendation. Proceedings of the 2019 World Wide Web Conference, 2506-2516. DOI: 10.1145/3292500.3330668 [27] Lipianina-Honcharenko, K., Wolff, C., Sachenko, A., Kit, I., & Zahorodnia, D. (2023).
Intelligent method for classifying the level of anthropogenic disasters. Big Data and Cognitive Computing, 7(3), 157. DOI: 10.3390/bdcc7030157 [28] Turchenko, I., Osolinsky, O., Kochan, V., Sachenko, A., Tkachenko, R., Svyatnyy, V., & Komar, M. (2009). Approach to neural-based identification of multisensor conversion characteristic. In Proceedings of the 5th IEEE International Workshop on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS 2009) (pp. 27–31). IEEE. https://doi.org/10.1109/IDAACS.2009.5343030