Figure 1 shows the network structure of BGFM. The graph flexibly represents higher-order associations between features. Edges in BGFM are useful feature interactions obtained by model aggregation. After resolving the data imbalance, beneficial feature interactions are selected. After learning by the attention mechanism, different feature interactions are given different weights and jointly output for final prediction.

The BGFM will update the network step by step. The input values are processed through feature embedding as the initial data for BGFM, as e i 1 = e i , where e i k is the latest feature embedding for the k -th layer. The model initially has no pre-input edge information, so edges are first obtained by the interaction selection component. The resulting edge information is then aggregated to update the feature embeddings in the remaining regions.

Existing methods for unbalanced data learning can be divided into three categories. BGFM uses a data-level solution, while others need more flexibility and robustness. BGFM greatly simplifies the workload of model training, improves efficiency, and gets various classifiers. Borderline-SMOTE is an algorithm extended for SMOTE. It considers the effect of noisy samples, and the algorithm uses only a few classes of samples with the attribute Danger on the border to obtain new samples, yielding a balanced distribution of the training sample set. After solving the problem of perceived data imbalance, two main components exist in each layer of BGFM. Both of them are described in detail next.

We devised a mechanism to obtain favorable pairwise feature interactions in the paper. The mechanism is the inference of connections between perceptual features through the graph structure, which models higher-order connections between features. However, the edges connecting two nodes v i v j ∈ E exist deterministically, greatly simplifying the selection process compared to the direct introduction of gradient descent-based optimization techniques.

This limitation is resolved by replacing the set of edges E by heightened neighbors P , which P i j is explained as the likelihoods of v i v j ∈ E . It shows that the interaction between the features is very important. A different graph structure P k needs to be learnt at each k -th layer and comparing it with the previously derived graph. These treatments provide higher performance. Specifically, each layer of the model’s graph structure is fixed, culminating in fixed-form outputs. However, our model is characterized by adaptive learning and can model associations of beneficial features.

This section aims to design a metric function to obtain beneficial feature interactions. The P i j v i v j metric function calculate f s e i e j s the weights of the edges. NMF-based functions are used to evaluate the edge weights (He and Chua, 2017). The product of elements of these feature vectors is converted into a scalar using Multilayer Perception (MLP) with one hidden layer, which can be calculated as Equation 1.

where W 1 s , W 2 s , b 1 s and b 2 s are the inputs to the multilayer player. ReLU and Sigmoid activation functions are represented by δ ⋅ and σ ⋅ , respectively. It is worth noting the order of the inputs to f s is invariant, as f s e i e j = f s e j e i . The same pair of nodes have the same edge weights at this point. This successive graph structure modeling allows the gradient to backpropagate. Since there is no truth graph structure, the gradient here is defined by the deviation between the model’s estimated and actual values. Feature interactions are treated as one, and the weights are estimated using MLP. Euclidean distance or other distance metrics can also be chosen (Zhang W. et al., 2021).

After selecting the beneficial feature interactions, the feature representation is updated by performing an interaction aggregation operation. For the target feature node v i , the attention coefficient of each feature interaction is measured while aggregating its beneficial interactions with its neighbors. The learnable projection a and nonlinear activation function LeakyReLU are applied to measure the attention coefficients as Equation 2.

This implies the significance of interactions between features v i and v j In this paper, we only compute c i j of the node j ∈ N i . N i represents the neighbors of the node v i , which is the sum for features that are useful to interact with v i . In the paper, the following function Softmax is used to normalize them in all choices of j , as shown in Equation 3.

where Y i is the output value of the i -th node, the output values of the multi-classification range vary from 0 to 1; k is the number of nodes that the network finally outputs, that is, the number of categories that can be classified. This makes it easy to compare the coefficients obtained between different feature nodes. After obtaining the normalized attention coefficients, the linear and nonlinear combinations of links between features are computed as subsequent new feature inputs as Equation 4.

where a i j measures the attentional coefficient of feature i and feature j interactions, and b i j indicates the probability that such feature association is helpful. a i j The attention coefficient is computed via the soft-attention mechanism and b i j is computed by the hard-attention mechanism. The information about the selected feature interactions is controlled by multiplying them and making the input values of the feature interaction selection mechanism learnable by gradient backpropagation.

To capture the diverse polysemy of feature interactions in different semantic subspaces and to stabilize the learning process, this paper extends our mechanism by applying multi-head attention (Li et al., 2017; Marcheggiani and Titov, 2017; Wang et al., 2019). Specifically, H individual attention mechanisms perform the update of Equation 4 and then concatenate these features to produce an output feature representation as Equation 5.

where | | denotes the cascade, a i j is the normalized value obtained through the h -th attention machine with W a h is the linear transformation matrix of the former. Optionally, the feature representation can be updated using average pooling, as shown in Equation 6.

The results for k -th layer is e 1 k e 2 k … e n k , which is a collection of n feature representation vectors. Because the representations acquired in multiple layers model different orders of interactions, they play different roles in the ultimate result. Thus, they are connected in series to get the definitive expression for every feature (Beck et al., 2018) as Equation 7.

Finally, all the feature vectors are pooled equally to get the result at the graph level, and the final prediction is made using the projection vector p . The obtained results are computed using Equations 8, 9:

The research focuses on exploring the effects of multiple influencing factors (including user, system, and contextual ones) on the perceived QoE of game users. A general taxonomy of the various factors in the literature is drawn upon, and further references are made to the taxonomy of existing game-related studies in terms of game QoE. Finally, an empirical test method is derived (Pornpongtechavanich et al., 2022; Jiang et al., 2019). Specifically, this gaming dataset considers the effects of three different system factors (latency, packet loss rate, jitter, and additional network parameters), user skills (user-personal factors in terms of gaming experience), and context (in terms of action categories and social context). The game entity under study is Glory of Kings, a game in which the interaction is mainly based on the UDP protocol, which requires a high level of real-time and user engagement. Due to the lack of a dataset of user game perception, the testing process in the study’s laboratory environment was determined after reviewing the relevant literature.

In joint efforts of team members and participants, data from 789 games were collected, with each piece of data representing three dimensions of user, service, and environmental data. Each dataset has 21 features, consisting of 4 pieces of user data (player ID, age, gender, and skill level), 16 pieces of in-game and post-game service data, and a user-perceived score for the last one.

In this paper, to demonstrate the effectiveness of the proposed algorithm, we compared it with algorithms from four categories: (A) linear methods, (B) FM-based methods, (C) DNN-based methods, and (D) aggregation-based methods. The specific eight comparison algorithms include LR (A), Standard FM (Rendle, 2010) (B), NFM (He and Chua, 2017) (C), AFM (Xiao and Ye, 2017) (B), AutoInt (Song et al., 2020) (D), Fi-GNN (Cui et al., 2019) (D), InterHat (Li et al., 2020) (D).

LR is a linear regression, modelled using only a single feature; Standard FM is able to model second-order interaction links of features; NFM designed a dual interaction layer and DNN to handle nonlinear features and model higher order feature interactions; AFM introduces the attention mechanism to give weight to the interaction of different features; AutoInt is to improve the efficiency of the model in learning higher-order feature interactions through self-attentive networks; Fi-GNN uses gated graph neural networks to model higher-order feature connections as fully connected graphs; InterHat uses the attention machine to select features, and raw feature multiplication produces higher-order feature interactions; LTFM is an extended FM-based model that considers the effects of temporal and spatial information projections and feature interactions on the final game perception results.

The following five assessment metrics are used in the experiments of game user perception evaluation: AUC (Gospodinova et al., 2023; Li et al., 2022), Precision (Annadurai et al., 2024; Chan et al., 2024), Recall (Wang et al., 2024; Zheng et al., 2024; Hong et al., 2024), and F-measure (Li et al., 2024a; Li et al., 2021; Li et al., 2023a; Guo et al., 2024a; Guo et al., 2024b; Ma and Tong, 2024; Sultan et al., 2024; Li et al., 2024b; Li et al., 2024c; Li et al., 2023b; Li et al., 2023c).

The AUC curve is taken as the area under the ROC curve. The larger the value, the better the model performance. Precision is used to calculate the proportion of correct predictions among all samples with positive predictions. Recall is the ratio of positive class samples correctly judged by the classifier to the total number of positive class samples. Usually, accuracy is inversely proportional to recall. A composite metric, F-measure, is introduced to balance the effects of precision and recall and to evaluate a classifier more fully. When both precision and recall are high, the value of F-measure is high.

The performance of the BGFM model after balancing the data was first analyzed experimentally, as shown in Table 1. It is concluded that there is an improvement in the BGFM model performance compared to the standard FM.

It can be deduced from Table 1 that FM performs well for sparse feature data. However, since poor game perception is a minority class occurrence, FM is impairing the correctness of the final judgment by judging the minority class as the majority class. Due to the data volume limitation, we consider preprocessing the data and de-rationalizing the generation of new data from the existing data to achieve the result that the data classes to be judged are basically the same. The balanced data is then imported into our model. The results show that using Borderline-SMOTE oversampling to balance the data category distribution is beneficial for the final perceptual evaluation.

The performance comparison of these methods on the game-aware dataset is shown in Table 2, from which the following observations are obtained: the BGFM proposed in this chapter achieves the best performance on the game perception dataset. The enhanced efficiency of the BGFM compared to the four classes (A, B, C, and D) of methods is particularly significant. BGFM employs a mechanism for choosing and aggregating beneficial feature interactions, and its performance is superior and easy to manage. Taking all aspects together, BGFM is superior to existing algorithms. In the following, the four types of algorithms, A, B, C, and D, will be specifically analyzed.

Aggregation-based methods outperform the other three classes of models, demonstrating the advantages of selection strategies in getting higher-order relationships. However, the LTFM model performance still performs well, suggesting that the importance of projected information that considers both temporal and spatial information interacting with and capturing features for the final game perception results is favorable for the final perceptual evaluation. However, this model only expands the data dimensions and captures hidden feature interactions based on the properties of FM for sparse data, which fails to address the performance fluctuations caused by the imbalance of data categories and captures higher-order beneficial feature interactions. The BGFM solves these problems well.

Compared to the powerful aggregation-based baseline AutoInt and Fi-GNN, BGFM still offers a significant performance improvement and can be considered important for game-aware prediction tasks. This enhancement is due to the combination of GNN with FM. Treating features as nodes, two-by-two interactions between features as edges, and each input as a graph, GNN’s aggregation strategy solves two of FM’s problems: suboptimal feature interactions that lead to model overfitting and the difficulty of modeling higher-order feature interactions. GNN introduces the concept of feature interaction and a beneficial interaction selection method that greatly improves the model’s performance.

The attention mechanism assigns different weights to interactions. AFM outperforms FM, demonstrating the necessity of considering feature interaction weights. Although NFM uses DNNs to model higher-order interactions, they do not ensure an improvement over the base model and the improved model with the addition of an attention mechanism, possibly because of their implicit feature interaction learning approach. AutoInt performs better than AFM because the multi-head attention mechanism in the model takes into account the richness of feature interactions in multiple spaces.