Group Semantic Recommendation System using Attention Neural Network

Abstract Recommender systems (RS) are commonly used in areas such as online orders, travel, and music to suggest items that match user interests. With the rapid growth of social interactions and online activity, their use has naturally extended to both personal and Group Recommendation Systems. A group recommender system focuses on providing recommendations based on a group's shared preferences, rather than relying on a single individual’s choices. To overcome these challenges, we propose a novel BADLGRS developed. The GRL model was used to construct a tripartite graph representing interactions between the number of items, users, and group interactions. To effectively capture semantic group features, this phase introduces a novel model, GRUANN. The GCN model with two layers was used to learn user preferences under the GPL. A novel BADLGRS model is evaluated across four datasets and demonstrates superior performance when compared to existing methods. Specifically, it achieves accuracies of 0.893, 0.567, and 0.095, and a MAP of 0.095. Finally, the results of the BADLGRS model consistently outperform existing models in group recommendation tasks. Introduction Recommender-System (RS) have emerged as a topic solution to the difficulty of data overload, assisting users in locating pertinent information from massive databases by making tailored suggestions based on their prior preferences [1]. Usually, user interactions, including implicit or explicit ratings, are the source of these preferences [2]. While many real-world situations involve group interactions, such as eating with co-workers, shopping online orders, or organizing a trip with friends, traditional RS models tend to concentrate on single-user scenarios [3, 4]. This emphasizes the necessity of Group Recommender Systems (GRS), which are made expressly to offer suggestions that meet the preferences of a group as opposed to an individual [5]. With the growth of social networking sites and mobile devices, which make group interactions easier, the significance of GRS has increased [6]. Nevertheless, despite their importance, many current GRS models rely on conventional methods, including content-based modes, collaborative filtering (CF), and different types of models used [7, 8]. Many of the models are less useful in real-time applications like social media or collaborative decision-making because they frequently overlook the dynamic, complex and changing preferences of group members [9]. Additionally, these models frequently underrepresent social aspects, including history, connections and trust. New opportunities for GRS development have emerged with the rise of deep learning (DL) techniques, which have transformed domains such as computer vision (CV) and natural language processing (NLP), enabling the extraction of group recommendations [10,11,12]. Systems may develop more intricate representations of user preferences because of DL's ability to accurately characterize non-linear interactions of user and objects [13,14,15]. Moreover, DL models may use enormous volumes of data, including text and images, to find latent patterns that conventional models would overlook [16]. Because of this, DL is a desirable method for creating more efficient GRSs that are able to record the rich, multi-modal data that is frequently seen in actual user interactions [17,18,19]. Several deep learning models, used in recent years like Generative Adversarial Networks (GAN), Neural Networks (NN), and Long Short-Term Memory model were used to predictions [10, 12, 20,21,22,23,24,25]. While attention methods have been included to improve performance by concentrating on the most pertinent aspects for suggestions, these models are also good at simulating the temporal dynamics of user-item interactions [13, 15, 26,27,28,29,30]. Hybrid models, such as MLP and GRU, have also been investigated to increase recommendation accuracy by combining many deep learning architectures. Even with these improvements, dynamic, changing user preferences are still difficult for current deep learning-based GRS models to handle [31]. Many of the models in use today ignore the possibility of changing user preferences over time in favour of set user features [32]. By introducing two-way deep learning models that dynamically capture changing preferences, the proposed study overcomes this restriction and recovers the general efficacy and performance of GRSs in practical applications. The suggested BADLGRS differs from current techniques by including a amount of innovative components. In order to better capture and express both individual and group preferences, BADLGRS uses a dual-stage design that combines GRL and GPL, in contrast to conventional single-stage models. A unique aspect that improves the semantic comprehension of user preferences within the group context, something lacking in many previous models, is the use of a TG to represent the interactions between groups, users, and things. Additionally, deeper semantic feature refinement is made possible by the incorporation of a GRUANN, which is more sophisticated than the feature extraction methods employed in DFM-AVG and AGR. Furthermore, compared to more straightforward aggregation techniques used in current systems, the PLM used in our method for aggregating group preferences provides a more precise and complex modelling of collective behaviour. Objectives To maximize group recommendations, this work presents a unique method to GRS called the bi-stage adaptive deep learning-based group recommendation system (BADLGRS), which incorporates graph-based deep learning models. While CNNs, GANs and transformers, traditional deep learning models are frequently employed for picture super-resolution tasks, the proposed model stands out by utilizing a two-stage deep learning architecture designed especially for group recommendation problems. The proposed work aims to achieve the following main objectives: - 1) The effectiveness of the GRS relies on the accurate representation of group-based features extracted in this model. It effectively represents the group, user, and uses an undirected tripartite graph to perform. - 2) To accurately learn the GRS model in three different models trained to optimize the function, which are the GPL, TGCN, and PLM models. - 3) To verify BADLGRS with four different large and real-world application data to explore its efficiency in giving accurate performance. This research work explained the different sections given below. Section "Related Work" includes the literature survey. Section "Proposed Methodology" presents the suggested methodology. The results and discussion are included in Section "Results Analysis". Section "Conclusion" concludes with a discussion of the future scope. Related Work A personality-aware GRS algorithm is suggested [33], based on pairwise preferences. Users' personality traits are identified using 5 personality styles. Abbasi-Moud et al. [34] presented a fuzzy model for tourism attractions and for generating recommendations. Yannam et al. [35] suggested using slope one to create group recommendations (GRS). Jun and Peilin [36] presented a user dynamic profile information optimization based on GRS. The technique raises the level of satisfaction of all participants in the group. The dataset used in the experiment is Zhihu, and the simulation results show that the method is effective. Ni et al. [37] proposed two features extracted for RS: items and the number of users, based on deep learning. The depiction of item features is done using a based technique. Huang et al. [4] introduced a user-interaction approach, preferred in graphs, enabling accurate learning of group semantic features through advanced representation transfer learning to understand group-based interactions. Zhang et al. [38] proposed GRS, which is based on feature interaction (GRUIFI) and user importance. The GRS uses a multi-head attention network to extract high-order feature interactions among group members. Yang et al. [39] proposed a hybrid collaborative filtering approach to address sparsity issues. An algorithm to alleviate the sparsity difficulty is proposed for RS, while tackling the zero values. The algorithm substitutes the zeros according to the constraint conditions and the rules of convergence through multiples. The equations are then used in a collaborative way, and a hybrid approach is adopted to achieve better recommendation results. This approach uses an attention mechanism to explore the interpersonal relations within the group and come to a common standstill regarding the group members' preferences [40]. Zan et al. [41] developed an attention model of GRS for different people's attraction to user-based items. In this research, a neural embedding layer was used. The goal of the UDA mechanism is to use a difference operator to gather information among group users. The aggregation of group members' preferences is performed using an aggregation technique. The model allows assigning different weights to each user within a group. Jia et al. [42] suggested a group recommendation approach by using a hypergraph convolutional network (HCR). Liang et al. [43] suggested extracting fuzzy attention network group relations based on groups and user items. Pujahari et al. [44] developed a combination of a hidden Markov model and a collaborative filtering model for user preferences. Abolghasemi et al. 45 presented a novel group-based RS using pairwise favorite information, which could generate better group-based recommendations. Shi et al. [46] presented a Deep reinforcement learning model designed to make sequential RS. This model can extract long-term (LT) and short-term (ST) relevant recommendations for the changes in group interests. Regardless, the exploration/exploitation dilemma can be a challenge in DRL. The KD-based graph attention FI model (KD-GAFIM) is a lightweight combines knowledge distillation (KD) with graph neural networks (GNNs) [47]. To retain a modest model size while supplementing the feature vector with feature dependencies, the method makes use of graph attention networks. Additionally, KD-GAFIM increases the effectiveness of user CTR prediction by sharing the embedding layer of the instructor model. Furthermore, model interpretability was used to inform the proposal of a feature compression technique that finds high-contributing features for inference and model refinement, depending on how well they perform in model interpretability. Federated Heterogeneous Graph Contrastive Learning (FedHGCL), a novel approach presented by Wang et al. [48], builds various augmented perspectives for contrastive learning (CL) to improve FedRec by leveraging heterogeneous information. First, in order to improve recommendation performance, federated heterogeneous graph contrastive learning was implemented, in which each user locally builds a tiny CL-based multi-view framework. Second, in order to balance the training data for every user, a user sampling technique was created for client-side data augmentation (DA) and server-side model updates. Presenting a machine learning-based approach based on the real-time soil fertility context acquired using the Internet of Things (IoT) enabled soil fertility mapping, Khan et al. [49] aimed to improve the fertilizer recommendation system's accuracy. To help with real-time soil fertility mapping and context-aware fertilizer recommendations, an Internet of Things architecture was also proposed. Khan et al. [50] proposed an innovative approach to mobile phone rating categorization using federated learning and TF-IDF features. To categorize the recently generated Flipkart dataset, the proposed method involves building a federated deep neural network (FDNN). This method incorporates TF-IDF feature extraction, balancing, data cleaning, and federated learning prediction. Aung et al. [51] proposed a U-Net hybrid model to enhance feature extraction and improve accuracy. In this analysis, multiple mechanisms were used. Spatial Attention (SA), Dilated Efficient Channel Attention (Dilated ECA), Squeeze-and-Excitation (SE) Block, and Convolutional Block Attention Module (CBAM) were all components of a U-Net backbone. The purpose of carefully coupling these modules was to enhance feature representations at both local and global scales. Table 1 Comparison of different existing models with the Group Recommendation System. Problem Statement In this section, we will focus on GRS, methods for guiding users to a decision, and choosing amongst a myriad of options. As per the existing study, baselines for GRS are broadly divided into two categories: recommender-based memory systems and model-based GRS. The above methods simply sum users' preferences without accounting for group-based interactions among multiple users. Later, group-based decision models were presented to identify the best preferred items for each group information will be extracted. GRS were used with many of the latest DL methods, such as CNNs, RNNs, LSTMs, and MLPs. Furthermore, to gain a deep analysis of the temporal and sequential dynamics of user preferences, hybrid designs that combine these models are used. To improve the performance and accuracy of user preference representations, GRS has also included attention mechanisms that focus on the most important aspects. Feature selection techniques have also been shown to improve recommendation performance by selecting the most relevant features from the user-item interaction matrix. However, many existing models lack the ability to capture and adapt to dynamic user requirements, leaving a gap in accurately representing and adjusting to user preferences in real-time applications. A neural network model is used to improve accuracy. Still, many of the issues were reported in the GRS to resolve the DL models. The proposed work is performed based on two stages of a learning scheme to address these challenges in GRS using advanced DL models. Proposed Methodology In this framework, we have proposed a novel BADLGRS model, and a sequential network is used, such as GRL, tripartite graph (TG), and GPL. Next, a GRUANN is used to extract the semantic items. To optimize the various semantic features will be used for GRL model. In GPL stage, a two-layer graph convolutional network is used to acquire user preferences. Finally, group preferences are learned, and the network's two layers are optimized using the Pairwise Learning Method (PLM). The validated across four datasets, and the results using different metrics are reported in this research work. Figure 1 presents a step-by-step flow of the proposed methodology. Group Representation Learning (GRL) In this phase, GRL model is used to extract the semantic features, and user interaction is performed using GRUANN. Optimal parameters were used to extract the semantic users in the GPL stage. Next phase, user semantic features are created using an undirected tripartite graph \({U}_{YG}=\left(n,e\right)\). In \(n\) represents the number of nodes and \(e\) means edge value. Then the node set is represented as \(n={G}_{s}\cup {U}_{s}\cup {I}_{s}\), and the edge set is represented \(e=\left\{\left({Y}_{g},a\right)|{Y}_{g}\in {G}_{s}\right.\), \(a\in {U}_{s}\) where \(a\) is a member of \(\left.{Y}_{g}\right\}\cup \left\{\left({Y}_{g},b\right)|{Y}_{g}\in {G}_{s}\right.\) and \({Y}_{g}\) has interacted with \(\left.b\right\}\cup \left\{\left(a,b\right)|b\in {U}_{s}\right.b\in {I}_{s}a\) and has interacted with \(\left.b\right\}\) [52]. The mathematical equation of semantic values is below. The \(R\left(.\right)\) is a ReLU function used, \({U}_{s2}\) and \({h}_{s2}\). These are the parameters used. Group Semantic Feature learning (GSFL) is Based on the GRUANN The GRUANN model used to extract the best semantic-based features and optimized user features was represented. The proposed model used two types of layers: encoder and decoder. To improve the efficiency and performance of the training data, we used an attention layer and a weight-parameter decoder for the semantic feature extraction layer. The encoder and decoder of GRUANN are shown in Fig. 2a (Encoder diagram) and 2(b) (Decoder Semantic model), respectively. Figure 2 c shows the flowchart of the architecture of the GRUANN. Variational Sampling: By sampling latent variables at every node, the sampling step enables the model to represent the uncertainty of group preferences. Particularly when working with sparse data or changing user interactions, this is essential for understanding the diversity in group preferences. Graph Attention Method: The model may dynamically concentrate on the most important attributes and interactions among people, objects and groups due to the graph attention method. The model learns more precise group-level semantics by allocating distinct attention weights to various nodes and interactions. GRU-based Decoder: For group recommendation tasks where group dynamics change over time, the GRU aids in identifying temporal patterns and sequential relationships in the interaction data. The semantic module layer of the decoder further incorporates semantic importance. Every node \(v\in J\), neighbourhoods of the node is represented to \({N}_{n}=\left\{{q}_{1},{q}_{2}....,{q}_{z}\right\}\), and \(z\) indicates the cardinality \({N}_{n}\) [53]. The \(qi\) is nodes and denoted as \(W\left(qi\right)\). The initial feature \(qi\) denoted as \(\widetilde{q}i\)\({\kappa}_{qi}\) and is fully connected layers of function \({F}_{cl}1\) to created an \({d}_{d}\)-dimensional vector points as \({q}_{i}^{w}\) for \(qi\) and is represented as, The layer \({\mu}_{l}\) consists of eight parameters \(GRU{A}_{2}^{l}\) and is denoted as \({U}_{1}^{l}\in {\mathfrak{R}}^{{d}_{d}\times {d}_{0}},\)\({h}_{1}^{l}\in {\mathfrak{R}}^{{d}_{d}},{S}_{1}^{l}\in {\mathfrak{R}}^{{d}_{d}\times \left({d}_{g}+{d}_{0}\right)},\)\({R}_{1}^{l}\in {\mathfrak{R}}^{{d}_{d}},{U}_{2}^{l}\in {\mathfrak{R}}^{{d}_{d}\times {d}_{0}},\)\({h}_{2}^{l}\in {\mathfrak{R}}^{{d}_{d}},{S}_{2}^{l}\)\(\in {\mathfrak{R}}^{{d}_{d}\times \left({d}_{g}+{d}_{0}\right)}\) and \({R}_{2}^{l}\in {\mathfrak{R}}^{{d}_{d}}\). where, \(U\Delta =\left({u}_{\Delta }^{1},{u}_{\Delta }^{2},\dots ..{u}_{\Delta }^{n}\right)\) training functions used and matrix represented an \(S=\left({s}_{1},{s}_{2},...,{s}_{n}\right)\) and activation function of \(f\) used the hidden functions of proposed model. Here \(x\left(.\right)\) represents the sigmoid activation function \({U}_{s}\) and \({h}_{s}\) were used to train the model parameters. The proposed model used combination of GRU model of semantic values are represented \({m}_{s}\) and \({I}_{sf}\) used for to gain the each and every component values. GRUANN model of softmax functions and hidden values were in below equations. \({I}_{sf}^{s}\) is denoted as \(\left|{I}_{sf}^{s}\right|\), the training time parameter represents an \({U}_{s,j}\). The loss function \(LF\) of GRUANN is given below equations. In, \(\lambda q\) is the significance weights and manages the different components that are 0 and 1. Figure 2 c illustrates a unique hybrid model that combines variational autoencoder components with graph attention layers and a GRU-based decoder to improve group semantics learning. To prioritize the most pertinent interactions, feature vectors from users, things, and groups are aggregated using attention techniques in the GRUANN encoder block. To better manage missing or sparse data, variational sampling is used to represent the uncertainty in group preferences. By taking into account the sequential relationships in user-item interactions, the GRU-based decoder refines the group semantic characteristics and generates very accurate group suggestions. The suggested model can learn both local and global aspects thanks to this design, which enhances its capacity to produce suggestions for dynamic group situations. Group Preference Learning (GPL) Temporal Graph Convolutional Network Temporal Graph Convolutional Network (TGCN) is employed in the GPL network stage and is trained using PLM. It also optimization of the two-layer network architecture. The Graph Convolutional Network is designed with a two-layer in both ways user and group preferences (PS). Each layer consists of a single convolutional layer and two fully connected (FC) layers. One layer is dedicated to learning group preferences, while the other focuses on learning based on user activities. Figure 3 represent an network of GPL architecture with GPM, PLM and UPM modules. During training time, data was \({M}_{b}G\) of model groups \({G}_{s}\) at a time. The training data was generated based on the GPM \({T}_{ss}\). The GPL stage's architecture is shown in Fig. 3. The data \({M}_{b}G\) of groups is formed at a time during training. To extract group-specific characteristics, the use of convolutional layers is fully connected (Conv1). These attributes are processed by the UPM and GPM to identify preferences at the individual and group levels, which are then utilized to optimize suggestions. Figure 3 depicts the procedure and shows how the modules interact with one another. Below Table 2 performance of pseudocode for the training set of the Group Preference model (GPM). where, \({g}_{sf}\) and \({i}_{sf}\) are the semantic items, \(n\) represents the negative samples. Users are generated based on \({G}_{s}:{U}_{s}=\left\{{u}_{s}|\right.\) each \({g}_{s}\in {G}_{s},{u}_{s}\in {U}_{s}\) and \({u}_{s}\) is a member \(\left.{g}_{s}\right\}\). A training set can be used by applying \({U}_{s}\) user performance. The dimensions of the input data sent to the TGCN are indicated by the height and breadth of the convolutional kernels applied to a 1D tensor structure during the GRL stage. The input tensor to the TGCN in the suggested model has dimensions of \(O\times A\times E\), where \(O\) is represent an number of nodes (of the tripartite graph's users, items and groups), \(A\) value represent an number of channels that represent each node's feature dimension and \(E\) is the sequence length, or temporal dimension in the case of sequential data. To match the TGCN's input format, the learnt semantic characteristics from the GRL stage are processed and rearranged. Each node learns a feature representation as part of the graph attention process, which yields these characteristics. In order to forecast group preferences, this information is then structured as input for the GPL stage. Hence, \(F{C}_{1}\) and \(F{C}_{2}\) are represented as, From Eqs. (15) and (16), a flat vector \(Conv\) is denoted as \(\overleftrightarrow{Conv}\in {\mathfrak{R}}^{\left(\left[2{d}_{0}-{h}_{c}{w}_{c}/{c}_{l}\right]+1\right){c}_{n}}\). In this model we have used some training parameters represented as \({U}_{2}^{{g}_{s}}\in {\mathfrak{R}}^{{n}_{fc}^{1}\times \left(\left[2{d}_{0}-{h}_{c}{w}_{c}/{c}_{l}\right]+1\right){c}_{n}},\)\({h}_{2}^{{g}_{s}}\in {\mathfrak{R}}^{{n}_{fc}^{1}},{u}_{3}^{{g}_{s}}\in {\mathfrak{R}}^{{n}_{fc}^{2}\times {n}_{fc}^{1}},\) and \({h}_{3}^{{g}_{s}}\in {\mathfrak{R}}^{{n}_{fc}^{1}}\). The number of functions \(F{C}_{1}F{C}_{2}\) are denoted as \({n}_{fc}^{1}\)\({n}_{fc}^{2}\) and \({S}_{v}\) semantic features of scaling. The loss function \(L{G}_{pm}\) is defined as GPM. In \(L{U}_{pm}\) of UPM loss functions used. PLM The proposed work is based on top-K groups, with the top-ranking items reported, so that PLM is employed to optimize. In GRS, the models mainly used PLMs are Bayesian personalized ranking and regression-based pairwise loss models. Computational Complexity Computational Complexity of GPL The GRL mode was compared with computational complexity can be summarized in two main components: complexity of generating an undirected tripartite graph \({U}_{TG}\) and computational time of training GRUANN. The complexity of the time \({C}_{c}\) of the GRL model can be computed by combining the two parts as, Computational Complexity of GPL Consider the groups \({N}_{g}\), users \({N}_{u}\), and items \({N}_{i}\). In a recommendation system. Each one of the users has rated \({B}_{u}\) items \(\left({B}_{g}<<{N}_{i},{B}_{u}<<{N}_{i}\right)\) grouped as \({B}_{g}\) items. In \(O\left({N}_{g}{B}_{g}\right)\) and \(O\left({N}_{u}{B}_{u}\right)\) are used the GPM and UPM of training the both models. required to train the GPM and user preferences module (UPM). In this case, the overall number of interactions between group-user and item is represented as. The proposed work is based on top-K items are grouped based on depending the ranking of each and every user for which the model parameters are optimized using PLM. The PLM used for GRS is the Bayesian personalized ranking [26] and the regression-based pairwise loss. The aggregated with a weight coefficient \(\alpha \) denoted below, The hyper-parameter α is the weight and \(\phi \) represent an proposed model. Computational Complexity Semantic feature vectors generated by GRL are used to link the GRL and GPL phases. Users, objects, and groups' combined data is represented by these vectors, which are subsequently forwarded to the GPL stage for more processing and preference learning. The architectural design enables more reliable and efficient group recommendations by guaranteeing a seamless transition from graph-based learning to group preference prediction. Table 3 presents the proposed algorithm for BADLGRS. Results Analysis The model is deployed on the widely used Python library 'Tensorflow'. Four datasets were used to evaluate the models' efficiency: Mafengwo, CAMRa2011, and MovieLens-1 M. The performance of both models is evaluated using 5 different metrics HR (hit ratio), mean average precision (MAP), and normalized discounted cumulative gain (NDCG). These measures will look at the top-K recommendation performance of the K value considered as K = 1, K = 5, K = 10 and K = 15. Table 4 is the general hyperparameters of the proposed model. Extensive testing was used to get the hyperparameter values shown in Table 5. A grid search was conducted across a variety of hyperparameters, learning rates and feature vector dimensions, to guarantee optimal performance. The model's generalizability is assessed using the same architecture on the CAM Dataset, Mafengwo Dataset, ML-Rand Dataset, and ML-Simi Datasets, which were used in this research work. To maximize performance, particular hyperparameters were adjusted for every dataset. Because of the CAM Dataset's extensive interactions, four attention layers and a latent feature dimension of 128 were employed. Due to its low interaction complexity, the MFW Dataset benefited from three attention layers and a latent feature dimension was used in this research work. Due to its sparsity, the ML-Rand Dataset employed two attention layers and a latent feature used. Three attention layers and a latent feature dimension of 64 were used for the ML-Simi Dataset in order to capture user interactions based on similarity. Additionally, the batch size and learning rate were modified to fit the unique features of each dataset. These dataset-specific modifications guarantee that the suggested model maintains a consistent architecture while being successful across various datasets. Additionally, the model's capacity for generalization was assessed during training. Finally, the model training accuracy and performance on the four datasets were used to choose the final hyperparameters. Datasets Used For the implementation of the proposed model, compared with different datasets that are namely MFW, CAMRa2011, ML-Sim, and ML-Rand, these four datasets are freely available. MFW: In this work, we used the MFW dataset, which consists of the user's own venue data and the number of Gathering groups he/she joined or created on the tourism site (http://www.mafengwo.cn). The most popular database, many people visited the travel site. CAM: This is a movie rating dataset used as an example by a group of users/households in an event-driven manner. The data is collected from the website (http://2011.camrachallenge.com/2011). ML: Two datasets are chosen from the (grouplens.org/datasets/movielens/) site dataset. Next, consider two data sets, namely ML-Simi and ML-Rand. User group data similarities are low, and within-group similarities are high, which are part of the data set. The users are randomly divided in ML-Rand. ML-Simi: Using k-means clustering or other pertinent clustering techniques, users were grouped according to similar interests (such as movie genres or ratings). This guarantees that the groups reflect comparable tastes. ML-Rand: To provide variation in the group composition for robust recommendation tasks, groups were formed by randomly choosing a collection of users and objects. MovieLens dataset (for ML-Simi and ML-Rand): To ensure consistent user-item interactions, incomplete items were eliminated from the raw data. In the study, 80% of each dataset was utilized for training, while 20% was reserved for testing. Enough data is available for both model training and assessment due to this division. The datasets were cleaned to eliminate any missing or unnecessary information before splitting, and if required, ratings were standardized to a common scale. To ensure data quality, outliers and inconsistencies were also handled. To ensure consistency in comparison, resolution characteristics, such as rating scales and item categories, were standardized across all datasets. Table 5 presents the statistical analysis for each of the four datasets, which are compared with different statistical results. Performance Metrics The performance of the proposed model is compared with different state-of-the-art models, and we have considered different metrics represented in Fig. 4. The HR and NDCG metrics are compared with different K values like 1, 5,10 and 15, and these values are compared to existing models. Hit Ratio (HR) measures the top k items with high scores that are picked to form the top recommendations. NDCG: This metric is also compared with different existing models. The recommendations are calculated using some other metrics like accuracy, recall, f1-score, and MAP, which were compared with different existing models using different datasets. All the evaluation metrics are explained mathematically as below. represented as: Performance Analysis The efficiency of the GRL model was compared with different deep learning models that are deep factorization average [54], Some of the researcher were used dynamic based probabilistic matrix using neural network (DPM-NN) [4], social enhanced AGREE (SoAGREE) [55], DFM least misery (DFM-LM) [4], some attention mechanisms used for group recommendation (AGR) [4], DFM maximum satisfaction (DFM-MS) [4], attentive group recommendation (AGREE) [55], GCPP [45], DRL [46]and consensus model (COM) [40] compared with two metices like HR and NDCG using four different datasets. Analysis Based on CAM Dataset For the top-K recommendation evaluation, the performance compared with HR and NDCG metrics and these models are evaluated using CAM dataset comparison in Fig. 4 a-b. The results analysis of the proposed model gave better performance using HR metrics in two datasets NDCG and CAN. The goal of the proposed model is to construct a graph of the interaction between the group, the users and the items. When compared with existing models, the DFM-LM, DFM-MS, and DFM-AVG models give very poor accuracy on the datasets above. The reason is that the group decision making process is not represented in the models. The accuracy of the compared models (SoAGREE, AGR, AGREE) is average due to the lack of an optimization method for the network layer. The proposed mode compared with HR@15 metrics, with 90.9% achieved, and other models are reported to have very poor performance. Second dataset NDCG and HR @15 of the proposed model achieves 53%, and compared with other existing models are reported to have very poor performance (47%, 46% and 26% etc.). After comparing the existing and proposed model gave better on the CAM data set other models, both in terms of HR and NDCG. MFW Dataset Comparison with HR and NDCG Metrics To assess the top-K recommendation performance, two performance metrics were used: HR for the MFW dataset (Fig. 5 a),MFW dataset comparison with HR metrics (Fig. 5 b). Here the proposed mode compared with metrics with HR@10 gives better accuracy at 71% and existing models SoAGREE @64%, DFM-LM@63%, DFM-AVG@62, COM@54, DFM-MS@61, AGREE@64, DPMF-CNN@63, AGR@64, GRMTDL@68, DRL@69, and GCPP@61 gave poor accuracy performance. Next metrics NDCG@10 proposed model archives an 49%, whereas the other models attain SoAGREE@43%, DFM-LM@35%, DFM-AVG@34%, COM@37%, DFM-MS@31%, AGREE@42%, DPMF-CNN@37%, AGR@44%, GRMTDL@46%, DRL@47%, and GCPP@44%. After comparing these models, our proposed model gives better performance in the MFW dataset only. Analysis of HR Metric with ML-rand Dataset Figure 6 a represents an analysis of HR metrics for the ML Rand dataset considered. To measure the top-K recommendation and performance of the second metric, NDCG is given in Fig. 6 b for the dataset ML-Rand. At HR@5, metrics is compared with proposed model gave high accuracy at 76%, and similarly, compared to different models that are SoAGREE@72%, DFM-LM@71, DFM-AVG@52, COM@42, DFM-MS@71, AGREE@72, DPMF-CNN@71, AGR@72, GRMTDL@74, DRL@75, and GCPP@73, model accuracy performed. Next, the NDCG@5 metrics value of the proposed mode gave an 59% performance, compared with NDCG@5 metrics, and reported poor accuracy. Therefore, the suggested model has outperformed other existing models with the dataset ML-Rand, compared to HR and NDCG. Analysis of HR Metrics with ML-Simi Dataset The performance of the HR metric also compared using ML-Simi dataset is shown in Fig. 7 a, and in Fig. 7 b, the performance of the NDCG metric on the ML-Simi dataset is displayed. In HR@1, the proposed model gave a high HR of 44%, accuracy and other models also compared SoAGREE@39, DFM-LM@33%, DFM-AVG@32, COM@24, DFM-MS@32, AGREE@39, DPMF-CNN@37, AGR@40, GRMTDL@42 DRL@34, and GCPP@32. Similarly, next metrics of NDCG@1, the proposed model achieves an 43%, and other models' performance are SoAGREE@39%, DFM-LM@33%, DFM-AVG@32%, COM@24%, DFM-MS@32%, AGREE@39%, DPMF-CNN@37%, AGR@40%, GRMTFL@42%, DRL@42%, and GCPP@40%. Hence, the suggested model is better in ML-Simi dataset compared to other models. Performance for Group Preference Learning The performance of the Group Preference Learning model is evaluated based on several models that are advanced version of AGR (AGR*), the advanced version of AGREE (AGREE*), SoAGREE, and the advanced version of SoAGREE (SoAGREE*) on the MFW dataset in terms of HR and NDCG. HR performance metric on the MFW dataset is illustrated in Fig. 8 a. NDCG metric performance on the MFW dataset for top-K recommendations is illustrated in Fig. 8b. The model SoAGREE* is a recommendation accuracy metrics compared, but performs poorer than the proposed model. This model was built by using an average learning method, SoAGREE*constructed. Compared to other techniques, such as (AGR, AGR*, AGREE, AGREE*, and SoAGREE), performed less well in NDCG and HR. This is because the models adopted utilized some ineffective learning techniques to learn the user preferences accurately. Moreover, the GPL module is not aware of how user preferences for a group affect the performance of a particular recommendation on a specific item. Performance Analysis of BADLGRS In Table 6. The proposed model is compared with eight baseline models using K = 5, 10, and 15 on both datasets. Generally, single-attention-based baseline models exhibit moderate performance compared to the other models. In this analysis, we used the top K items (5, 10, and 15), but K@15. The proposed model achieved 89% accuracy for ML-Simi and 42% for ML-Rand. In comparison, DFM-LM@72%, DFM-AVG@71%, COM@74%, DFM-MS@the 71%, AGREE@76%, DPMF-CNN@75%, AGR@76%, and MAGRM@81% on ML-Simi, and DFM-LM@31%,DFM-AVG@31%, COM@32%, DFM-MS@30%, AGREE@34%, DPMF-CNN@33%, AGR@34%, and MAGRM@40% on ML-Rand, respectively. In general, the proposed model outperforms the current models in the accuracy of the recommended items on both datasets. Table 7 shows the recall that the proposed and existing models achieve on the ML-Simi and ML-Rand datasets for measuring top-K recommendation effectiveness. The proposed model at K@10, performs recall of 84% and 27% on the ML-Simi and ML-Rand data sets, respectively. In comparison, the baseline methods achieve recall values of 45%, 49%, 51%, 51%, 49%, 59%, 63%, and 71% on ML-Simi, and 19%, 14%, 14%, 15%, 29%, 20%, 19%, and 27% on ML-Rand, respectively. In general, the proposed model outperforms previous models on both datasets, demonstrating its efficiency at retrieving the corresponding item in top-K recommendations. In order to assess the effectiveness of the model compared using top-K recommendations, the performance of the proposed models and the existing models tested on the two datasets, namely ML-Simi and ML-Rand, is presented in Table 8. The proposed model would be able to obtain a MAP of 49% on the dataset ML-Simi and 14% on the dataset ML-Rand at K@5. In comparison, the baseline methods obtain MAP values of 26%, 32%, 27%, 38%, 39%, 37%, 38%, and 45% on ML-Simi, and 7%, 4%, 8%, 8%, 11%, 13%, 12%, and 14% on ML-Rand, respectively. Table 8 shows the performance of top-K recommendation on two datasets: ML-Simi and ML-Rand, using MAP metrics. The proposed model is evaluated over different time intervals with other existing models. Figure 9a represent an months wise data compared with proposed and existing model and Fig. 9b represent proposed model compared with f1-score. The results indicate that the performance of the proposed method is similar to some of the previous methods. This is primarily because of the lack of historical data on users' preferences for training. Complexity Analysis In computational time complexity of both the proposed model compared with existing models group recommendation system models is illustrated in Fig. 10. The proposed method requires a computation time of 0.15, which is lower than all compared baseline methods. In contrast, the existing methods DFM_AVG@0.31, DFM_LM@0.29, DFM_ms@0.28, COM@0.25, DPMF_CNN@0,24, AGR@0.23, AGREE@0.20, SOAGREE@0.19, GCPP@0.18, GRMTDL@0.17, and DRL@0.17 respectively. Statistical Test The statistical analysis is performed using the Wilcoxon test, a nonparametric test for ranking and paired samples. In this context, a low p-value indicates that the proposed method is statistically significant. The results of the statistical test are presented in Table 9. Discussion The GRS model compared with different datasets using the top K recommendations is calculated in Table 10. In this work, a method is shown based on different metrics of evaluation. Badlgrs-based Group Recommendation System (GRS) is based on a sequential learning approach, which includes two steps: GRL and GPL. A tripartite graph (TG) is used to model the Group–User–Item interactions in the GRL stage, while the proposed GRUANN model aggregates the semantic features. The GRUANN architecture has two parts: the encoder part that includes attention mechanisms, and the decoder part that includes GRU units. During the GPL phase, a TGCN is used to learn group preferences and its training is conducted by PLM strategy. For better convergence and model performance, Adam optimizer is used to optimize the loss function of TGCN. The model enhances the GRL stage by introducing a tripartite graph (TG) to capture complex Group–User–Item (GUI), thereby enabling a more detailed understanding of group behavior. Furthermore, BADLGRS enhances the GPL by utilizing the intrinsic graph structure in the recommendation problem via a TGCN. By using the PLM strategy, the TGCN can differentiate positive and negative preferences, which leads to proper learning of group preferences. The model is also optimized for training with the Adam optimizer, a custom loss function, and a streamlined training process, which enhances the model's ability to efficiently and effectively learn from its data while maintaining a strong performance in the generalization tasks. In general, the proposed model performs better than the existing recommendation systems. Comparative performance of the different models is shown in Fig. 11 and Table 11. BADLGRS provides better performance in terms of accuracy and effectively understanding group behavior and preferences. It improves recommendation precision by identifying and suggesting the items that best match the interests of each group, thereby enhancing the overall recommendation quality. The model also achieves better recall by ensuring that important and relevant items are not missed, unlike many traditional recommendation approaches. Limitations Although the proposed model demonstrates strong performance in an experimental study, its scalability for large-scale and real-world applications in significant challenge. Scalability is an issue because the model's complexity can cause it to perform poorly on large-scale graphs (millions of users/items). Furthermore, the model can't manage ad hoc group formations in real-time circumstances because it presently depends on established groups. Additionally, the model has issues with cold-starting for new users or objects and data sparsity, which might affect the quality of recommendations. Addressing these limitations can further improve the effectiveness and real-time applications of Group Recommendation Systems (GRSs). Future work will mainly focus on large applications such as AI, e-commerce, tourism, business, etc. Ablation Study In this ablation study, we will compare four different methods and four different datasets. Each of these modules is singled out for analysis to identify its effectiveness. The HR values of the proposed model and the ablation study results are shown in Fig. 12. FLOPs Table 12 presents a computational complexity analysis of the proposed model compared with various existing models across four datasets. Inference Time (ms) indicates how long it takes the suggested model to provide predictions following input. This will depend on the computational configuration and the volume of input (number of users and objects). The floating point operations per second needed by the suggested model are shown as Resource Usage (GFLOPs). The amount of RAM used by the model during inference is referred to as memory utilization. Other methods to obtain this include using profiling tools or just monitoring the RAM use while testing. The model's processing time (ms) is the amount of time needed to extract features and do any other pre-processing required before the recommendation-generating process really starts. Run Time (ms) is the overall amount of time needed for BADLGRS to provide suggestions, encompassing all input to output phases. It blends processing time with inference time. The ratio of accuracy to time (e.g., time per recommendation) is used to assess efficiency. This must be determined using the inference time and performance measurements, including accuracy, recall, and F1-score. The results show that the suggested model outperforms other cutting-edge models like DFM_AVG, DFM_LM, AGREE, and GCPP in terms of complexity efficiency. These results show that the proposed model is very effective for real-time illness detection in real-world scenarios, combining excellent accuracy with significantly lower memory and processing requirements. Conclusion This work presents a new BADLGRS that improves group recommendation. The framework can be divided into two modules: GRL and GPL models. For the GRL model, it is a semantic feature vector and learning of the user items is represented. In the GPL and TGCN network, a novel learning approach, named PLM, is trained. This technique is used for both training the model and optimizing the 2-layer network architecture. In this research, we used four datasets to validate the proposed and state-of-the-art models using different metrics, including HR, NDCG, MAP, accuracy, recall, and F1 score, to evaluate their performance. The experimental results analysis shows that the GRS of the proposed model achieves an accuracy of 0.893, a recall of 0.567, and a MAP of 0.095. Moreover, the proposed model, compared with other existing models, has a time complexity of 0.15 s., but this is higher in the real-world application domain. Therefore, a light-weight model will be developed for the future to address the challenges. In addition, Explainable AI will be used to enhancement of the mode. Further, this study would like to extend the recommendation system to other areas and evaluate its performance. Data Availability The data supporting this study's findings are available on request from the corresponding author. References Misztal-Radecka J, Indurkhya B. Bias-aware hierarchical clustering for detecting the discriminated groups of users in recommendation systems. Inf Process Manag. 2021;58(3):102519. Wei G, Wu Q, Zhou M. A hybrid probabilistic multiobjective evolutionary algorithm for commercial recommendation systems. IEEE Trans Comput Soc Syst. 2021;8(3):589–98. Jia JJ, Yao YW, Chen WH. A group recommendation algorithm based on non-negative matrix factorization. Comput Eng Sci. 2022;44(5):933. Huang Z, Liu Y, Zhan C, Lin C, Cai W, Chen Y. A novel group recommendation model with two-stage deep learning. IEEE Transactions on Systems, Man, and Cybernetics: Systems 2021. Kujur A, Raza Z, Khan AA, Wechtaisong C. Data complexity based evaluation of the model dependence of brain MRI images for classification of brain tumor and Alzheimer’s disease. IEEE Access. 2022;10:112117–33. Gupta D. A comprehensive study of recommender systems for the internet of things. J Phys: Conference Series. 2021;1969(1):012045. Li K, Wang CD, Lai JH, Yuan H. Self-supervised group graph collaborative filtering for group recommendation. In Proceedings of the sixteenth ACM international conference on web search and data mining 2023;69–77. Wu X, Xiong Y, Zhang Y, Jiao Y, Zhang J, Zhu Y, et al. Consrec: Learning consensus behind interactions for group recommendation. Proc ACM Web Conf. 2023;2023:240–50. Ullah F, Pun CM, Mohmand MI, Mahendran RK, Khan AA, Alhammad SM, ... Farouk A. Privacy-aware secure data auditing for cloud-based intelligence of things environment. IEEE Int Things J 2025;12(11):15288–15303. Xu J, Chen Z, Li J, Yang S, Wang H, Ngai EC. Aligngroup: Learning and aligning group consensus with member preferences for group recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management 2024a;2682–2691. Javed U, Shaukat K, Hameed IA, Iqbal F, Alam TM, Luo S. A review of content-based and context-based recommendation systems. Int J Emerg Technol Learn (iJET). 2021;16(3):274–306. Xu W, Wu Q, Liang Z, Han J, Ning X, Shi Y, ... Zhang Y. SLMRec: Distilling large language models into small for sequential recommendation. arXiv preprint arXiv:2405.17890 2024b. Liu Y, Zhu S, Yang T, Ma J, Zhong W. Identify then recommend: towards unsupervised group recommendation. Adv Neural Inf Process Syst. 2024a;37:96101–26. Afoudi Y, Lazaar M, Al Achhab M. Hybrid recommendation system combined content-based filtering and collaborative prediction using artificial neural network. Simul Model Pract Theory. 2021;113:102375. Liu Q, Wu X, Wang Y, Zhang Z, Tian F, Zheng Y, et al. Llm-esr: Large language models enhancement for long-tailed sequential recommendation. Adv Neural Inf Process Syst. 2024b;37:26701–27. Khairy M, Mahmoud TM, Abd-El-Hafeez T. The effect of rebalancing techniques on the classification performance in cyberbullying datasets. Neural Comput Appl. 2024;36(3):1049–65. Sharma S, Rana V, Kumar V. Deep learning based semantic personalized recommendation system. Int J Inform Manage Data Insights. 2021;1(2):100028. Som A. Automated student group collaboration assessment and recommendation system using individual role and behavioral cues. Front Comput Sci. 2021;3:728801. Omar A, Abd El-Hafeez T. Quantum computing and machine learning for Arabic language sentiment classification in social media. Sci Rep. 2023;13(1):17305. Li N, Gao C, Jin D, Liao Q. Disentangled modeling of social homophily and influence for social recommendation. IEEE Trans Knowl Data Eng. 2022;35(6):5738–51. Yang Y, Wu L, Liao Y, He Z, Shao P, Hong R, Wang M. Invariance matters: Empowering social recommendation via graph invariant learning. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval 2025;2038–2047. Lin G, Hua Z, Feng T, Yang S, Long B, You J. Unified semantic and ID representation learning for deep recommenders. arXiv preprint arXiv:2502.16474 2025. Duan J, Zhang PF, Qiu R, Huang Z. Long short-term enhanced memory for sequential recommendation. World Wide Web. 2023;26(2):561–83. Badawy A, Fisteus JA, Mahmoud TM, Abd El-Hafeez T. Topic extraction and interactive knowledge graphs for learning resources. Sustainability. 2021;14(1):226. Kumar C, Kumar M. Session-based recommendations with sequential context using attention-driven LSTM. Comput Electr Eng. 2024;115:109138. Zhang C, Zhang X, Shi T, Xu J, Wen JR. Test-Time alignment for tracking user interest shifts in sequential recommendation. arXiv preprint arXiv:2504.01489 2025. Salih HS, Ali MH, Khan MI. IoT-Enabled cloud storage data access control model based on blockchain technology. Int J Theoretical Appl Computational Intell 2025. Huang Z, Ni J, Yao J, Xu X, Zhang B, Chen Y, ... Xue C. Social group recommendation with TrAdaBoost. IEEE Trans Computational Soc Syst 2020;7(5):1278–1287. Zare G. TiLLM-Rec: Temporal-Interval-Aware Large Language Model for Sequential Recommendation under Irregular User Interactions. Available at SSRN 5453594 2025. MamdouhFarghaly H, Abd El-Hafeez T. A new feature selection method based on frequent and associated itemsets for text classification. Concurrency Comput Pract Exp. 2022;34(25):e7258. Cho J, Hyun D, won Lim D, Jae Cheon H, Park HI, Yu H. Dynamic multi-behavior sequence modeling for next item recommendation. Proc AAAI Conf Artif Intell. 2023;37(4):4199–207. Vasavi B, Dileep P, Srinivasarao U. Optimized aspect and self-attention aware LSTM for target-based semantic analysis (OAS-LSTM-TSA). Data Technol Appl. 2024;58(3):447–71. Abolghasemi R, Engelstad P, Herrera-Viedma E, Yazidi A. A personality-aware group recommendation system based on pairwise preferences. Inf Sci. 2022;595:1–17. Abbasi-Moud Z, Hosseinabadi S, Kelarestaghi M, Eshghi F. CAFOB: context-aware fuzzy-ontology-based tourism recommendation system. Expert Syst Appl. 2022;199:116877. Yannam VR, Kumar J, Babu KS, Patra BK. Enhancing the accuracy of group recommendation using slope one. J Supercomput 2022;1–42. Jun Y, Peilin L. A New Method of Group Information Recommendation Based on the User Dynamic Profile Information Optimization. In 2021 7th International Conference on Information Management (ICIM) (pp. 57–61). IEEE. 2021. Ni J, Huang Z, Cheng J, Gao S. An effective recommendation model based on deep representation learning. Inf Sci. 2021;542:324–42. Zhang J, Cheng J, Zhou Y, Yang Q. GRUIFI: a group recommendation model covering user importance and feature interaction. J Internet Technol. 2021;22(5):1143–55. Yang X, Zhou S, Cao M. An approach to alleviate the sparsity problem of hybrid collaborative filtering based recommendations: the product-attribute perspective from user reviews. Mob Netw Appl. 2020;25(2):376–90. Chen L, Cao J, Chen H, Liang W, Tao H, Zhu G. Attentive multi-task learning for group itinerary recommendation. Knowl Inf Syst. 2021;63(7):1687–716. Zan S, Zhang Y, Meng X, Lv P, Du Y. UDA: a user-difference attention for group recommendation. Inf Sci. 2021;571:401–17. Jia R, Zhou X, Dong L, Pan S. Hypergraph convolutional network for group recommendation. In 2021 IEEE International Conference on Data Mining (ICDM) (pp. 260–269). IEEE. 2021. Liang R, Zhang Q, Wang J. Hierarchical Fuzzy Graph Attention Network for Group Recommendation. In 2021 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) (pp. 1–6). IEEE. 2021. Pujahari A, Sisodia DS. Handling dynamic user preferences using integrated point and distribution estimations in collaborative filtering. IEEE Trans Syst Man Cybern Syst. 2022. https://doi.org/10.1109/tsmc.2022.3148675. Abolghasemi R, Viedma EH, Engelstad P, Djenouri Y, Yazidi A. A graph neural approach for group recommendation system based on pairwise preferences. Inf Fusion. 2024. https://doi.org/10.1016/j.inffus.2024.102343. Shi Y, Ren X, Wang X. MAGNET: Multi-Interest Attentive Group Recommender with Deep Reinforcement Learning. In 2023 2nd International Conference on Artificial Intelligence and Intelligent Information Processing (AIIIP) (pp. 280–286). IEEE. 2023. Yang B, Zhou J, Zhang S, Xing Y, Jiang W, Xu L. A lightweight knowledge distillation and feature compression model for user click-through rates prediction in edge computing scenarios. IEEE Internet Things J. 2024;12(3):2295–308. Wang Y, Wang Y, Zhang Y. Federated heterogeneous graph contrastive learning for privacy-preserving recommendation. Tsinghua Sci Technol 2026. Khan AA, Faheem M, Bashir RN, Wechtaisong C, Abbas MZ. Internet of things (IoT) assisted context aware fertilizer recommendation. IEEE Access. 2022;10:129505–19. Khan AA, Driss M, Boulila W, Sampedro GA, Abbas S, Wechtaisong C. Privacy preserved and decentralized smartphone recommendation system. IEEE Trans Consum Electron. 2023;70(1):4617–24. Aung TMM, Khan AA. Enhanced U-Net with attention mechanisms for improved feature representation in lung nodule segmentation. Curr Med Imaging. 2025;21(1):E15734056386382. Guo L, Zhu Y, Gao M, Tao Y, Yu J, Chen C. (2024, August). Consistency and discrepancy based contrastive tripartite graph learning for recommendations. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 944–955. Salem FM. Gated RNN: the gated recurrent unit (GRU) RNN. Recurrent neural networks: from simple to gated architectures. Cham: Springer International Publishing; 2021. pp. 85–100. Zeng B, Shang Q, Han X, Zeng F, Zhang M. RACMF: Robust Attention Convolutional Matrix Factorization for rating prediction. Pattern Anal Appl. 2019;22(4):1655–66. Cao D, He X, Miao L, Xiao G, Chen H, Xu J. Social-enhanced attentive group recommendation. IEEE Trans Knowl Data Eng. 2019;33(3):1195–209. Hu L, Cao J, Xu G, Cao L, Gu Z, Cao W. Deep modeling of group preferences for group-based recommendation. Proc AAAI Conf Artif Intell 2014;28(No. 1). Funding This research did not receive any specific grant from a funding agency in the public, commercial, or not for the public sector. Author information Authors and Affiliations Contributions Gopisetty Rathnamma, methodology, software, validation; Ulligaddala Srinivasarao, formal analysis, investigation, resources; Kommanaboyina Sai Vijaya Lakshmi, data curation, writing—original draft preparation; Vadige Sathish Kumar, writing—review and editing, visualization, project administration. Corresponding author Ethics declarations Ethical Approval NA. Conflict of interest The authors declare no potential conflict of interest. Additional information Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. About this article Cite this article Rathnamma, G., Lakshmi, K.S.V., Kumar, V.S. et al. Group Semantic Recommendation System using Attention Neural Network. Cogn Comput 18, 105 (2026). https://doi.org/10.1007/s12559-026-10650-2 Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s12559-026-10650-2