Enhancing Context-Aware Recommendation Using Trust-Based Contextual Attentive Autoencoder

Context-aware recommender systems are intended primarily to consider the circumstances under which a user encounters an item to provide better-personalized recommendations. Users acquire point-of-interest, movies, products, and various online resources as suggestions. Classical collaborative filtering algorithms are shown to be satisfactory in a variety of recommendation activities processes, but cannot often capture complicated interactions between item and user, along with sparsity and cold start constraints. Hence it becomes a surge to apply a deep learning-based recommender model owing to its dynamic modeling potential and sustained success in other fields of application. In this work, a trust-based attentive contextual denoising autoencoder (TACDA) for enhanced Top-N context-aware recommendation is proposed. Specifically, the TCADA model takes the sparse preference of the user that is integrated with trust data as input into the autoencoder to prevail over the cold start and sparsity obstacle and efficiently accumulates the context condition into the model via attention framework. Thereby, the attention technique is used to encode context features into a latent space of the user's trust data that is integrated with their preferences, which interconnects personalized context circumstances with the active user's choice to deliver recommendations suited to that active user. Experiments conducted on Epinions, Caio, and LibraryThing datasets make it obvious the efficiency of the TACDA model persistently outperforms the state-of-the-art methods.


Introduction
The demand for recommender systems has arisen as a result of the information surplus created through the huge volume of data.The recommender system aims to forecast a user's unknown likes and dislikes depending on that user's known interests in specific products.The traditional approach for recommender systems [1], like content-based and collaborative filtering, seems to be useful and provided a rational rate of accuracy in the past.These approaches, on the other hand, are unable to describe extensive nonlinear correlations that frequently facilitate useritem interactions.With the advancement of deep learning [2,3], recommendation engines now use a variety of approaches, that are superior to classical algorithms appropriate to their potential to discover feature delineation from scratch.Existing research in this area is still restricted, as it could not take advantage of context data that is abundant in real-world circumstances.Context adds extra details to the interaction between the user and the item, thereby improving the suggestion quality [4].The attention technique [5] makes it simple to include context in this technique.In addition to this, the rating given by each user to the items in the database is very sparse making it tough to find the correlation between the user and the item which leads to meager recommendations.To overcome these issues, the user's demographic details are obtained through social media websites else the site's sign-up form, is intended to utilize a hybrid strategy to recommend to a new user, combining a collaborative filtering approach with a demographic suggesting technique.
Autoencoders [6] are deep fully connected neural networks susceptible to acquire codings, which are latent abstractions of inputs.The code part of the autoencoder is often substantially less dimensional than the actual input.In the Denoising AutoEncoder(DAE), the input layer is added with noise, and the training of the network is performed to restore the actual input at the output layer to avoid overfitting.Since the autoencoder's latent layer captures the inherent depiction of the data, it permits it suited for deployment in the recommendation process.This allows the model to take sparse input of the user-item rating vector and learns the latent feature associated with it in the code part and reconstruct the output layer as a dense vector of predicted rating value [7,8].Hence denoising autoencoder is employed as a basic architectural component in the TACDA model where the context details and trust information provides adds dimensions to real-life scenarios.
Simultaneously, there must be attention to the backdrop scenario (i.e.) context data that entails offering a suggestion.Location, purchasing time of the product, partner, customer sentiment, and so on are all examples of context.When the context changes, the user's opinions on the item may change.Hence the conventional process of recommendation takes the context detail into account by contextual modeling or post or pre-filtering techniques to enhance the effectiveness of the recommendation.To provide intensive suggestions to the user according to the current context of the user, the TACDA model accumulates contextual data through the attention framework [9].Its ultimate goal is just to emphasize or concentrate attention on a particular segment of data.It accepts a specific rating as well as the context that concerns it where the output is a depiction of the rating that focuses on knowledge related to the specified context.This attention framework is used for two purposes: initially, to offer accurate analysis of a higher-dimensional input by just examining its subsets.Finally, concentrate on certain sections of the input that are relevant.Hence Trust-based Attentive Contextual Denoising Autoencoder (TACDA) model employs the attention framework to incorporate the contextual features into the coding part of the user's rating.
However, the accuracy of recommendation provided by the autoencoder that incorporates context by attention framework can be enhanced by integrating the trust information [10] available in the social network that eliminates the sparsity and cold-start problem.Numerous investigations were recommended to add knowledge of trust into the model of recommendation [11,12], depending upon the fact that users' choices are habitually prompted primarily through their friends [13].These shreds of evidence imply that trust relationships are effective in helping to model user choice and improve recommendation efficacy.Motivated by these factors, a customized recommendation model TACDA, an autoencoder framework is proposed which is enhanced by a contextual attention mechanism.It resolves the sparsity and cold start issues by incorporating additional social trust factors that improve the Top-N recommendation quality.
The following are the primary contributions of this paper: • A Trust-based Attentive Contextual Denoising Autoencoder (TACDA) is the first approach that generates the most accurate recommendation corresponding to the current context associated with the active user by utilizing context data through the attention framework in the autoencoder.• To overcome the sparsity and cold start issue, both explicit values of trust and the implicit trust data computed from the similarity value are accumulated into the TACDA model to effectively enhance the Top-N recommendation accuracy.• Employing authentic datasets from Caio, Epinions, and LibraryThing datasets on recommendation tasks, we completely investigated our TACDA model.The findings show that the TACDA model is more constructive than existing state-of-the-art models.

Related Work
Research in recommender systems considers a variety of approaches, algorithms, and their combinations including collaborative filtering approaches, hybrid approaches, knowledgebased approaches, content-based approaches, and context-aware approaches [1].When a large amount of information about the items is available, identifying similar items based on characteristics can allow them to be recommended [14].Enhancements such as detailed analysis of users and items, the accumulation of contextual information, and the facilitation of multi-criteria ratings have increased the applicability of content-based recommender systems to a wide range of applications [1,20].Also, identifying explicit content feature extraction in the target data file benefits improving the accuracy of content-based recommendations [15].Collaborative filtering-based recommendations using patterns on user information, without requesting intrinsic information about the target is widely explored in a variety of recommendations in social networks, e-commerce, and books [16,17].Su and Khoshgoftaar in a study provide a comprehensive review in which they explore all stat-of-art approaches available for Collaborative filtering-based recommendation systems to summarize their advantages and limitations [54].Due to the scarcity of data [18,19] in the input ratings, the step of finding similar users quite often fails in Collaborative Filtering-based Recommender Systems.Combinations of different approaches, such as collaborative filtering recommendations and content-based recommendations, are also been investigated to overcome the shortcomings of each technique and obtain improved recommendations [56].In place of the similarity weight, a trust metric can be used to estimate and propagate trust across the network.Abdul-Rahman et al. [31] propose a real-world trust model that learns social trust characteristics based on the user's social connections with people who share similar preferences.This model relies on trust to overcome the complexities and uncertainties of social interactions.The effectiveness of Trust-aware recommender systems in overcoming the traditional problems with recommender systems is demonstrated by Massa et al. [32] through the proposed Trust-aware recommender system architecture.The performance of the proposed recommender systems proves that collecting a few trust statements is more useful than collecting an equivalent count of ratings.Trust-aware recommender systems [21] help to address the major challenge of embracing trust in recommender systems.It provides a different perspective on user preferences than item scores and improves the accuracy of recommendations.The network's trust and user information can be used to guess anonymous ratings and provide personalized recommendations.According to studies, trust-aware recommender systems are more resistant to shilling attacks and can be used to generate recommendations for new users.The TrustSVD [11] approach is extensively used to provide accurate recommendations using the single value decomposition method, which is enhanced by injecting explicit trust data into the model.Evolutionary algorithms under artificial intelligence algorithms are investigated for optimization in recommender systems.An ant colony-based trust-aware recommendation model named TCFACO [23] uses ant colony optimization (ACO) to choose the set of effective users and their related weights.In TrustMF [12] approach trust information is employed in matrix factorization to improve the accuracy of the model but fails to find the non-linear relationship between product and person.Deep learning and machine learning algorithms, which are elements of artificial intelligence algorithms that investigate the non-linear relationship between person and product, are being investigated for the development of knowledge-based recommender systems [52,53].Zhao et al. proposed an e-commerce recommender system that automatically sets price discounts while recommending a product using Linear Regression [24].A recommendation system with improved scalability to help knowledge workers in determining valuable new content is developed by Verma et al. [25] using Support Vector Machines (SVM) and Active Features-based Model Selection (AFMS).A deep neural network (DNN) guided by examplesrules is proposed by Alashkar et al. [26] for a fully automated makeup recommendation system.This fully automatic recommendation system provided effective recommendations in both statistical and perceptual ways with high synthesis accuracy.A modified DNN, which can proficiently learn the nonlinear relations among persons and products with reduced complexity, is proposed by Chen et al. [27] for an industrial-scale recommender system.Convolution Neural Networks (CNNs) are proficient in extracting both global and local features from diverse data such as visual data and textual data.Elkahky et al. [28] propose a recommendation system that performs Cross Domain User Modeling by mapping items and users to a hidden space using CNN, wherever the correlation between user and product is expanded.A multi-layer neural network architecture called ONCF is proposed by The et al. [29] to perform collaborative filtering.Above the embedding layer, the proposed model uses an exterior composition to characterize the pairwise correlations to result in a two-dimensional interaction map in which the higher-order correlations are learned using a CNN.A conversational recommender system, a two-stage approach proposed by Christakopoulou et al. [30] makes use of a Recurrent neural network (RNN) to offer unified interactions with users for effective extraction of user preferences to produce better recommendations.RNNs allow the recommender framework to analyze the sequential emergence and temporal dynamics of data.
An autoencoder [7,8] is a class of deep learning algorithms that uses unsupervised learning to learn data encodings.It uses non-linear transformations for dimensionality reduction [55] and is widely explored in the recommender framework to learn lower-level features and impute the matrix of interaction in the reconstruction layer [7].Zang et al. [8] proposed Denoising Autoencoder for collaborative filtering (CDAE) to use an exceptional weight matrix for every user which has a prominent impact on the model's execution.CDAE's parameters are learned by lessening the reconstruction error and are widely made used for ranking the predictions.Advanced autoencoder-based architectures that showed improved performance than CDAE include Muli-Variational Autoencoder and Multi-Denoising Autoencoder by Liang et al. [33] uses Bayesian inference approach to estimate parameters and exhibited improved results than what was produced using likelihood functions.
Trust-based recommender systems using Autoencoder named AutoTrustRec proposed by Bathla et al. [34] extract non-linear relations between the rating and trust information to generate recommendations.Pan et al. [35] modeled a new framework for Correlative Denoising Autoencoder that used the structure of three autoencoders to learn features under multiple domains, rater, trustor, and trustee to generate recommendations.The requirement of high computing power is the drawback of this approach.Many researchers have used DAE to solve spare problem issues in a recommendation system.When compared to traditional DAE models, such as the matrix factorization method [42], approaches that focus on how to learn user preferences directly from users' rating information [43] illustrate significant improvement.Pan et al. [44] developed a correlative DAE (CoDAE) that takes into account correlations among users with various roles when learning robust interpretations from sparse ratings and social network inputs while performing recommendations.In an approach proposed by Bathla et al. [45], recommendation accuracy is improved by feeding direct and indirect trust values into a neural network via a shared layer in an autoencoder.The Deep-MSR recommender system by Yengikand et al. [46], utilizes an MLP and stacked autoencoder to overcome the partial expressiveness of the Dot product function in describing various impacts of different latent factors present in the user-item interaction matrix.To address issues such as data dispersion and cold start in recommendation systems, Kashani et al. [47] propose combining trust information with negative matrix decomposition.Deng et al. [48] address this using a recommendation system that follows two-phase learning.In this case, a Deep autoencoder is used in the first learning phase to learn the initial values of the MF model for the second learning phase to offer a higher quality of recommendations.Zhang et al. [49] propose a two-layer neighbor selection scheme to address recommendation system challenges while improving the effectiveness of neighbor selections by identifying the most prominent and trustable neighbors.Wan et al. [50] propose a recommendation system that addresses the matrix initialization and cold start challenges in recommendation systems by employing a deep marginalized Denoising Autoencoder.The model first employs linear DMF and nonlinear DMF to learn features from the user and item rating matrix to improve initializing accuracy.Pan et al. [51] propose an approach for learning low-and high-level features from social connections using a Sparse Stacked Denoising Autoencoder and a matrix factorization technique to confront data sparsity and imbalance problems in recommender systems for social networks.In TDAE [22] to enhance the recommendation accuracy by deploying deep neural networks, here denoising autoencoder-based rating prediction incorporates the trust data into the model.IS_AE [10] performs item splitting for gaining the knowledge of context into the autoencoder system with the accurate recommendation but results in increased computation time.
Context-aware recommender systems that use Contextual Modeling, Contextual Prefiltering, and Contextual Post-filtering, can make accurate recommendations using particular contextual conditions of the user [36].Time-aware recommender systems are a subset of Context-aware recommender systems that focus on using time as contextual information to identify overlapping networks among users and aid in reducing the effect of sparsity [37].Xia et al. [38] propose attention-based neural collaboration filtering for recommender systems, where long-distance dependent information and key information are provided as attention to the neural network to handle the cold start and data sparsity issues.Personalized recommendations are obtained using an autoencoder with an attention mechanism, proposed in [9] where it gathers context information into the model but lacks accuracy in sparsity and cold start problems.

Stacked Denoising Autoencoder
Autoencoders [7], an unsupervised learning approach utilized for learning the firm latent representation of features from the partially corrupted layer of input which is regenerated in the output layer from the discovered hidden features.The autoencoder has two functions: a decoder and an encoder.
The encoder encodes the input values to a hidden representation employing an activation function as Eq. ( 1): The decoder uses the activation function that decodes the hidden representation to its output in which the network is trained to equate the input as in Eq. ( 2) where ρ(.) indicates the method of activation, input data is represented by x, (a 1 , a 2 ) and (V 1 , V 2 ) denotes the biases and weights respectively.In Stacked autoencoders are introduced, where multiple layers of the encoder are connected preceding the decoding layers.This approach allows the encoder networks' last layer to identify the latent representation in the lowest dimension.Experimentally it improves the overall quality of the model.Here each layer's parameters are trained individually, whereas the rest of its networks' parameters stay unaltered.Emerging research in deep learning promotes the stacking of trained sparse autoencoders to initiate deep neural networks.
For predicting the rating, data sparsity is an issue that disturbs the prediction accuracy.Some of the existing works addressed the sparse inputs by calculating the missing values.The inputs are sparsified by the autoencoder.The edges connecting the hidden and the input layer are assigned weights when the neurons are randomly inhibited in the layer of input during the forward propagation, resulting in sparse inputs.Similarly, the output layer neurons are initialized to zero to restrict the edges and their associated weights among the hidden layer and the output layer during back-propagation.The outcome of sparsification is calculated by tuning the α value of the loss function.
The actual method of restraining the edges of weights among the layer of input neurons and layer of hidden neurons is to alter the absent values to zero.Hence to stop the process of autoencoder as of recurring zero, a function for observed loss is used which neglects the unidentified loss values.Here, missing data are unidentified data in the forward propagation technique; hence the error throughout the backpropagation needed to be trained the model using their construction errors is unnoticed.This operation is alike to removing the neurons having missing values in the works proposed by [7].The above method and the masking noise are integrated with the autoencoder.The sparsification method for the inputs is explained in Fig. 1.The deep learning model used supervised learning for implementing the process.The original information is updated as zero by adding a mask of noise values.The value truly occurs and it has been expected as missing, therefore the autoencoder can be enforced to study the missing data.
Here, the input x is measured as two components, one is the x i that is attained by the addition of noise on x i and the other component is x i which has no change.The loss function has been altered to highlight the denoising component in the model.The loss function encompasses 2 components depending on a fine-tuning parameter α as in Eq. 3.
where α represents the parameter for tuning used in the squared error for denoising, (1 − α) is the fine-tuning parameter used in the squared error for reconstruction, y ∈ R M represents the corrupted component of y, B(y ) represents the set of corrupted values of y, M(y) is the set of the unaltered values of y, y i and y i represent the ith outputs in the network.

Attention Mechanism
The attention technique is a prevalent prototype in current years and it is extensively used in various responsibilities like computer vision, recommendation, and information retrieval.
The key belief is to duplicate the attention technique of humans.In our brain when sensory signals are processed, it focuses on convinced areas.Human beings can identify things better by focusing on these subjects.According to this faster screening skill of humans identifies the competence and accuracy of the attention technique.The attention mechanism based on deep learning is considered a matrix of weights using similar scales as in input value.In the commencement, every weight resembles input data; and the product of the weight and input value obtains the result.Lastly, the degree of attention needed for the input data hinges on the value of the weights.Also, the weights are comparative to the degree of attention, which means that a bigger weight creates robust attention.The key idea of attention is to study how to allocate attentive weights for a set of features: Higher values for the weights indicate that certain features are extremely informative for the work.
The attention technique was presented by [41], to report the bottleneck issue that rises when a fixed-length encoding vector is used, and the decoder is having partial contact with the input data.This is intended to create some issues in complex and lengthy sequences, as the dimensionality is expected to be the same for the shorter and simple sequences.
The attention technique has various steps like calculations of the alignment scores, the weights, and the context vector: 1. Scores from alignment: The encoded hidden states h i are taken and also the output from the previous decoder S n-1 to calculate score P n,i , which shows that the elements of the input are aligning with the elements of the output at a particular position n.The model is indicated by a(.) that is constructed using a feed-forward network 2. Weights: The weights are calculated using softmax on the alignment scores 3. Context vector: The context vector denoted by C n is given to the decoder.It is calculated from the N hidden states from the encoder.
The attention mechanism is used in a lot of recommendation works.The interaction between the inputs is considered and the network in this attention mechanism is formed by a couple of interaction layers and grouped with the interaction vector.The self-attention technique has been projected by Google which substitutes RNN and CNN.It finds the parallelization of models that are trained using actual attention which results in a good performance for the translation of machines.The attention technique is outstanding while considering NLP which increases the effectiveness while extracting the features.

Proposed Methodology
The architecture of the Trust-based Attentive Contextual Denoising Autoencoder (TACDA) is depicted in Fig. 2.
Firstly, the implicit trust values are computed to overcome the sparsity problem in the available explicit trust values.Secondly, the proposed model TACDA is presented that integrates user preferences, implicit and explicit trust values, and user and item contextual attributes that enhance the accuracy of Top-N context-aware recommendation.

Problem Definition
Considering the transactional dataset containing a collection of users U = {1, 2, ..., u} with a collection of items, I = {1, 2, ..., n}, and a users' log of the prior various choices of products O = (u, i, r ui , et u , d ui ) where r ui signifies the rating provided by a user u for the item i on a specific date d ui , as well as explicit trust values, exist between users et u respectively.The main objective is to offer a range of items N to every other active user u q focusing on their present context scenario associated with the user uc k and the item ic m that has been extracted from the date of purchase d u q i is given as an attention layer input that will optimize her or his contentment.In many circumstances, the data includes unobserved and missing preferences that are indicated by Õ.Consider O u q denoting the range of item choices in the range of training data for a particular user u q , as well as the user's u q unobserved choices Õu q .To a user u q , the candidates to be offered are the items in Õu q .The recommender's purpose is to select a set of items that are contextually related from the candidate set that has the greatest estimated score for each user u q .

Determining Implicit Trust
Recommendation accuracy can be significantly increased by incorporating trust values into the system [10].Generally, the trust data that exist between users that is available explicitly in the dataset is sparse.As a result, manipulating users' implicit trust relations has become increasingly popular and challenging.A familiar method to prevail over this issue is to incorporate similarity metrics to measure the associations between users.In [39] the trust value relating to the users is manipulated by Pearson's correlation coefficient (PCC), resulting in an incoherent overall similarity degree of 1 although a single co-rating exists between those users.As a result, when calculating the correlation, [40] acknowledged this problem and normalized the correlation to [0, 1]: where d denotes the co-rating numbers among u a and u b .Likewise, [40] recommended a correlation threshold ϑ to manipulate trust values that are calculated implicitly, where it is utilized only if the ϑ value is minimal when compared to the correlated value: here t u a ,u b signifies the binary representation of trust value that is obtained implicitly with user u a and u b .In this model, the available trust values are expanded by the additional trust values inferred implicitly from the ratings given by the user which results in a matrix of dense trust values.The essential point is the trust value computed implicitly from the rating vectors of 2 distinct users, where the significance ϑ is less, it is feasible to mine added trust relationships amid the users implicitly.The optimal ϑ value is acquired by varying different values of ϑ TACDA and its corresponding results in experiments were examined.

Trust-Based Attentive Contextual Denoising Autoencoder
The TACDA model presumes that there are n items and u users.It takes the rating vector R Uq ∈ R n given by the user U q on the entire item in the dataset together with the explicit trust vector E T ∈ R x and implicit trust vector I T ∈ R y as the three vectors of input to the model.The rating vector R Uq is corrupted using the drop-out/mask-out method [10] to acquire the corrupted vector RUq .The vector RUq with E T and I T vectors is given into the TACDA model that produces the latent vector h z Uq ∈ R l by the encoding function ω(•).Moreover, the associations among trusted values and the preference of the user are substantially nonlinear with the distinct distribution.To address this constraint and find more features from trust data, both the values of trust are fed into the TACDA models' all input and hidden layers, to enhance the top-n prediction accuracy [10] as shown in Eq. (9).
where h z Uq ∈ R l is the learned latent representation of the model, V ∈ R p×(n+x+y) and V ∈ R l×p represents the matrices of weight, r q ∈ R n represents the preference of the user in R, et q ∈ R y and it q ∈ R x represents the trust values of u q ,g ∈ R p and g ∈ R l represent vectors of biases, z represents the last layer of the encoder (i.e.) the latent code part of an autoencoder and ω indicates ReLU function since it operates fine on sparse rating vector.

The latent code vector h z Uq
RUq is fed as an input vector to the attention framework, where the features of context associated with the user and the item are provided in the model.The primary aim of the attention framework is to observe the input depiction focused on the offered specific contextual condition.It is important to note that the proposed TACDA is adaptable and adequate as several contextual features as required.Though, in this model, only two contextual features (k and m) are considered for simplicity.This attention framework [9] employs a weighted context of the user uc k and the context of the item ic m to the h z Uq latent layer for any attributes of contextual condition k and m with the activation method ρ(•) as represented in Eq. ( 10) (10) where V h ∈R l×l is a weight matrix, where l denotes the size of the attention layer, which is equal to the size of the hidden layer.V k ∈R l×|k| and V m ∈R l×|m| are matrices of weight corresponding to the size of contextual conditions, h z Uq RUq which is the latent layer output.
The best outcome of the attention framework is achieved when the function of activation is set to tanh.
The outcome of the attention framework t U q RUq is given as input into the softmax layer.At last, the outcome of the softmax layer is integrated with the outcome of the latent layer h z Uq RUq to perform element-wise multiplication.Thereby it generates the attention frameworks concluding output that is represented as A U q RUq .
where t U q RUq is the outcome of the activation function of the attention framework and RUq is the outcome of the latent layer.The function definition of the softmax layer is The attention framework works by applying a biased mean into the latent layer depiction, to the weight depicting the significance focused upon the contextual condition.A decoding function φ(•) is used to reconstruct the original form of the input's intrinsic latent depiction along with the provided contextual condition.
where RUq ∈ n , V ∈R n×l , and g ∈ R n represent the matrices of weight and biases as the same in the dimension of the input layer.The decoding activation function φ(•) is set to the sigmoid function.Since it restricts the limit of the output to the range between 0-1.Hence it provides users with the likelihood correlated upon every item in the reconstructed layer, which is utilized for customized recommendation ranking.The model's parameters are learned by significantly reducing the mean squared error concerning the denoised input layer RUq and the reconstructed output layer RUq .min

Online-Recommendation Phase
The TACDA model follows the following Algorithm (1) for generating the recommendations to the active user Uq.

Description of the Dataset
Employing authentic publicly available datasets like Ciao, Epinions, and LibraryThing [10], the developed TACDA model is examined against the present state-of-the-art technique.
The information in the dataset given in Table 1 is imported from Ciao.com, Epinions.comwebsites, and a book review website librarything.com.The users of this website tend to give a rating to the product purchased ranging from 1 to 5 and establish a trusting relationship with the rest of the users by choice.Thereby a network of social trust is developed in a form of binary representation, where 1 represents trust and 0 represents distrust linkage between two users accordingly.According to the specified Date field available in the dataset two context dimensions namely, Time and Season are extracted.Here the context conditions of time feature such as afternoon, night, evening, and morning, is considered as the user contextual features uc k.Similarly, the context condition of the season such as winter, rainy, summer, and fall are considered as the item's contextual features ic m .

Evaluation Metric
To show that integrating reviews and ratings improves top-N suggestions generated with highly relevant products, decision-support knowledge retrieval metrics such as recall (fraction of the entire amount of items relevant that are retrieved) and precision (fraction of items relevant among the received items), are utilized.Here, the user is provided N items to match his preference from unrated training data.Hence, in the TACDA model, the Mean Average Precision (MAP) is used to determine the efficiency of the top-N item suggestion.
Precision and Recall: For a specific top-N suggested item K u , where K N ,u items are considered by the user from the test data.

Mean Average Precision (MAP):
Average precision (AP), is a ranking metric that gives weight to items that are precisely suggested.
here rel(f), an indicator method outputs 1 when a product at rank f was considered, otherwise sets to zero and Precision@f indicates K i 's value of precision@f in the top N prediction.Subsequently, (MAP@N) denotes the mean of the entire AP scores of users.
Discounted Cumulative Gain (DCG@N): For each consumer, it's explained as: The DCG normalized on the ideal iDCG@N is mentioned as Normalized Discounted Cumulative Gain at N (NDCG@N), which takes into account the ideally suggested items' ranking.

Experimental Setup
The dataset is split for experimentation by 70% in training, 20% in cross-validation, and 10% in the test set.The metrics of evaluation comprise R@10, p@10, M AP@10, N DCG@10 since accurately evaluating the Top-N recommendation outcome.The proposed system TACDA is developed using the TensorFlow library's Keras API in Python.In this dataset, all of the ratings are scaled between -1 to + 1.In the experiments conducted on each dataset, the hyperparameters for all comparative approaches are carefully tweaked by the GridSearchCV methodology to ensure that every approach gives the best value for proper comparisons.The robustness of the system to the epoch, noise variations, neuron number in the latent space, the rate of learning, and the optimization procedure are all examined consequently.Since the variations in noise and count of hidden neuron size affect the results, more experiments were carried out to establish the ideal significance of corruption ratio and latent layer size, which has been described in 5.2.1.In the training process, the TACDA model was observed to converge when the epoch was set to 300.It is also examined with different learning rates λ {0.015, 0.025, 0.001, 0.002, 0.01,0.02)and determined to be optimum when λ = 0.001.Finally, the framework is refined by employing Adam Optimizer having a batch range of 40, and regularized by setting weight decay l 2 = 0.0001.
For accurate comparisons, the TACDA model is assessed against the baseline methodology in each of the key domains: Neural Network models (U-AutoRec, CDAE, AutoTrustRec, TDAE, CoDAE, ACDA, IS_TDAE), Model-rely techniques (TrustMF, TrustSVD).To conclude the model-based techniques, Librec, a recommender library is employed.For TrustMF the regularization parameter λ is set to 0.001, the degree of trust regularization parameter λ t is controlled by setting λ t = 1, and the number of latent features d is set to 10 across all the datasets.In TrustSVD the parameters are set to λ = 0.8, λ t = 0.9 for optimal results.Finally, for all the neural network-based baseline model implementations, the parameter values are set to be the same as the TACDA model (to the possible extent).The existing models are illustrated as follows.
1. TrustSVD [11]: This approach is extensively used to provide accurate recommendations using the single value decomposition method, which is enhanced by injecting explicit trust data into the model.2. TrustMF [12]: This approach is simple and employs trust information in matrix factorization to improve the accuracy of the model.3. U-AutoRec [7]: Collaborative DAE considers users' k-hot encoded vector of rating values as input specified to the items.4. CDAE [8]: For Top-N recommendation, Collaborative DAE adds on the auxiliary data in the means of users' latent factor as input together with the rating value.5. AutoTrustRec [34]: This approach generates recommendations by extracting non-linear relations between the rating and trust information.6. CoDAE [35]: This approach modeled a new framework for Correlative Denoising Autoencoder that used the structure of three autoencoders to learn features under multiple domains, rater, trustor, and trustee-to generated recommendations 7. TDAE [22]: To enhance the recommendation accuracy, here denoising autoencoder-based rating prediction incorporates the trust data into the model.8. ACDA [9]: This approach incorporates both the contextual features associated with the user and the item into the autoencoder by attention mechanism.9. IS_TDAE [10]: This approach generates context-based fictitious items via the method of item splitting that incorporates additional trust data into the model.

Impact of Latent Layer Size and Corruption Ratio
The accuracy of the framework is most influenced by the latent space representation, which is correlated to the number of neurons in that layer.To conclude the optimal count of neurons in the latent space, the TACDA approach is assessed with various values of k.The experimental findings determining the corruption ratio are shown in Fig. 3 and the experimental findings determining the latent unit size are shown in Fig. 4. We also investigated various corruption ratio (α) values spanning from 0.1 to 1.0 in 0.2 intervals.The outcomes, as shown in Fig. 3, demonstrate that when the corruption ratio rises, the efficiency declines.As a result, we fixed the corruption ratio α = 0.5 in Caio, α = 0.7 Epinions, and α = 0.6 LibraryThing datasets.The graphs in Fig. 4 show that once k = 200, the model's efficiency stays consistent, and thus no significant progress in efficiency beyond that level at the intensity of raise in time of training.

Results and Discussion
The experimental outcome of the TACDA method against the respective state-of-the-art method is depicted in Tables 2 for the Caio dataset, Table 3 for the Epinions dataset, and Table 4 for the LibraryThing dataset with the finest results noticeable in boldface.Figure 5, 6 and 7 depicts the NDCG@N and MAP@N values on the 3 respective datasets as indicated in Table 1.From the outcome, it is observed that the impact of the NDCG, MAP, precision, and recall scores were consistently trustworthy throughout the entire dataset.On all datasets, the TACDA model performed better, representing its suitability for Top-N recommendation processes.
Tables 2, 3 and 4 show that the TACDA model outperforms the others by at least 15.6% because it can acquire deeper cognitive properties with its non-linear interaction mechanism and can more effectively accumulate attributes from ratings, trust, and context data.Here the autoencoder consistently assimilates both implicit and explicit trust value into the rating vector to address sparsity and cold start issues in the system, and it provides context-based recommendations by accumulating the contextual data via the attention technique.IS_TDAE outperforms 6.3% of other trust-based models except for TACDA by generating contextual recommendations via the item splitting process but the time taken for generating recommendations is comparatively high due to the increase in the size of the fictitious item that is fed  TrustMF and TrustSVD are the approaches of reducing the dimensions, outperform other approaches in specific instances of computational speed although result in poor accuracy, according to the efficiency measure in terms of average training length.As user contextual features and item contextual features are incorporated in the attention mechanism during the training phase of the TACDA model together with trust values, the Top-N recommendation time is relatively high amid greater MAP@N and NDCG@N accuracy values with accurate contextual circumstances based Top-N item suggestion generation related to the other systems.U-AutoRec, CDAE, AutoTrustRec, TDAE, and ACDA have an average outcome in aspects of accuracy and execution time.The training period of IS_TDAE is relatively low while the accuracy is low when compared to the TACDA model.
Finally, it is demonstrated that our proposed TACDA model performs better accuracy in recommending Top-N items that accurately satisfy the active user's present context situation with firm implicit and explicit trust data incorporated with attentive contextual features associated with both the user and the item.

Ablation Study
To extensively examine the effectiveness of the proposed TACDA model focusing on the contributions of the different components, an ablation investigation is carried out.Therefore two distinct model variants have been implemented where each data set's experimental parameters are identical.These different variants are compared with the Sect. 3 specified approach with its default settings.Figure 8 displays the impact of ablation across all datasets.
• Default TACDA approach: As demonstrated in Sect.3, the Top-N recommendations in this framework make use of the entire features components of the existing TACDA model.• Without Context feature: To suggest recommendations, this model variant cannot integrate the contextual features extracted from the date of purchase into the latent space of the user's preference via the attention mechanism.Therefore, in the online recommendation phase, the items are recommended to the active user without considering the exact contextual circumstance of the user.This is done exactly to measure the accuracy of the Top-N recommendations concerned with the contextual attributes that focus on the user-specific significant context.Hence, considering the integration of contextual features through the Fig. 8 MAP@10 results of the model variants on all the datasets attention method in the proposed TACDA model enhances recommendation accuracy by at least 26% on all datasets by forecasting and offering the precise item based on the active user's context circumstances.• Without Trust values: In this variant, the explicit and implicit trust values are not incorporated into both the hidden and the input layer that attempts to merge these two diverse sources of information and create more accurate user semantic representations.Utilizing this variant the recommendation accuracy is further degraded due to the sparsity in the rating data.Here, primarily the rating score in the output prediction layer is the user-item interaction, which lowers the model's performance by at least 15% on all datasets while comparing the default TACDA technique, as shown in Fig. 8.
Owing to space constraints, we have only presented the MAP@10 outcomes for all datasets in Fig. 8.According to Fig. 8, the proposed TACDA approach performs better at selecting the right item according to the context condition of the user when contextual features and trust values are used utilized.Thus, to cope with the cold start problem, the model TACDA incorporates the users' trust information relying on sparse ratings to augment each user's features.The results obtained demonstrate that the model TACDA can offer realistic Top-N recommendations to the user.

Conclusion and Future Enhancement
Due to a lack of sufficient ratings and contextual information associated with the user or item, collaborative filtering-based recommendation systems that use previous ratings of users to estimate their future preferences have underperformed.Deep learning algorithms that can perform representation learning on user data can proficiently overcome this by providing better recommendations.In this paper, an Autoencoders based efficient deep learning framework named TACDA is proposed for providing enhanced Top-N recommendation that accurately suits the active users' context situation utilizing autoencoder.Here the sparse rating data is incorporated with trust data into the TACDA model to overcome sparsity and cold start issues where contextual information is being discovered by augmenting user and item context data via attention framework.The proposed work is extensively evaluated on a wide range of datasets.The evaluation results demonstrate that the proposed approach outperforms the state-of-the-art approaches in terms of obtaining more accurate recommendations.
This investigation can be expanded in a variety of ways, and we intend to perform it in future tasks.First, different forms of loss functions will be analyzed, such as pairwise and listwise losses, which are effective for ranking operations, especially when dealing with inherent user preferences.Secondly, other features, such as textual reviews and productrelated specifications, might be employed as auxiliary or side information.

Fig. 1
Fig.1The technique of sparsifying inputs.The rating matrix for the user item is used to create the input vector.The missing data in the vector are changed to zero, and also masking noise is used to corrupt input.Before performing backpropagation, the errors occurring because of missing values are changed to zero.Here, denoising errors are multiplied by weight by α, and therefore the reconstruction errors are multiplied with the weight(1 -α)

Fig. 2
Fig. 2 Framework of the proposed TACDA model

Fig. 4
Fig. 4 Determination of latent unit size

Table 1
Statistics of data

Table 2
Performance evaluation results for Caio data

Table 3
Performance evaluation results for Epinions data

Table 4
Performance evaluation results for LibraryThing dataset