An adaptive RNN algorithm to detect shilling attacks for online products in hybrid recommender system

1.1 Background of study

The recommender system (RS) gives recommendations to help potential buyers for choosing complex information and different products. RSs are becoming more and more indispensable with increasing number of choices available online [1]. Many e-commerce platforms such as JD.com and Amazon have launched many online services in recent years. These platforms are mostly used by consumers to obtain and share different products through online reviews [2]. To purchase products or services, the online reviews have become a significant data source for consumer. For example, among 2010–2016 the amount of customers reading online reviews enhanced to 91 from 71% according to the consumer market report [3]. The consumer read the online reviews to create the top purchase decision, when consumers decide to buy the products or services. However, it is very tough for buyers to read the entire online reviews within a limited time. From the online reviews, useful information are obtained which help to select the desirable products and rate these products [4]. Therefore, it is important for researchers to focus on techniques and methods for rating products depending upon online reviews to support purchase decisions of consumers [5].

Nowadays, depending upon the online reviews many schemes for rating products include the data fusion procedure from the existing research results. The sentiment orientation of each online review can be automatically identified and other products ranking can be explained by using this method [6,7,8]. Collaborative filtering (CF) and content based (CB) are the two algorithms mainly used in the RS. Nowadays, in the recommendation field the CF is very popular and it plays a vital role. In RSs there are number of hazard attacks. Depending on the purpose of attackers, the attacks can cause various damages to vulnerable systems. The attacks directly affect the performance of RSs [9]. In RSs, the biased user’s representation not only makes the system vulnerable to attacks but also leads to improper recommendations to its users [10]. The large number of attacks results in system collapse under such conditions. It is very tough to stop unscrupulous users from injecting fake profile into a system [11]. The attack profiles must be detected and removed to ensure the trustworthiness of RS among its users. Two attacks such as push and nuke are presented in this article. The shilling attacks can be categorized as push and nuke attacks. To promote or demote the predictions is the main objective of these attacks which are made for targeted items. The rating of m-dimensional vector is considered by attack profile, where the total amount of products in the system is represented as “m.” Figure 1 depicts the four parts of profile. The null partition “I _ɸ” are those items in the profile with no ratings. Depending on the attack type a rating designed will be given by single target item “I _t” to convert its recommendation; usually this will be both minimum (r _min) and maximum (r _max) possible rating. During the attack for special treatment a few attacks are needed to classify a group of products. This superior set “I _s” typically gets high ratings to make the profiles like user who desire these items. At last, to complete the profile, set of filler items “I _F” is included. This method is used for choosing the products in “I _F” and “I _S” and the ratings given to these things characterize an attack model.

Figure 1

Shilling attack model.

In existing work, the researchers has studied the implementation of algorithm from an area of artificial intelligence (AI) and machine learning (ML) algorithm with RSs [12,13]. The ML algorithms such as Bayes network, clustering, ordinary least square regression, logistic regression, random forest, support vector regression, and k-nearest neighbour are the oldest ML techniques. These algorithms are used to give consumers good recommendations in RSs [14,15,16]. However, due to the number of methods and variations presented in the literature, the ML field does not have a clear classification method and also there is demand of strong ML algorithms and more robust to outperform. As a consequence, when developing an RS, it becomes difficult and confusing to choose an ML algorithm. One of the most robust ML algorithms is Ensemble model. It produces highly accurate results and improves the accuracy and performance [17]. Ensemble can give high accuracy with strong predictive power which may help to select some best products, however, selection of model is really hard and time consuming.

The proposed algorithm uses a hybrid approach in which an ML algorithm is integrated with an optimization algorithm to construct a more powerful and effective prediction model for RS. In this work, adaptive recurrent neural network model is developed by the integration of opinion mining approach with recurrent neural network (RNN) [18] and Bald Eagle Search (BES) optimization algorithm [19]. The two key approaches used – RNN and BES – are discussed in the subsequent subsections.

1.2 Problem statement

In the previous research, some researchers have focussed on sentiment analysis to enhance the accuracy of collaborative recommender system (CFRS). On the other hand, other researchers have used clustering algorithms and used classification approach to detect the shilling profiles (or PN-Attacks) in CFRS. Also, some have used neural network approach to enhance the accuracy of CFRS. The existing research works either focus on enhancing accuracy or to detect the shilling attacks in CFRSs. There is no single algorithm that can do both, that is enhance and detect the attacks. Furthermore, the existing algorithms consider only product ratings to identify attacks. To bridge this gap, in this work, a hybrid approach is used in which RNN is applied on user reviews to detect the PN-Attacks. Also, modified density peak clustering algorithm (MDPCA) is used for clustering to find the most preferred user from each cluster and Gaussian kernel is used to enhance the performance of the proposed model, named as adaptive recurrent neural network (ARNN) model. The BES Optimization Algorithm is used to initialize and update the weights of RNN. The proposed ARNN model is developed for hybrid recommender system (HRS).

1.3 Contributions

The major contributions/highlights of this research are as follows:

1. The pre-processing is used to find and eliminate the unwanted attacks in consumer reviews of Amazon product datasets.

2. The output of the pre-processing is given to the feature extraction by word embedding technique.

3. Developing an MDPCA on consumer reviews of Amazon product datasets to ensure a well-formed cluster.

4. The Gaussian kernel is used to enhance the performance of the density peak clustering algorithm (DPCA) to evaluate the local concentration.

5. An efficient ARNN is used to obtain the maximum preferable user from each cluster. The RNN model parameter values (weights) are initialized through the fitness computation of BES “search agents.”

1.4 Organization of the article

The article is organized as follows. Section 2 discusses the related work. Section 3 discusses the methodology used. Section 4 gives the implementation of proposed model with discussion about the attacks, pre-processing stage, feature extraction by word2vect method, MDPCA technique for clustering, and ARNN method. The performance analysis depicting the results is presented in Section 4. Finally, Section 5 gives the conclusion and future direction.

3.1 RNN

RNN is a type of neural network with feedback connections among nodes. It has an ability to model a dynamic system. RNN has three different layers, i.e. input, output, and hidden layers, as shown in Figure 2. RNN uses the following recursive formula:

where “h _t ” is the hidden layer vector, “W _h” is the weighting matrix, “x _t ” is the input vector, and “y _t ” is the output vector.

Figure 2

Structure of RNN.

Long-term dependency issue is presented in RNN, and the issues of weight matrix and long interval time keep multiplying recurrently with earlier outcomes. This may cause exploding gradient and vanishing gradient issues. To avoid this issue, long short-term memory (LSTM) is used, which will enhance the performance.

In the LSTM, every neuron is a memory cell. Every neuron includes three gates, namely, input gate, forget gate, and output gate. The three gates are discussed as follows:

1. The forget gate F(t) defines unwanted data. At the previous unit “t – 1,” by entering the output “h _t−1” and adding the input “x _t ” with current time “t,” softmax function “s(t)” is given as follows:

   (3) F ( t ) = σ ( W f ⋅ [ h t − 1 , x t ] + b f ) ,

   (4) s ( t ) = ∑ j exp ( z i ) exp ( z j ) .

2. The input gate explains which new data to be recollected in cell state.

   (5) i ( t ) = σ ( W i ⋅ [ h t − 1 , x t ] + b i ) ,

   (6) C ˜ ( t ) = tanh ( W c ⋅ [ h t − 1 , x t ] + b c ) .

   To obtain the updated information the values of “i(t)” and C ˜ ( t ) are multiplied by the sigmoid function that we want to add to the cell state.

   (7) C ( t ) = F ( t ) × C ( t − 1 ) + i ( t ) × C ˜ ( t ) .

3. The output gate explains which data will be output in the cell state. The cell state is first triggered in the “tanh” layer before being multiplied by “o(t).” At time “t” the multiplication result is the output data “h(t)” in the block of LSTM.

The available data are categorized into three non-overlapping sets for the purpose of training, testing, and validation. At first, the LSTM model in consumer reviews of Amazon products dataset is trained. To choose the best parameters as well as the performance in the proposed model the validation is employed. Finally, the same dataset is utilized for testing purpose and the performance and accuracy are verified. The weight and bias value of all the three gates is updated by BES Optimization Algorithm [19]. To validate the co-sequences of every phase of hunting is the main behaviour of bald eagle.

3.2 BES optimization

The main behaviour of the bald eagle is to validate the consequences of every phase of hunting. The fitness function of BES is calculated by:

The hunting behaviour of BES can be classified into three stages, i.e. select, search, and swooping.

1. In the select stage, the BES finds and picks the best area as best bias within the chosen search space.

   (11) b new , i = b best + α × r ( b mean − b i ) ,

   where random number is denoted by “r.”

2. In the search stage, the best position as best weight value for the swoop is mathematically calculated by

   (12) W i , new = W i + y ( i ) × ( W i − W i + 1 ) + x ( i ) × ( W i − W mean ) .

3. In the swooping stage, the bald eagle blows from best weight in the search space and best bias in the best area. Both these are calculated and mathematically illustrated as follows:

3.3 Proposed framework

The proposed framework for HRS is implemented in four phases. First of all, the user data are collected which contain ratings and reviews of various products from dissimilar users. The user data include both genuine and attack profiles. The collected data then enter into Phase 1 and are pre-processed. Dissimilar products are eliminated and feature extraction is applied using Word2Vec technique. In Phase 2, MDPCA is applied to form clusters and the most preferable user from every cluster is obtained. The proposed ARNN model is applied on the reviews and ratings of this most preferable user, in phase 3. Finally, the classification is done in Phase 4. If the rating is found genuine, then it is stored in the database to generate the recommendation otherwise it is discarded considering it to be an attack. The higher rating “5” is labelled as “PUSH ATTACK” (P-Attack) and the lower rating “1” is labelled as “NUKE ATTACK” (N-Attack). The rating is labelled “NORMAL” otherwise (i.e. for ratings “2,” “3,” and “4”). Figure 3 presents the proposed framework for the HRS.

Figure 3

Proposed framework for HRS using the ARNN model.

4.1 Phase 1A: Pre-processing

In the pre-processing [41] stage, products with dissimilar features are eliminated. These dissimilar products are measured as outliers and are eliminated because if the products are rated through significantly smaller number of users, then considering such products into subsequent phases of RS will affect its whole performance. The user reviews of filtered products are then given to the pre-processing unit. The steps involved in pre-processing are removal of stop words, lower case, punctuation, and lemmatization.

1. Stop word removal: Stop words such as articles (a, an, the), prepositions (on, at, in, there, this etc.), conjunctions (for, but, and, or, etc.) are removed from the given reviews. These words are removed because, generally, they do not contribute any significant value to the review.

2. Lower case: All words of the review are converted into lower case. Similar words are then combined and the dimensionality is reduced. For example, word Star will be converted as star and if it appears twice in the sentence, it will be combined as one.

3. Punctuation removal: This is a classic method in data mining and information retrieval. The punctuations are removed from the text. However, the presence of punctuation marks often denotes the existence of some sentiment but it does not contribute mathematical value to the text.

4. Lemmatization: The process of merging many words to one is known as lemmatization. This lemmatization analyses a word morphologically and removes its inflectional ending.

Table 1 shows an example of pre-processing a review from the given dataset.

4.2 Phase 1B: Feature extraction

A machine cannot understand the text directly. To make the text understandable, each word of text is represented in a real-valued vector. This procedure of representation of a word into real-valued vector is called word-embedding. In word-embedding technique, the words with similar meanings have similar representation. TF-IDF is one such technique of word-embedding. This technique is discussed and used in the proposed hybrid deep neural network model for CFRS in the previous chapter. The main drawback of this technique is that it cannot maintain the sequence of words. This problem is handled by another word-embedding technique known as Word2Vec [42]. This technique enables similar words to have similar dimensions and, consequently, helps to bring context. Each sentence in this technique is known as a “Vector.” This means that all individual words of a sentence are treated as one – one feature of a vector. For the given vectors, a vocabulary is created by taking only non-repeated vectors followed by creating the one-hot encoding of the vectors. Consider reviews as follows:

R1: This dish is very spicy and hot

R2: This dish is not spicy and not hot

R3: This dish is yummy and good

Then, vocabulary is created as: Vocab: {This, dish, is, very, spicy, and, hot, not, yummy, good}. One-hot coding of the Vocab is created as shown in Table 2.

In the above example, each word occupies one of the ten dimensions. This means that there is no similarity among the words irrespective of their literal meanings. Word2Vec technique establishes the association between the similar meaning words. Word2Vec is an unsupervised model where the corpus can be passed without giving any label. But, since it is basically an internal neural network, it leverages a supervised classification model to get the embeddings from the corpus. There are two methods in word2vec for word embedding, known ashpen Continuous Bag of Word (cBOW) and Skip-Gram (SGram).

cBOW technique tries to find the target word using context word. The hidden layer is eliminated and the words with the same position are analysed. This way of procedure is known as bag of word. The continuously distributed words are thus named as cBOW. In this method, the context word is given as input “X” and the target words are predicted. Figure 4 represents the cBOW structure.

Figure 4

cBOW structure (Source: https://arxiv.org/pdf/1301.3781.pdf Mikolov et al., 2013).

Another technique is known as SGram. In this technique, context word is searched using the target word. This method tries to find the words in the same sentence. We utilize the present word with hidden layer projection as an input that may calculate the words within the range. This method predicts multiple words from a single given word. In the SGram model, target word is given and the context words are predicted. Consider a review: “The dish is very spicy and hot;” and a context window size of 2, the target word is “spicy,” then the model tries to predict [“very,” “and”], and so on. The words are fed in pairs of (X,Y) where “X” is the input and “Y” is the label. This is done by creating “Positive Input Samples” (PIS) and “Negative Input Samples” (NIS). PIS have training data in the form [(TARGET, CONTEXT), 1], where “TARGET” is the centre word, “CONTEXT” is the surrounding word, and “1” is the label which indicates the relevant pair. NIS has training data in the form [(TARGET, RANDWORD), 0], where “TARGET” is the centre word, “RANDWORD” is the randomly selected word, and “0” is the label which indicates the irrelevant pair. Figure 5 represents the structure of SGram method.

Figure 5

SGram structure (Source: https://arxiv.org/pdf/1301.3781.pdf Mikolov et al., 2013).

In the proposed framework, SGram model of Word2Vec technique is used. The model is implemented in the python platform using “genism” and “pickle” modules. For the given dataset, the total number of epochs is 20. The total number of embeddings is 5,000 with a vocabulary size of 6,408 and a window size of 4.

4.3 Phase 2: MDPCA for clustering

The objective of clustering is to sort the data into groups depending on characteristics similarity. Thus, information from dissimilar clusters is different from each other and information from similar clusters have the same properties. To find the cluster centres is the main idea of DPCA that fulfil two significant assumptions which are given as: (i) by lower density neighbour the cluster centre is bounded, (ii) the centre of cluster is away with a higher local concentration from any other cluster centre point.

To calculate the cluster centres the DPCA uses two measures such as local concentration and distance. Distance among two data points should have been determined by the two measures. For computing the distance between two data points, first DPCA accepts the Euclidean distance. However, the distance can cause several misclassifications if the dataset is inseparable and complex. Thus, we enhance the DPCA which is known as MDPCA. The Gaussian kernel is one of the modified algorithms which is developed to calculate the distance. Moreover, for calculating the local concentration the Gaussian kernel is used in place of cut-off kernel to improve the performance of the proposed method. Thus, the MDPCA allocates each and every residual point to the similar cluster as its nearby neighbour with high density after that cluster centre has been found (Figure 6).

Figure 6

MDPCA.

4.4 Phase 3: ARNN model (using experimental modelling)

The most preferred user from every cluster is passed on to the ARNN model. The dataset is split into training dataset and test dataset. In the training dataset, the user ratings “2,” “3,” and “4” are labelled as “NORMAL”; higher ratings “5” are labelled as “PUSH ATTACK,” and lower ratings “1” are labelled as “NUKE ATTACK.” The ARNN model is applied to both training dataset and test dataset. The results obtained for the test dataset are then compared with the training dataset and attacks are predicted. The implementation of ARNN model uses a hybrid approach by integrating RNN with BES Optimization and works with two procedures. One is Forward Feed and the other is Backward Feed. The implementation is done in the Python platform using the “keras” module. Because of the gradient issue, LSTM RNN is used. During the Forward Feed, using the BES algorithm, the fitness value is computed.

The algorithm enters into the select stage and the best bias is found to be 25–75%. The value of alpha ( α ) is taken as 0.25. Now, the algorithm enters into the search stage and the best weight is picked. The random value required to pick the best weight is computed using random.uniform() function, that is available in python. Finally, both best bias and best weight are computed. After the weights are initialized using the BES Optimization Algorithm, the hidden layers are calculated. At this point, the long-term dependency issue of vanishing gradient and exploding gradient is generated.

LSTM is used to resolve the issue. Forget gate, input gate, and output gate are calculated, and hidden vectors are updated. The back propagation is done using the “Adam” optimizer (available in the “keras” module of python). The stop criterion is set to 0.0001, so as to achieve minimum error. If the stop criterion is reached, then the execution stops and resultant values are passed on to classification phase, else the procedure is iterated again. During the ARNN model implementation, the LSTM dropout is 0.2. Dropout is a technique used to prevent a model from over fitting. Dropout works by randomly setting the outgoing edges of hidden vectors to 0 at each update of the training phase. The activation function used is softmax (given by equation (4)). Figure 7 explains the implementation of ARNN model briefly.

Figure 7

Proposed ARNN model for HRS.

4.5 Phase 4: Classification

At output layer, when the stop criterion is met, finally the computed activation value for the test set is compared with the training set and the ratings are classified as genuine or an attack. The genuine rating is labelled as “NORMAL,” higher rating “5” is labelled as “PUSH ATTACK,” and the lower rating “1” is labelled as “NUKE ATTACK.” The “NORMAL” ratings are stored in the database to generate recommendations, while “PUSH” and “NUKE” attacks are discarded.

The Pseudo code for ARNN model is given as follows:

Pseudo code 1: Implement ARNN

1. Collect user ratings and user reviews of various products from dissimilar users.

2. Apply pre-processing steps (stop word removal, lower case, punctuation removal, and lemmatization) to eliminate dissimilar products.

3. Apply feature extraction using Word2Vec–SGram technique

4. Apply MDPCA to form clusters and find most preferable user from every cluster.

5. Apply procedure RNN_BES (call pseudo code 2).

6. Classify the user ratings – “NORMAL,” “PUSH,” “NUKE.”

5.1 Dataset

Dataset in https://www.kaggle.com/datafiniti/consumer-reviews-of-amazon-products is used. There are about 35,000 consumer reviews for Amazon electronic products such as Kindle, Fire TV Stick, and so on. The dataset includes basic product data, review text, rating, and more for each product. The user ratings are collected in the rating scale of 1–5. There are 22 attributes in the dataset. Out of the available 22 attributes, userId, userReview, and userRating are used for attack detection. The Review data type is text and user rating data type is integer. The platform used for the experiment is Python.

The algorithm enters into the select stage and the best bias is found as 25–75%. The model was created with 75% of the data and the test data use 25% for the best fit model.

5.2 Experiment and results

First, necessary packages “keras,” “pandas,” and “genism” for the proposed model are imported. “Pandas” package is used to read the dataset. Pre-processing is done by replacing NAN cells into 0 and removing unwanted data. Feature extraction by word embedding technique Word2Vec is done by using “genism” package. After that, clustering of data using MDPCA is done. Finally, the clustered data are fed into ARNN for classification. After the classification, Push and Nuke ratings are discarded from the dataset and recommendations are generated from genuine ratings. Performance measures are calculated for existing algorithms and their results are compared with the proposed ARNN algorithm. Table 3 shows the result for various performance measures for both proposed (ARNN) and existing techniques (DNN, RNN, MV-EDM, CNN, and SDAE-PCA). It depicts the accuracy performance for both proposed and existing techniques. As a measure of recommendation quality, accuracy measures the degree to which recommendations are in alignment with the user’s preferences. The accuracy performance of the proposed method is 98.66% and for the existing methods it is as follows: SDAE-PCA is 82.60%, MV-EDM is 85.28%, CNN is 88.94%, DNN is 92.75%, and RNN is 94.28%. The proposed method ARNN gives better accuracy results (98.66%) than other existing methods.

As a concept F-measure is calculated by taking the harmonic mean of recall and precision measures. The significance of using F-measure is that the harmonic mean penalizes extreme values as compared to average means. The F-measure performance of the proposed method is 98.67% and that of the existing methods – SDAE-PCA is 87.76%, MV-EDM is 90.54%, CNN is 92.34%, DNN is 92.83%, and RNN is 94.20%. The value of proposed method ARNN is 98.67%, which is the highest among other existing methods. This implies that the proposed method ARNN is better than the existing methods.

In recommendation systems, it is desired to recommend only top N products to the user. To achieve this, precision metric is computed by considering topmost N results. The precision performance of the proposed method is 98.50% and for the existing methods it is as follows: SDAE-PCA is 83.67%, MV-EDM is 87.01%, CNN is 89.64%, DNN is 91.85%, and RNN is 93.89%. The proposed method ARNN got 98.50%, which is the better precision result than other existing methods.

Similarly, Recall is the percentage of true documents that are successfully retrieved. The recall performance of the proposed method is 98.83% and for the existing methods it is as follows: SDAE-PCA is 89.10%, MV-EDM is 90.12%, CNN is 93.67%, DNN is 93.84%, and RNN is 94.67%. The recall value for the proposed method ARNN is 98.83%, which is best among the existing methods.

Another metric, false alarm rate (FAR) is defined as a proportion of number of wrong alarms to total number of warnings. Table 3 shows that the FAR performance measure of the proposed method ARNN is 98.67% and for the existing methods it is as follows: SDAE-PCA is 26.78%, MV-EDM is 21.43%, CNN is 18.90%, DNN is 7.24%, and RNN is 5.74%. It can be concluded that the proposed method ARNN has least FAR than other existing methods and thus is best of all.

Figure 8 shows the comparative analysis of various performance measures – accuracy, F-measure, precision, recall, and FAR for existing methods – DNN, RNN, MV-EDM, CNN, and SDAE-PCA and the proposed method ARNN. Figure 9 represents the receiver operating characteristic curve for the proposed ARNN model. It is clear from the graph that the proposed method ARNN has highest accuracy, precision, recall, and F-measure values and least FAR value as compared to other existing methods DNN, RNN, MV-EDM, CNN, and SDAE-PCA.

Figure 8

Performance analysis of HRS.

Figure 9

Receiver operating characteristic curve.

5.3 Computational complexity

The computational complexity of the proposed ARNN model is O(n), where n is the length of the input sequence. Table 4 shows the computational complexity of ARNN model. The result shows that the ARNN model has less computational complexity because of high convergence speed of the BES algorithm.

5.4 Limitation of research work

This article confines the work to detect shilling profiles (push attacks and nuke attacks) in HRS but not on multi-criteria CFRSs. The multi-criteria CFRSs take into account multiple views collected on different features of a product. Thus, wrong recommendations are viable for the fraud reviews. The proposed ARNN algorithm can be extended to deal with such system.