Semantic Web Mining for Analyzing Retail Environment Using Word2Vec and CNN-FK

The Semantic Web is what its creator Tim Berners-Lee first intended for the World Wide Web: "The first move is to place the network data in a manner that machines will interpret or transform it into a shape. This generates a network that can be accessed explicitly or implicitly by computers, what I consider a Semantic Web [1]. This bold dream has therefore been streamlined to promote its start-up in consideration of the complexities of Site implementation. And why is this dream suddenly recovered? And why is e-business important? The Semantic Internet is not only accessible on the World Wide Web. It represents a set of technologies which works equally well on internal intranets as it attempts to solve a number of major issues in the current architectures of information technology [2]. The main feature of the Semantic Web in the sense of e-business is convergence [3]. The Internet and the World Wide Web have allowed knowledge systems to be linked globally.

Big data has recently been used mostly in retail management systems. The database generation is linked to variety, truthfulness, speed, volume and value. Compared to traditional mining techniques, process and management in these databases have increased capacity [4]. Most e-commerce firms use various methods in order to get customers away from the retail market by offering such services, such as cash back, stable exchanges and delivery cash, etc. In order to succeed in this dynamic market climate, retailers need to recognise their customers' problems and solve consumers' problems [5]. The various social media patterns are also regularly understood by retailers.

Standardization efforts are now underway to create increasing communication application frameworks in various areas [6]. They also, however, proliferated and concentrated on the same principles of syntactic interoperability that just slightly removed obstacles in most contexts [7]. Logical interoperability is the focus of Semantic Web. This provides the building blocks that facilitate the connections between computer-managed objects [8]. For example, the two sets of data relating to the same object or to more complex information like that, all objects of one type would be followed by something else. These relationships tell computers that there is a fundamental reality [9]. Computers themselves cannot infer such connections, and therefore it is necessary to train them the origins from which data are immediately incorporated [10]. Nevertheless, a major initial initiative is required, but it affects wide network environments [11]. However, computers have more paths to follow with these relationships between meanings as they try to determine whether two data objects are appropriate or which are the object of the invoice.

In the Semantic Web project, the research community has made a tremendous effort and we are now beginning to see realistic gain in several fields, especially the one with a strong influence on e-business [12]. The big prospects are foreseen in the interoperation of information systems which vary from inter- which intra-agency interactions [13]. The problem is then how the Semantic Web can assist in the creation of a webbed economy where members exchange knowledge and incorporate internal business processes in an open and trustworthy manner through spontaneous inter-organizational interactions [14].

According to a new survey, gross eCommerce departments in the United States just amounted to 119.1 billion dollars for the fourth quart of 2020. In the course of product search and purchase, however, there is one major problem with which customers have to contend [15]. Many e-shops have the same items, but knowledge about the goods differs widely across various e-shops [16]. In addition, there are no global product identifiers and deals are usually not interlinked. Consequently, the buyer cannot find all the details required and the best prices for their searched items. There is a lot of product aggregators, such as Google Product and PriceGrabber2 to combine and categorise the products of numerous E-shops and several different retailers to have a better users experience. However, it is not easy to define the same product across thousands of e-shops and integrate the data in one representation [17]. In addition the product aggregators have to handle the task of sorting all the items by taxonomy in order to enable better navigation and product search.

The retail sector has seen major changes in emerging market outlets and an extremely diverse worldwide customer base. Customers have deep implications for how retail businesses prepare and improve their supply chain processes [18]. These unwilling customers shop and expect a great shopping experience anywhere and everywhere. However, other retail firms concentrate on how to find new solutions in the fields of procurement, refilling [19], supply and pricing. Retail companies begin to investigate objectively how best their supply chain processes can be managed to meet ever-changing customer needs and associated expertise while reducing business costs and encouraging high efficiency.

A major problem relating to semantic progress in the definition of the Web Interface to ensure efficient consistency with their information systems in remote business services. Ontology usually is known as a web service of semantology [20]. The measurement of the semantiquity between web services can therefore be reduced to the measurement of semantiquity [21] among ontologies. Advanced computational mechanisms such as machine learning and deep learning have a strong influence to get good classification results [22].

In this proposed work, the output of words2vec with CNNs in retail sector to identify documents required and unnecessary is examined. The output of the CNN-FK model with word2vec is evaluated in this paper according to the data dimension and the number of periods. We have conducted comprehensive studies on large collections of retail environment to test the classification accuracy of CNN-FK in words2vec model. The experimental results showed a significantly improved accuracy of the classification model by using words2vec with which semantic relations between words are established.

In this paper proposing a model with a neural network architecture and a word representation mechanism called Word2Vec used. Initially word2vec handles the text data and represents it as a feature map. And feature map is given to Convolution neural network (CNN), it extracts the features and made classification. The rest of the papere is as follows section-2 address the existing literature, section-3 deals with the proposed model, section-4 illustrates the results and discussions and finally section-5 concludes the paper.

The proposed model with a neural network architecture and a word representation mechanism called Word2Vec used. Initially word2vec handles the text data and represents it as a feature map. The subsequent section explains the word2Vec mechanism and CNN model briefly. Here Figure 1 shows the proposed model working.

In deep neural networks, hidden layers are an essential aspect of learning features from input data sets. Deep neural networks are modified neural networks with many hidden layers. Data was feeded into the network after it was pre-processed in this model. This technique was widely used in the form of sequence patterns of words in the documents. Word embedding was carried out in order to convert word into feature vectors. Hidden layers learn the basic characteristics and supply the output units.

1.png

Figure 1. Proposed model framework

3.1 Word2Vec

Word2vec is the technique / model for generating word embedding to best reflect word. It catches several correct syntactic and semantic zed word connections. It's a shallow neural network of two layers. Input is modelled, while neurons comprise both the secret layer and the output layer. The Figure 2 represents the shallow vs deep learning model for considering the hidden layers.

2.png

Figure 2. Shallow vs. Deep learning

Word2vec is a two-layer network where there is input one hidden layer and output. Word2vec stands for terms on the representation of vectors in space. Words are expressed as lines, such that identical definitions together and separate concepts are placed further apart. Words are described in lines.

3.2 Convolution Neural Network Fisher Kernel

It is to be able to learn from a probability model, the logistical function must first be articulated. As mentioned above the CNN is a biased model and therefore the function P (X|Φ) cannot be represented. In order to show that the most advanced softmax layer of CNN is inspected, the generative loglike feature can be expressed. Remember that the softmax function of the CNN looks like:

$p\left(C_{k} / x, \emptyset\right)=\frac{\exp \left(w_{k}{ }^{T} x^{1}+b_{k}\right)}{\sum_{j} \mathrm{~T} \exp \left(w_{k}{ }^{T} x^{1}+b_{j}\right)}$                  (1)

where, Ck stands for the k-th class, wk, bk are tuneable weights and bias respectively. In the case of CNN x is an image, and xˆ are activations of the penultimate CNN layer. As it was shown by Neethu and Rajasree [5], the softmax function could be seen as an expression for the Bayes rule with tuneable parameters wk. And bk:

$p\left(C_{k} / x, \emptyset\right)=\frac{\exp \left(w_{k}^{T} x^{1}+b_{k}\right)}{\sum_{j} \times \exp \left(w_{k}^{T} x^{1}+b_{j}\right)}$$=\frac{p\left(^{x} / \emptyset, C_{k}\right) p\left(\emptyset, C_{k}\right)}{\sum_{j} \mathrm{x} p\left(^{x} / \emptyset, C_{j}\right) p\left(\emptyset, C_{j}\right)}$$=p\left({ }^{C_{k}} / \emptyset_{, x}\right)$                   (2)

From there it is also possible to see that the joint probability P (x, Φ, Ck) (i.e. The nominator of Eq. (2) is equal.

$\begin{aligned} p\left(C_{k}, \emptyset, x\right)=\exp &\left(w_{k}^{T} x^{1}+b_{k}\right) \\ &=p\left(^{x} / \emptyset, C_{k}\right) p\left(\emptyset, C_{k}\right) \end{aligned}$           (3)

To be able to derive a generative log likelyhood function one has to be able to express the generative probability P (X|Ω), where Ω is a newly introduced symbol for the set of model parameters. In the case of CNN it is thus proposed to move the variables C1... Ck into the set of model parameters (i.e. Ω = {Φ, C1... CK}) such that it is possible to express the probability of set of images X conditioned on Ω. At this point P (x|Ω) is defined as:

$\begin{aligned} P\left(x / \Phi, C_{1}, \ldots, C_{k}\right) &=P(x / \Omega)=\frac{P\left(x, C_{1}, \ldots, C_{k}, \Phi\right)}{P\left(C_{1}, \ldots, C_{k}, \Phi\right)} \\=& \frac{P(\Phi, x) \prod_{k=1}^{K}+P\left(C_{k} / \Phi, x\right)}{P\left(C_{1}, \ldots, C_{k}, \Phi\right)} \end{aligned}$                  (4)

where the independence of the probabilities P (C1|Φ, x)... P (CK |Φ, x) is assumed. Note that this assumption arises from the probabilistic interpretation of the softmax activation function whose formula is given in Eq. (2).

Assuming that samples P (x|Φ, C1... CK) are independent P (X|Φ, C1, ..., CK) then becomes:

$P\left(X / \Phi, C_{1}, \ldots, C_{k}\right)=\prod_{i=1}^{n} \mathrm{~K}+P\left(x_{i} / \Phi, C_{1}, \ldots, C_{k}\right)$                 (5)

If we were to follow Fisher Kernel framework, at this point the expression for the derivative of the loglikelyhood of P (X|Φ, C1, ..., CK) would follow where K is constant. However, there are several caveats that make this step very challenging.

Unknown P (Φ, x) and P (C1, ..., CK, Φ) The probabilities P (Φ, x) and P (C1, ..., CK, Φ) from Eq-4 are unknown. Note that it would be possible to assume uniform prior over P (C1, ..., CK, Φ), however P (Φ, x) depends on the data x, which is a property that cannot be omitted in the Fisher Kernel setting.

Loglikelyhood derivative Even if there was a possibility to overcome the issues with unknown probabilities, obtaining the derivatives of the loglikelyhood function with respect to the parameter set Ω would be a very challenging task.

Instead of formulating unrealistic assumptions, that would help us with obtaining the final evaluable formulation of P (X|Φ, C1, ..., CK) we opt for defining our own function Λ(x, Φ, C1, ..., CK) that has similar properties as the probability P (x|Φ, C1, ..., CK):

$\wedge\left(x, \Phi, C_{1}, \ldots, C_{k}\right)=\prod_{k=1}^{K} \mathrm{~K}-P\left(x, \Phi, C_{k}\right)$                   (6)

Function Λ in our formulation of Fisher Kernel based features retakes the role of prob- abilities P (x|Φ, C1, ..., CK) and K is constant. The FK classifier uses derivatives of the generative loglikelyhood function with respect to its parameters. Here since we use our own defined function Λ(x, Φ, C1, ..., CK), we call the expression, that is the equivalent to the generative likelihood L(X|Φ, C1, ..., CK) as pseudo-likelihood is defined as:

$\hat{\mathcal{L}}_{\wedge}\left(X, \Phi, C_{1}, \ldots, C_{k}\right)=\prod_{i=1}^{n} \mathrm{X} \wedge\left(x_{i}, \Phi, C_{1}, \ldots, C_{k}\right)$                   (7)

At this point it is important to note that in the case of CNNs, the set of samples X actually consists of only one observation, which is the image xi, i.e. in our case n = 1 at Threshold T. If the contents of Eq. (3) are plugged into the pseudo-likelihood formula Eq. (7) the following is obtained:

$\hat{\mathcal{L}}_{\wedge}\left(X, \Phi, C_{1}, \ldots, C_{k}\right)=\prod_{i=1}^{n} \mathrm{X} \prod_{k=1}^{K} \mathrm{~T} \exp \left(\omega_{k}^{T} \hat{x}+b_{k}\right)$                (8)

Taking the logarithm of Eq. (8) the corresponding pseudo-log likelyhood function is formed.

$\log \hat{\mathcal{L}}_{\wedge}\left(X, \Phi, C_{1}, \ldots, C_{k}\right)=\sum_{i=1}^{n} \sum_{k=1}^{K} \mathrm{P} w_{k}^{T} x^{1} \omega_{k}^{T} \hat{x}_{i}+b_{k}$                (9)

After taking a derivative of the function log ˆ(X, Φ, C1... CK) with respect to its Parameters Φ, C1, ..., CK , Fisher Kernel based features could be obtained.

Recall that the derivatives of Eq. (9) are not the proper Fisher Kernel features, be- cause of our decision to replace probabilities P (x|Φ, C1, ..., CK) with Λ(x, Φ, C1, ..., CK) which cannot be regarded as generative probability measures. However our selection of Λ(x, Φ, C1, ..., CK) could be justified.

The purpose of the generative probability is to assign higher values of P (x|Φ, C1, ..., CK) to images x that are more likely to be observed. Our function Λ computes the product of normalized class posteriors P (x, Φ, Ck). Λ thus reaches high values once P (x, Φ, Ck) are also elevated. This means that from the view of Λ the images that are likely to appear are those that contain objects from classes C1, ..., CK. We hope that the fact that Λ assigns high values to the images that contain actual visual objects could be regarded as a justification of our choice of Λ function as a suitable replacement of P (x|Φ, C1, ..., CK).

The loglikelihood gradients should have a 'meaningful' shape from the Fisher Kernel point of view. This implies that its directions should be structured so that linear classification can be carried out in this space. The gradients of model trained to optimise generative log-like characteristics are used in the case of Fisher Kernel. This quality of gradient directions is guaranteed by the fact that the probability of the model reaches its height. However it is not clear if these characteristics are also present in the gradients of the described pseudo-likelyness.

The consistency of this pseudo-likehood formula Eq. (9) is another positive characteristic of ŠINE. Since the terms have been omitted exponentially, the expression is a simple sum of linear functions. It is then a simple job to obtain its derivative.

The fact that n = 1 in formula Eq. (7) could also be regarded as a theoretical issue, since the Fisher Kernel was originally defined to compare the sets of samples X which generally have more than one element. This issue could be for instance resolved by picking random crops or flips of the original image x and adding them to the set X. This extension would be another step for improving our proposed method, which is not covered in this thesis. Also note that in our particular case variables X and x are ambiguous and both represent image x.

To conclude, the gradients of Λ cannot be regarded as Fisher Kernel features, because of the reasons mentioned in the previous paragraphs. However the substitution of probability P (x|Φ, C1, ..., CK) by our own function Λ is the only difference between the Fisher Kernel and our proposed method. Thus, in this thesis we choose to term the gradients of Λ Fisher Kernel based features, because of the evident resemblance to the original method.

3.png

Figure 3. CNN basic architecture

The CNN basic structure is represented in Figure 3. Word2Vec has been shown to be a valuable embedding technique for capturing significant relations based on its experimental performance. In the first stage of the analysis process, Word2Vec also has demonstrated strong representational capabilities for relevant feature consideration. A CNN-FK algorithm was combined with Word2Vec to perform a binary classification on applicants for input data. In the fields of retail environment data classification and identification, some recent studies have also been performed on Word2Vec but not accurate. The proposed model uses word2vec model with CNN-FK for web mining in retail environment.

3.3 Word2Vec and CNN-FK algorithm for classification of unstructured semantic web data

Algorithm Word2Vec based CNN-FK

{

Input: web data

Output: Classified data as Knowledge.

Step-1: Take the input data set and split it into training and testing data

Step-2: while training

For each word in the Training Data:

If it exists in the model:

For each word in the Content Tree:

      Calculate the correlation between the word vectors.

Step-3: Parent ß Word with Maximum Correlation

Add the word to the Content Tree as

the child of the parent

Step-4: Feature map ß{Set of nodes in the content tree}

Step-5: Apply CNN on the feature map

{

    Layer-1 Convolution ReLU

    Layer-2: Pooling.

    Layer-3: Flattening.

    Layer-4: Full Connection

}

Step-6: Return classified data

}

The vectorized value of a tokenized word is specifically transferred to the input layer by CNN-FK. In neighbouring terms, the input layer is extracted within a given window size. The procedure was referred to as a convolutionary operation and the features extracted during the convolution were used to construct a characteristic chart. CNN then carried out a max-pool over time, which removed the greatest value from the characteristics derived. The process allowed CNN-FK to extract a feature. CNN repeated the procedure by adjusting the window size to extract some functionality. The extracted features were supplied to the completely connected layer and functions of drop-out and softmax were performed.

This paper uses the word embedding technique 'Word2vec' to create sequence segmentation methods in order to transform data sequences into text vectors, thereby gaining insight into word associations of different lengths. A CNN-FK architecture is then developed via a convolutionary neural network with an embedding layer to carry out the identification task. Word2vec will translate the inner connection between words into numerical characteristics.

3.4 Deriving Fisher Kernel based features from CNN

In the following section it is shown how to use the gradients of the pseudo-loglikelyhood derived from a CNN in combination with a SVM solver and use it for classifying images. (the pseudo-loglikelyhood formula is given in equation Eq. (9). The kernel function KΛ that uses gradients of the pseudo-loglikelyhood and compares two sample sets Xi and Xj is the same as in the case of Fisher Kernel:

$K_{\wedge}\left(X_{i}, X_{j}\right)=U_{X_{i}}^{T} I^{-1} U_{X_{j}}$                 (10)

in this particular case UX stands for the derivative of the pseudo-log likelihood of the CNN Eq. (9) w.r.t. its parameters Ω evaluated at particular point (image) Xi:

$U_{X}=\left.\nabla \Omega \log \hat{L} \wedge\right|_{X_{i}}$                    (11)

Furthermore it is again possible to utilize the cholesky decomposition of the matrix I and express kernel function K(Xi, Xj) as a scalar product of two column vectors ΥXi and ΥXj. Where,

$K_{\Lambda}\left(X_{i}, X_{j}\right)=\gamma_{X_{i}}^{T} \gamma_{X_{i}}$                (12)

$\gamma_{X}=L U_{X}, \quad I=L^{\prime} L$                   (13)

After obtaining ΥX the ℓ2 normalization follows:

$\gamma_{X}^{\ell_{2}}=\frac{\gamma_{X}}{\left\|\gamma_{X}\right\|_{2}}$                       (14)

Note that for brevity the vector Υℓ2 will be simply denoted as ΥX. Classifier formed by using derivatives of the pseudo-loglikelyhood of CNN will be termed CNN-FK classifier, the vectors ΥX Fisher Kernel based features or shortly CNN-FK features.

The algorithm explains how the web data has been analyzed using proposed model. Initially word2vec handles the text data and represents it as a feature map. And feature map is given to Convolution neural network (CNN), it extracts the features and made classification.