Keywords

1 Introduction

Computing in terms of computer technology refers to the process of executing tasks with the help of a computer device. A few characteristics of the computing process are: solutions must be precise and valid, there should be unambiguous and correct control sequences, and finally, formulation of the mathematical solutions of the problems should be easy. Computing methods are categorized into two folds: hard computing and soft computing. Hard computing is the traditional practice that draws on the postulates of precision, certainty, rigour, and inflexibility. It requires a well defined analytical model and often takes a significant amount of computation time. On the other hand, soft computing is different from conventional computing, it includes the concept of approximate models and provides solutions to tricky real-world problems. Unlike hard computing, soft computing deals with imprecision, uncertainty, approximations, and partial truths. Soft computing incorporates modern theories and practices such as expert systems, fuzzy logic, genetic algorithms, artificial neural networks, and machine learning. Some of the key differences between hard and soft computing are: hard computing draws on binary (two-valued) logic and deterministic in nature, while soft computing works upon formal (multi-valued) logic and stochastic reasoning; hard computing requires exact data for its mechanism, while soft computing can tackle ambiguous and noisy data; hard computing executes sequential computations, on the other hand, soft computing is capable of performing parallel computations; hard computing requires explicit programs to be written, while soft computing can emerge its own programs.

Sentiment analysis is one of the soft computing techniques that perceives positive, negative, or neutral opinions, known as polarity within a piece of text. This text can be a clause, sentence, paragraph, or a whole document. Let us take customer feedback as an illustration, sentiment analysis weighs the inclination of customers towards a product or service, which they express in textual form as comments or feedbacks. For example, consider the following feedbacks by two different customers (Table 1).

Table 1 Comments and corresponding sentiments
Full size table

The goal of sentiment analysis is to take a piece of text as input, analyze it, and returning a metric or score that estimates how positive or negative the text is. The process can be understood as context-based mining of text to identify and extract subjective information from source data. It helps businesses to judge the social opinions of their products, brands, and services by monitoring the online activities of their customers. However, the analysis of web and social media platforms is limited to trivial sentiment analysis and count-based metrics: akin to engraving the surface and overlooking other important insights that ought to be discovered.

Sometimes, sentiment analysis is coupled with text analytics and people often consider them as the same or related processes. Though both procedures extract meaningful ideas from customer data, both are the essential constituents of the customer experience management module, but, they are not the same thing. As we know, the former classifies a piece of text or expression as positive, negative, or neutral and determines the degree of this classification, the latter is concerned with the analysis of the unstructured text, extracting apt information, and converting it into productive business intelligence. Text analytics deals with the semantics of the text: involving the grammar and the relationships among the words. In general terms, text analytics draws out the meaning, while sentiment analysis develops an insight into the emotions behind the words. Sentiment analysis has an upper hand over text analytics that the former can be applied to non-text feedbacks such as emoticons or emojis. A ‘grinning face with big eyes’ emoji is coupled with a higher sentiment score than the emojis of ‘frowning face’ and ‘zipper-mouth face’.

Recently, text and sentiment analysis has received tremendous attention, especially due to the availability of gigantic data in the form of unstructured text available on social media, E-commerce websites, e-mails, blogs, and other similar web resources. This requires analyzing large volumes of unstructured text, extracting relevant information, and determining people’s opinions and expressions. Nowadays, the majority of business firms are using text and sentiment analysis techniques to understand the feedbacks of their customers and to gain information about the degree of customers’ inclination towards their products and services. Therefore, sentiment analysis provides valuable insights and helps the firms to formulate effective business strategies. However, the massive data derived from social media and other sources are unstructured, highly dimensional, and involve uncertainty and imprecision. This kind of massive text usually contains white spaces, punctuation marks, special characters, @ links, hashtag links, stop words, and numeric digits etc. This unstructured data must be cleaned before being fed to the classification models. These types of unnecessary expressions or characters can be removed using data pre-processing libraries available in Python. Thanks to soft computing techniques, we are equipped to handle uncertainty imprecision, partial truth, and approximation.

The rest of the chapter is organized as follows: Sect. 2 covers the literature review, followed by Sect. 3, data collection and methodology, Sect. 4 presents experimental results and discussions, followed by the final concluding section.

2 Literature Survey

Pang and Lee [20] presented an exhaustive survey on opinion mining and sentiment analysis. They explored research works that promise to directly enable opinion-oriented information-seeking systems. They focused to give more attention to contemporary challenges raised by modern sentiment-aware applications rather than already available traditional fact-based analysis models. Prabowo and Thelwall [21] proposed a hybrid approach to sentiment analysis based on rule-based classification and machine learning. They proposed a complementary and semi-automatic approach where every classifier supports other classifiers. They tested their hybrid model over movie reviews, product reviews, and MySpace comments and reported that a hybrid model is capable of improving classification effectiveness in terms of micro-and macro-averaged F1 measure. The authors suggested that in real-world applications, it would be better to have two rule sets: the original and induced rule sets. Barbosa and Feng [5] investigated the writing pattern of Twitter messages and meta-information of the words that constitutes them. Based on this data, they proposed the automatic detection of sentiments on tweets. They utilized biased and noisy labels of tweets provided by a third party and used this source as training data. They combined these labels by utilizing various strategies and compared their model with already existing techniques. The authors claimed that the solution proposed by them can handle more abstract representation of tweets and proved to be more robust and effective. Agarwal et al. [1] studied Twitter data for sentiment analysis. They proposed two models: one binary model to classify tweets as positive and negative and one 3-way model to classify them as being positive, negative, and neutral sentiment. They performed experiments with the unigram model, feature-based model, and kernel-based model. The authors used the unigram model as a baseline and reported an overall gain of 4% for these classification tasks. They claimed that the feature-based and tree kernel-based models outperformed the unigram baseline. In their concluding remarks, the authors stated that the sentiment analysis for Twitter data is the same as sentiment analysis for other genres. In their work, Gräbner et al. [12] proposed a classification system of the reviews of hotel customers employing sentiment analysis. Given a corpus, they designed a process to collect words that are related semantically and developed a domain-specific lexicon. This lexicon served as the key resource to develop a classifier for the reviews. The authors claimed to achieve a classification accuracy of 90%. Liu [14] presented a minutely detailed work on sentiment analysis and opinion mining. The author gave an in-deep introduction and presented a thorough survey of the available literature and the latest developments in the realm. This work presents an excellent qualitative and quantitative analysis of opinions and sentiments and stands as a distinguished literary resource for practical applications. The author endeavored to develop a common framework to bring different research works under a single roof and discussed the integral constituents of the subject like document-level sentiment classification, sentence-level subjectivity and sentiment classification, aspect-based sentiment analysis, sentiment lexicon generation, opinion summarization, and opinion spam detection. Bagheri et al. [3] proposed an unsupervised and domain-and language-independent model for analyzing online customers’ reviews for sentiment analysis and opinion mining. Their generalized model was equipped with a set of heuristic rules to detect the impact of opinion word/multi-word. They presented a novel bootstrapping algorithm and proposed a metric to detect and score for implicit and explicit aspects of reviews. They claimed that their model can be used in a practical environment where high precision is required. Medhat et al. [17] gave a detailed analysis of sentiment analysis algorithms and their applications. Their work can be considered as the state of the art in the domain. The authors categorized a large number of research articles according to their participation in sentiment analysis techniques for real-world applications. They suggested that further research is needed to enhance Sentiment Classification (SC) and Feature Selection (FS) algorithms. They commented that the Naïve Bayes and Support Vector Machine (SVM) serve as the base or reference techniques for comparing novel SC algorithms. Fang and Zhan [10] proposed a general process for sentiment polarity categorization thereby giving detailed descriptions. They studied online product reviews from Amazon.com over the following major categories: beauty, books, electronics, and home appliances. Each review includes rating and review text among other data. The authors used a part of speech (POS) tagger at the preprocessing step and then computed sentiment score. They used the F1 measure to evaluate the performance of their proposed classification process and reported that the SVM model and the Naïve Bayes model performed almost the same. Mozetič et al. [19] exploited a big set of tweets from different languages. The tweets were labeled manually and they exploited them as training data. They proposed automatic classification models and reported that the performances of the top classification models are not statistically different. The authors concluded that it would be good to give more attention to the accuracy of the training data than the genre of the model being employed. They found that on applying to the three-class sentiment classification problem, there is no correlation between the accuracy and performance of the classification models. From the literature available on sentiment analysis, this work on human annotation is very unique. Saad and Saberi [22] presented a survey of sentiment analysis and opinion mining techniques, their applications, and challenges. The authors classified such techniques into three groups: machine learning approach, lexicon-based approach, and combination method. They collected data from blogs and forums, reviews, news articles, and social networks (Twitter and Facebook). They concluded that the unstructured data is a big hurdle in sentiment analysis and stated that algorithms of sentiment classification and opinion mining need further research for improvement. Ghag and Shah [11] mentioned that the bag-of-words is a popular tool of sentiment analysis. The authors classified the sentences extracted from the sentiments by reviewing their syntactic and semantic structures. They proposed some metrics like relative frequency, term frequency, and inverse document frequency to improve accuracy. They used text preprocessing techniques and claimed to achieve 77.2% classification accuracy.

Employing automatic text clustering and manual qualitative coding, Mäntylä et al. [16] analyzed around seven thousand research articles from Scopus and Google scholar and presented a computer-assisted literature review for sentiment analysis. They highlighted a very interesting fact that automatic sentiment analysis had been possible only with the availability of online subjective texts and therefore, 99% of the research work in this domain took place after 2004. According to the authors, computer-based sentiment analysis started by analyzing product reviews available over the web, and it is now being applied over a wide range of domains like social media texts (Twitter, Facebook, etc.), stock markets, elections, disasters, medicine, and cyberbullying. They stated that sentiment analysis involves a multitude of data sources like tweets, comments, chats, emoticons etc. Alsaeedi and Khan [2] investigated applications and results of various sentiment analysis techniques over Twitter data. They explored machine learning, lexicon-based approaches, ensemble approaches, and hybrid approaches. They reported the following conclusions of their research work: when multiple features were taken, machine learning techniques resulted in the greatest precision; lexicon-based techniques performed good but they require manual efforts to create the archive, and the ensemble and hybrid-based algorithms performed better than supervised machine learning algorithms. Tyagi and Tripathi [26] also collected Twitter data and performed sentiment analysis. The authors extracted the features through the N-gram modeling technique and exploited the K-Nearest Neighbor algorithm to categorize sentiments into positive, negative, and neutral. Bhagat et al. [6] studied online product reviews, general tweets in Twitter, and movie reviews and carried out sentiment analysis of text messages using supervised machine learning techniques. They preprocessed the messages and applied Naïve Bayes, Decision Tree, and Support Vector Machine (SVM) techniques for their research. They proposed a three-tier framework: the first layer is the initialization layer for data collection and message preprocessing, the second layer is the learning layer which splits preprocessed data into training and test datasets and develops three machine learning models, the final layer evaluates the performance of the models based on precision, recall, F1-measure, etc. The authors concluded that the Decision Tree and SVM can be considered as good classifiers with lower mean square error.

3 Data Collection and Methodology

As we all know, Amazon is one of the leading E-commerce websites, where a large number of users’ reviews can be found. After purchasing the products, customers can post their reviews directly on the Amazon review portal. With such a massive amount of customers’ reviews, this provides an opportunity to study and investigate feedbacks of the customers about a specific product [8]. All such comments or feedbacks help the sellers and other potential customers comprehend product-related public opinions. In the present case study, we are taking reviews of Amazon customers for sentiment analysis. We propose a three-tier model that takes raw data from the Amazon portal as input and generates a comparative report over certain parameters. In the first phase, we fetch data from the portal, apply data preprocessing and cleaning techniques to repair and/or remove the dirty data. In the second phase, we apply TF-IDF and Skip-Gram models for statistical feature selection. This step filters out those input variables which exhibit the strongest relationship with output variables. In the final phase, we apply machine learning (ML) algorithms say, Random Forest, Naïve Bayes, and Support Vector Machine analyzing the processed dataset and to classify the customers’ reviews into the genres of positive, negative, and neutral. These algorithms are applied to the processed dataset to study the following performance parameters: accuracy, precision, recall, F measure, true positive, and false negative. Finally, our study compares the outputs of these three classifiers over the above-mentioned parameters. Figure 1 depicts the elements of our proposed model.

Fig. 1
figure 1

Proposed model and its constituents

Full size image

We hypothesize a four-fold methodology for our present research work: (a) data collection, (b) data preprocessing, (c) data representation, and (d) data classification. We now discuss the above steps in detail below.

3.1 Data Collection and Preprocessing

To conduct this case study, we gathered data from the Amazon web portal using an automated technique known as Scrapping. Scraping is a data extraction technique used for data collection from different websites. Scrapy is a free and open-source web-crawling framework which is written in Python and it is used for extracting data from websites. We applied the scraping process to extract Amazon reviews using the Scrapy library which permits the programmers to extract the data as per their requirements [18]. In our practical experiments, the scrapped data set consists of 300,000 mobile phone reviews from the Amazon review portal for various international brands. However, these reviews are in unstructured and unlabeled text form which requires pre-processing treatment. This is an essential step of the whole process as the accuracy of machine learning models depends on the quality of data we feed into them. The scrapped data set used in our research had many missing or null values. We dealt with these issues by utilizing the Imputation technique, a widely used tool in the realms of machine learning and data mining. The basic principle of the technique is to replace each missing value of an attribute with the mean of the observed values of the attribute, known as Mean Imputation (MEI), or a nominal attribute with its most commonly observed value, known as Most Common Imputation (MCI). For each attribute fi with missing values, the classifier Ci (. . .) takes as input the values of the other (n – 1) attributes {fj | j ≠ i} for an instance, and returns the value for fi for this instance [4, 15, 24]. Other preprocessing treatments applied to this data before feeding the data to the machine learning models are spellings corrections; stop words removal; removal of special characters and punctuations from text data; removal of multiple spaces; removal of numeric digits from the review texts; removal of all URLs, hashtags, and E-mail addresses; upper to lower case conversion; contraction to expansion; substitution of any non-UTF-8 character by space; stemming; and removal of rare words. To improve the performance of the classifier’s models, some of the irrelevant attributes (like reviews.dateSeen, reviews.sourceURLs, reviews.title, reviews.username, etc.) have been dropped after pre-processing.

Ultimately, after applying all of the above-mentioned preprocessing treatments, we receive accurate, useful, and clean text suitable for analysis and classification of sentiments. Table 2 presents the final extracted attributes and their description.

Table 2 Features information of Amazon reviews dataset
Full size table

Every product rating is based on a 5-star scale ranged from 1-star to 5-star with no existence of a half-star or a quarter-star. Figure 2 depicted below shows the distribution of reviews based on Amazon’s 1–5-star rating scales.

Fig. 2
figure 2

Distribution of Amazon’s star-rating scores

Full size image

As shown in the above figure, the most frequent review rating in our dataset is 5 stars, with more than 30% share in the entire dataset. Figure 3 illustrated below shows the attributes which are of the numerical type and their distribution in the data set.

Fig. 3
figure 3

Distribution of numerical data in the dataset

Full size image

It is clear from the above figure that reviews.numHelpful is a valuable attribute in our dataset, so we kept only those instances in the dataset for which more than 75 people found the review helpful. On the other hand, in reviews.rating attribute, the distribution is skewed towards 5 stars rating. The last two attributes, reviews.userCity and reviews.userProvince have NaN values i.e., a numerical value that is undefined or not present. Therefore, we have dropped these attributes from our dataset. One important attribute that is used for product identification is Amazon Standard Identification Number (ASIN). Our dataset has 35 different products which possess unique ASIN values and are used for training our classifiers. After analyzing ASINs and product name attributes, we observed that there’s a one to many relationships between the ASINs and the product names, i.e., a single ASIN is linked with one or more product names. Figure 4 shown below visualizes the individual ASIN and product reviews in a bar graph representation.

Fig. 4
figure 4

Review ratings and ASIN frequencies

Full size image

The above figure clearly shows that certain products have significantly more reviews than other products, which indicate a higher sale of those products. Based on this ASIN attribute frequency graph we can easily decide which products should be kept or dropped. Now, for better insight into the data or corpus, the Wordcloud visualization is an excellent tool in practice. The word that appears more prominent based on the frequency or importance in the text data is displayed with the bigger size in the Wordcloud visualization. In simple words, the word with larger size has more weight than the word with smaller size. After the completion of pre-processing of the dataset, we visualize words from the reviews’ text using Wordcloud feature as shown in Fig. 5 below.

Fig. 5
figure 5

Wordcloud of reviews’ text

Full size image

3.2 Data Representation

After the pre-processing of the unstructured text, the data representation is a vital step in sentiment classification. The extracted pre-processed reviews are mainly in text format but numerical representation in terms of metrics is needed to classify sentiments using the machine learning algorithms. Therefore, we have applied two different approaches to convert text data into some suitable form to be fed into the machine learning classifiers. The first approach is word embedding and second is the combination of Term frequency and inverse document frequency (TF-IDF). For word embedding, we applied the Word2vec model with skip-gram architecture. The skip-gram model predicts the source context words given a target word. It works as an unsupervised learning technique that is used to find the most suitable and related words for a given target word [13]. Skip-gram architecture provides more accurate and effective results when we have a corpus of bigger size, because, in the skip-gram approach, each context-center pair is considered as a new observation. The word vectors are adapted using Eq. (3.1), as given below:

$$w_{i,j} \left( {k + 1} \right) = w_{i,j} \left( k \right) - s\frac{\partial J}{{\partial w_{i,j} }}$$
(3.1)

where \(w_{i,j} \left( k \right)\) is word vector value in step \(k\) of the optimization process, \(j\) is our optimization function and \(s\) is the chosen step size. The optimization function is applied for selecting those words which can be represented using the Eq. (3.2) given below.

$$J = \mathop \sum \limits_{i,j = 1}^{V} f\left( {X_{i,j} } \right) \left( {w_{i}^{T} \tilde{w}_{j} + b_{i} + \tilde{b}_{j} - \log X_{i,j} } \right)^{2}$$
(3.2)

where \(V\) is the number of word tuples with the non-zero co-appearance count, \(X_{i,j}\) is the count of co-appearances, \(w_{i}\) is a word vector and \(\tilde{w}_{j}\) is word vector’s context, \(b_{i}\) and \(\tilde{b}_{j}\) are biases (again every word has two of them: one for the word and other for the context) and function \(f\) is a weighing function. The skip-gram architecture is illustrated in Fig. 6 given below.

Fig. 6
figure 6

Architecture of word to vector skip-gram model

Full size image

The TF-IDF algorithm is based on words’ statistics for feature extraction and represents how important a word or a phrase in a corpus. TF-IDF assigns a unique score to each word using a hybrid statistical method, in terms of the product of term frequency (TF) with inverse document frequency (IDF). The TF denotes the total number of times a given term occurs in the dataset against the total number of all words in the document, and the IDF measures the amount of information word provides [23]. In our case study, TF assigns a score to most frequently occurring words in the mobile review dataset and IDF assigns weight to the least frequent words in the same dataset. The TfidfVectorizer from the sklearn python library is used to fit the vectorizer on the corpus of the review texts to extract features and the model will transform the text data into the TF-IDF representation. Using the TF-IDF approach in a normalized data format, each corpus word can be represented using the following equation:

$$W_{ij} = tf_{ij} \times \log \left( {\frac{N}{{df_{i} }}} \right)$$
(3.3)

where, \(tf_{ij}\) is the number of occurrences of the ith word in the jth review and \(df_{i}\) is the number of reviews containing i and N is the total number of reviews.

3.3 Classifications

Classification is a supervised machine learning process that generally focuses on predicting a qualitative response by recognizing and analyzing given data points. This case study is focused on sentiment analysis by classifying the reviews into 1–5 ranked scales. To carry out this task, we have applied three different supervised classifiers: Naïve Bayes, Support Vector Machine, and Random Forest. These classifiers are then evaluated to provide a comparative analysis of various parameters for classifying reviews into positive and negative genres. The literature suggests that reviews with ratings 4 and 5 should be categorized as positive reviews while reviews with rating 3 should be labeled as neutral and reviews with ratings 2 and 1 should be treated as negative reviews. Since here we are interested in analyzing only positive and negative reviews, so we have neglected neutral reviews form the data set. The working of the individual classifier is explained in the next sections.

3.3.1 Naïve Bayes Classifier

The Naïve Bayes (NB) classifier is a simple and robust probabilistic classifier algorithm that is based on the Bayes theorem. It assumes that attribute values are independent of each other given the class. This assumption is known as the conditional independence assumption. Therefore, applying changes in one feature does not affect other features of the class [7]. Let D be our Amazon review data set for training the model then each tuple in the dataset is defined with n attributes and it is represented by: \(X = \left\{ {a_{1} ,a_{2} ,a_{3} , \ldots ,a_{n} } \right\}\). Let there be m classes represented by: \(\left\{ {C_{1} ,C_{2} , C_{3} \ldots ,C_{m} } \right\}\). For a given tuple X, the classifier predicts that X belongs to the class having the highest posterior probability, conditioned on X. The Naïve Bayes classifier predicts that the tuple X belongs to the class Ci if and only if \(P\left( {C_{i} |X} \right)\) is maximum among all i.e.:

$$P\left( {C_{i} |X} \right) > P\left( {C_{j} |X} \right) for 1 \le j \le m, i \ne j$$
(3.4)

Since we want to maximize \(P\left( {C_{i} |X} \right),\) the class Ci for which \(P\left( {C_{i} |X} \right)\) is maximized is called the maximum posterior hypothesis. According to the Bayes theorem,

$$P\left( {C_{i} |X} \right) = \frac{{P\left( {X |C_{i} } \right) \cdot P\left( {C_{i} } \right)}}{P\left( X \right)}$$
(3.5)

If the attribute values are conditionally independent of one another (Naïve Bayes condition), then

$$P\left( {X |C_{i} } \right) = \mathop \prod \limits_{k = 1}^{n} P(x_{k} |C_{i} )$$
(3.6)

where \(x_{k}\) refers to the value of the attribute \(A_{k}\) for the tuple X. If \(A_{k}\) is a categorical attribute, then \(P\left( {x_{k} |C_{i} } \right)\) is the number of tuples of class \(C_{i}\) in D having the value \(x_{k}\) for \(A_{k}\). The classifier predicts the class label of X is in the class Ci if and only if,

$$P\left( {X |C_{i} } \right) \cdot P\left( {C_{i} } \right) > P(X|C_{j} ) \cdot P\left( {C_{j} } \right)$$
(3.7)

3.3.2 Support Vector Machine

Support Vector Machines (SVM) is a supervised machine-learning based classification algorithm which widely deals with predictive and regression analysis. SVM algorithm aims to find a hyperplane in an N-Dimensional feature space that distinctly classifies the data points, while maximizing the marginal distance for the two classes (positive and negative) and minimizing the classification errors. The marginal distance for a class is the distance between the decision hyperplane and its nearest instance which is a member of that class [25]. The data points that lie closest to the decision surface (or hyperplane) are called support vectors and these points help us in building the SVM model. The loss function that helps in maximizing the margin is given below.

$$C(x, y\left( {f\left( x \right)} \right) = f\left( x \right) = \left\{ {\begin{array}{*{20}c} {0, } & {x < 0} \\ {1 - y*f\left( x \right)} & {else} \\ \end{array} } \right.$$
(3.8)
$$C(x, y\left( {f\left( x \right)} \right) = \left( {1 - y*f\left( x \right)} \right) +$$
(3.9)

The equation of the line in 2D space is \(y = a + bx\). By renaming \(x\) with \(x_{1}\) and \(y\) with \(x_{2}\), the equation will change to \(ax_{1} - x_{2} + b = 0\). If we specify \(X = \left( {x_{1} ,x_{2} } \right)\) and \(w = \left( {a, - 1} \right)\), we get,

$$F\left( x \right) = w \cdot x + b \,\,where\,\, w, x \in R^{n} \,\, and\,\, b \in R$$
(3.10)

The above Eq. (3.10) is called the equation of the hyperplane, which linearly separate the data.

The hypothesis function h in SVM classifier can be defined as:

$$h\left( x \right) = \left\{ {\begin{array}{*{20}c} { + 1 } & {if w.x + b x \ge 0} \\ { - 1} & { if w.x + b < 0} \\ \end{array} } \right.$$
(3.11)

The point above or on the hyperplane will be classified as class +1, and the point below the hyperplane will be classified as class −1. SVM classifier amounts to minimizing an expression of the form given below:

$$\left[ {\frac{1}{n}\mathop \sum \limits_{i = 1}^{n} { \hbox{max} }\left( {0,1 - {\text{yi}}\left( {{\text{w}}.{\text{x}}_{\text{i}} - {\text{b}}} \right)} \right)} \right]$$

The reason to choose this classifier in the present research is its robustness. It provides an optimal margin gap to separate hyperplanes and it gives a well-defined boundary for easy classification.

3.3.3 Random Forest

Random Forest (RF) machine learning algorithm is also known as a tree-based ensemble learning, which creates a forest of many decision trees. RF ensures that the behavior of each decision tree produced is not too correlated with the behavior of any other decision tree in the model. This final prediction can simply be the mean of all the observed predictions [9]. Therefore, the different decision trees obtained by using the RF algorithm are trained using different parts of the training dataset, which is the reason behind its unbiased nature and superior prediction accuracy.

4 Experimental Results and Discussions

For conducting the practical implementation of this case study, we used Jupyter Notebook with Python version 3.8. Various Python libraries have been used for data pre-processing and visual representation such as pandas, numpy, scrapy, matplotlib, seaborn, spacy, etc. For training and testing of machine learning classifiers, the corpus is divided into two subsets with a train-test split of 75–25% respectively. To evaluate the performances of the classifiers, the main parametric metrics employed in this research are accuracy, precision, recall, F-measure, true positive and false negative. In classification problems, precision (also called positive predictive value) is the fraction of relevant instances that are retrieved, while recall is the fraction of relevant instances that have been retrieved over the total amount of relevant instances. Accuracy of the model can be defined as the ratio of the correctly labelled attributes to the whole pool of variables. F-measure is a weighted average of precision and recall. As ASIN and review ratings are two important attributes available in our dataset, therefore we explored their relationship visualized in Fig. 7 below.

Fig. 7
figure 7

Relationship between ASIN and review ratings

Full size image

The above figure clearly reveals that the most frequently reviewed products in our case study have their average review ratings above 4.5. On the other hand, ASINs with lower frequencies in the bar graph have their corresponding average review ratings below 3. For analyzing the classification performance of machine learning models, we applied Naïve Bayes, Support Vector Machine and, Random Forest algorithms to the pre-processed dataset. The performance evaluation results of machine learning classifiers using Skip-gram and TF-IDF feature extraction techniques are shown in Table 3 and Table 4 respectively.

Table 3 Classifiers comparison using Skip-gram feature extraction method
Full size table
Table 4 Classifiers comparison using TF-IDF feature extraction method
Full size table

The results of the Table 3 above, clearly show that the Random Forest classifier achieves maximum accuracy in skip-gram model.

Table 4 reveals that TF-IDF significantly improves the accuracy along with other important parameters of Naïve Bayes and Support Vector Machine classifiers but does not perform well enough with Random Forest. After comparing the results of Table 3 and Table 4, it is clear that the TF-IDF approach improves the accuracies of Naïve Bayes and Support Vector Machine classifiers by 6% and 3.9% respectively but deteriorates the accuracy of Random Forest by a margin of 0.8%. Figure 8 compares the classification accuracies of individual classifiers against the Skip-gram and TF-IDF feature extraction techniques.

Fig. 8
figure 8

Accuracies comparison of Skip-gram and TF-IDF techniques

Full size image

As shown in the above figure, it is evident that in the case of Naïve Bayes classifiers, the classification accuracy obtained using TF-IDF is better than the value obtained using the Skip-gram technique. The Naive Bayes algorithm follows a probabilistic approach, where the attributes are independent of each other. Therefore, when the analysis is performed using a single word (unigram) and double word (bigram), the accuracy value obtained with TF-IDF is comparatively better than that obtained using Skip-gram. Similarly, it is clear that in the case of Random Forest, the classification accuracy value obtained using the Skip-gram technique is a little better than the value obtained using TF-IDF. As we know, Random Forest is an ensemble tree-based classifier and it aggregates the output obtained from different decision trees, the Skip-gram model which can predict the source context words given a target word gives better results. In the case of Support Vector Machine classifiers, the classification accuracy attained using TF-IDF is better than that obtained using the Skip-gram approach. Support vector machine is a non-probabilistic linear classifier and the trained classifier is used to find hyperplane for dataset separation, the TF-IDF which analyses the corpus word by word gives better results as compared to the Skip-gram model. Figure 9 shown below presents the Heatmaps of confusion matrices obtained.

Fig. 9
figure 9

Heatmaps of the confusion matrices

Full size image

The above figure depicts the four best confusion matrices obtained from various classifiers, which is a summary of the prediction results in our classification problem. The part (a) above, shows the trained model predicts True Negative of 86%, True Positive of 51%, False Positive of 4% and False Negative with 2%. Therefore, 86% and 51% are the correct predictions and 4% and 2% are incorrect predictions. We can see that we have a good ratio of correct predictions. Similarly, we can interpret the confusion matrices of other classifiers. Note that the SVM with TF-IDF shows maximum accuracy, as shown in part (d) of Fig. 9.

5 Conclusions

Analysis of sentiments is crucial for any online retail business enterprise to understand the opinions and feedbacks of its customers. This case study analyses the sentimental polarities of the scrapped user-reviews of Amazon customers through machine learning classifiers. The dataset used in the present chapter is collected from the Amazon review portal using the well-known scrappy library available in Python. We scrapped 300,000 mobile phone reviews from Amazon review portal for various international brands. This unstructured dataset had to be preprocessed first to convert it into a legitimate form so that machine learning classifiers can process it smoothly. The null and missing values were dealt with imputation technique and several preprocessing techniques like stop word removal, spelling corrections, stemming, special character handling, etc. were also exercised. Now, this preprocessed dataset which was in text form needed to be converted into numerical scores before submitting it into machine learning models. We employed two techniques for this preliminary step: Skip-gram and TF-IDF. After the above treatments, we put the processed dataset to three different classifiers: Naïve Bayes, Support Vector Machine, and Random Forest. The above-mentioned machine learning classifiers are then evaluated over some standard parameters say, accuracy, precision, recall, F-measure, true positive and false negative. The empirical results found to be very satisfactory.

The present sentiment analysis case study of the Amazon reviews can be considered a kind of novel work where various machine learning classifiers have been compared against two different feature engineering techniques. Empirical results reveal that all models are able to classify the user reviews into negative and positive classes with relatively high accuracy and precision. Calculated results exhibit that the Support Vector Machine model achieved the highest classification accuracy (98.61%) with TF-IDF feature extraction method. Next, the Naïve Bayes model with TF-IDF achieved the classification accuracy of 95.33%. And, the Random Forest model with the Skip-gram technique acquired 95.2% accuracy at the third position.