Introduction

Customers are increasingly buying on a variety of e-commerce platforms due to the quick innovation and widespread use of e-commerce technologies1. These platforms serve a variety of tastes and styles by providing a large selection of goods and services2. Online shopping is simple, but there are drawbacks to the virtual aspect of e-commerce platforms, including mismatched product descriptions and real goods, subpar quality, and poor after-sales support. Sentiment analysis (SA) of consumer reviews of e-commerce products is therefore essential3. Analyzing client feedback helps to increase customer satisfaction and service quality on these platforms, as well as offering helpful referrals for other customers4. Sentiment analysis is the process of mechanically analyzing subjective remark texts to derive a customer’s emotional preferences. It is sometimes referred to as opinion mining or text orientation analysis5. Conventional approaches to text mining involve machine learning and rule-based techniques. The lexicon-related approach is a rule-based strategy; machine learning classification methods include conditional random fields and deep learning (DL). Data mining (DM) techniques are used to collect data and find solutions to a variety of problems because of the enormous volume of data that is created online. Sentiment analysis encompasses text mining, computational linguistics, natural language processing, and other disciplines, and it is applied to e-commerce websites where users can voice their thoughts on a variety of subjects (NLP)6,7. This tactic is advantageous since it makes it possible to look over and gather feedback on different products. Finding out how a person feels about a product or service is known as opinion mining. The main strategies for sentiment analysis are hybrid models, lexicon-related approaches, and machine learning techniques8,9. Deep learning is being used in this study to further examine the intricacies of sentiment analysis, which has been studied extensively because of its importance and versatility10.

Sentiment analysis aims to obtain sufficient accuracy by extracting subjective information from text, such as customer evaluations11,12. The DL approach’s effectiveness and repute have made it more and more practicable. This study’s use of DL method to classify internet reviews into positive and negative sentiment categories offered a thorough description of the semantics of customer reviews. The analysis was conducted using Amazon cells and associated products from the Snap dataset13,14. Natural language processing now includes sentiment analysis as a crucial subfield, with uses in product reviews, social media monitoring, and customer feedback analysis. Gaining insight into the emotions expressed in textual data can help predict future market trends, user satisfaction, and public opinion. Even with great progress made in sentiment analysis techniques, there are still difficulties in interpreting the finer points of human language, especially in niche markets such as e-commerce15,16. To overcome these issues, this study suggests a novel sentiment analysis approach that is adapted to the special features of e-commerce. The proliferation of user-generated material and online communication platforms has led to an increased demand for strong sentiment analysis algorithms. Since sentiments are frequently entangled with vocabulary and context unique to the e-commerce industry, existing sentiment analysis approaches might not offer reliable insights in this setting. The goal of this project is to provide a framework that can effectively handle the convoluted language around e-commerce17.

An examination of current sentiment analysis techniques reveals a dependence on conventional methods like machine learning classifiers and bag-of-words models. These techniques might not work as well in e-commerce as they do in broader areas. Enhancing existing models for more accurate sentiment analysis is now possible because to recent developments in complex pattern recognition, dimensionality reduction, and semantic analysis18,19. By combining cutting-edge methods like Word2Vec for semantic understanding, Truncated Singular Value Decomposition (SVD) for dimensionality reduction, and Restricted Boltzmann Machine (RBM) for capturing complex patterns, the proposed system seeks to create a state-of-the-art sentiment analysis model for e-commerce20. Enhancing semantic understanding, resolving dimensionality issues, and utilizing RBM to capture complex sentiment patterns are among the goals. It is anticipated that the suggested sentiment analysis methodology will significantly advance e-commerce by offering a deep comprehension of emotions in textual data21. The results will provide precise and useful insights to businesses, governments, and scholars, enabling well-informed decision-making in the e-commerce industry.

Customer assessments were gathered for this study from a variety of sources, such as reviews, tweets, postings, and comments. To systematically extract pertinent data, particular search parameters were created, taking into account platform-specific characteristics, time frames, and keywords. The requirement for domain-specific relevance led to the choice to gather new data rather than utilize an already-existing benchmark dataset22. The goal of broadening the search parameters to encompass a variety of data sources was to offer a thorough grasp of consumer attitudes on various platforms. This work seeks to develop an efficient way for assessing consumer feedback on e-commerce products using the Weighted Parallel Hybrid Deep Learning (WPHDL-SAEPR) approach23. Several pre-processing processes are required for the WPHDL-SAEPR approach in order to guarantee data compatibility. Word embedding is done with the Word2Vec model, while sentiment categorization is done with the WPHDL model, which combines the Singular Value Decomposition (SVD) and Restricted Boltzmann Machine (RBM) models24. A dataset of customer reviews was used to analyze the simulation results obtained from the WPHDL-SAEPR approach.

Related research works

In the field of e-commerce, reviews have been analyzed using techniques like Long Short-Term Memory (LSTM) and Deep Learning Convolutional Neural Networks (DL-CNN) in conjunction with LSTM (CNN-LSTM). Tokenization of lowercase letters, stop word removal, and punctuation removal are some of the data pre-processing techniques used for data cleansing. After that, the clean dataset is examined using the CNN-LSTM and LSTM algorithms to categorize consumer sentiment as either favorable or unfavorable. For e-commerce product reviews, Liu et al.19 claim that Bert-BiGRU-Softmax’s deep learning technique, which combines hybrid masking, an attention mechanism, and sentiment modeling, works well. As an input layer, the Bert technique obtains semantic codes, calculates sentiment weights, and extracts multi-dimensional product features. A hidden layer is produced by the Bidirectional GRU algorithm, and attitudes are classified as positive or negative by the Softmax layer, which is outfitted with an attention mechanism.

The ERNIE-BiLSTM-Att (EBLA) SA method is supported by Huang et al.21. In their work, dynamic word vectors are produced via the Enhanced Representation via Knowledge Integration (ERNIE) word embedding and fed into the BiLSTM for text property extraction. Emotion classification uses the softmax method, whereas the Attention Mechanism (Att) maximizes the weight of the hidden state25. A two-channel hybrid model for sentiment classification gave rise to the MWVH technique, which aggregates the properties of several word vectors into a single set. Built with attention techniques, CNN and RNN each have their own channels. Better sentence vector quality are produced by this two-channel model, which makes up for the shortcomings of single network techniques26. In order to examine different textual features, the study also looks into previously developed LSTM techniques. An ensemble approach with LSTM-NN is proposed to transfer useful information in sentiment or opinion categorization. It uses both character-level and word-level tokenization. For effective prediction tasks, this hybrid network integrates CNN and BiLSTM. While a CNN makes predictions based on the engineering features, BiLSTM evaluates comments and product attributes27,28.

An ensemble machine learning technique is applied to create a sentiment analyzer for several categories of Amazon product reviews. Four distinct datasets are used using five machine learning techniques: Random Forests (RF), various trees, bagging, stochastic gradient boosting, and AdaBoost. The datasets are books, electronics, DVDs, and kitchenware. First, customer sentiments are analyzed using the NB approach, then user attitudes are divided into binary groups using SVM29,30. The dataset is then pre-processed for terms (TF) and inverse document frequencies (IDF) in order to assess features. Hierarchical spatial characteristics are well-suited for Convolutional Neural Networks (CNNs), whereas Random Forests (RF) offer reliable decision-making. When these models are integrated, feature learning and ensemble-based decision techniques are used to improve overall prediction performance31. CNNs are skilled at image-based tasks because they are able to extract features that are essential for precise prediction by capturing spatial hierarchies in data. In particular, when dealing with noisy data, RF’s ensemble method and random feature selection lead to a more stable and accurate model32. Its robustness and resistance to overfitting make it a great complement to CNNs. The Matthews Correlation Coefficient (MCC) is a complete statistic that takes into account false positives, false negatives, true positives, and false negatives33. This ensures a fair evaluation of the model’s performance, which is particularly important in situations involving binary classification. A powerful solution for a variety of machine-learning problems is provided by combining CNN for complicated feature learning, RF for group decision-making, and MCC for nuanced evaluation. This combination balances model complexity, interpretability, and resilience34.

Resolving the class gap is essential to the model’s efficacy. Dataset representation can be balanced by employing strategies like under sampling the dominant class and oversampling the minority class. Algorithms that incorporate class weight changes or resampling techniques during training improve overall performance and fairness in classification problems by strengthening their robustness and accuracy in predicting instances of minority classes35. Sustained model efficacy requires regular evaluation and adjustment of the balanced approach based on changing data dynamics.

The proposed model

The purpose of this research is to build a unique WPHDL-SAEPR algorithm for the purpose of classifying consumers’ feelings about online shopping product reviews. The WPHDL-SAEPR system is designed to successfully identify a variety of feelings that are expressed in evaluations of online retailers. It incorporates a number of processes, including data collecting, pre-processing, feature extraction, word embedding, and sentiment classification, among others. An illustration of the WPHDL-SAEPR approach’s overall flow may be seen in Fig. 1. Several tactics can be used to optimize the WPHDL-SAEPR model’s performance while preserving efficiency in order to manage its complexity in real-world applications. To start, preparing the data is essential to make sure the input is tidy, organized, and appropriate for the model. Text input is streamlined utilizing methods like tokenization, stop-word removal, and lemmatization prior to word embedding using the Word2Vec model.

Parallel processing is used to control the model’s complexity, which lowers computing load by enabling simultaneous operation of elements such the Singular Value Decomposition (SVD) and Restricted Boltzmann Machine (RBM). Furthermore, regularization methods like L2 regularization and dropout help avoid overfitting, guaranteeing that the model may still be applied to new data. Finally, tweaking hyperparameters is crucial to maximize model performance. It is possible to optimize the design to produce better outcomes and faster convergence by using grid search or random search approaches. By breaking up workloads into more manageable, smaller subtasks, the parallel hybrid design of the model ensures better scalability and effective use of computational resources.

Data collection

The initial step is to compile review data from a foundational benchmark dataset before applying the approach. Information was gathered from a number of sources, such as social media posts, reviews, tweets, and comments. Certain search parameters were set for themes and customer reviews before data extraction. Product reviews, Facebook posts, news feeds, and Twitter tweets are among the frequently used data sources. After extraction, the data analysis and mining system processes the data. In order to collect data for this study, a thorough method was used to obtain consumer assessments from a variety of sources, such as reviews, tweets, postings, and comments. To guarantee methodical and repeatable data extraction, precise search parameters were carefully established, taking into account keywords, time frames, and platform-specific requirements. The requirement for domain-specific relevance and granularity that complement the distinct focus of our study led us to choose to gather new data rather than utilizing an already-existing benchmark dataset like Snap. It’s possible that benchmark datasets don’t fully convey the subtleties and complexity present in our study goals.

In addition, the search was broadened beyond its original boundaries to include a variety of data sources, including news feeds, Facebook posts, product reviews, and Twitter tweets. With this extension, we hope to cover a wider range of user-generated content and offer a more thorough picture of customer sentiment on several platforms. The dataset is enhanced by this multi-source approach, which provides insights into customer opinions and experiences from several angles.

Data pre-processing

An important first step in assessing the textual components of the data is pre-processing it. Text data collection has been more difficult recently because of repetitions and redundancies in tweets, blogs, reviews, and other types of communication. The act of sifting and polishing this pre-processed data is called “data normalization”. Pre-processing the data entails a number of procedures, including tokenizing words, deleting stop words, and removing unnecessary spaces. During text pre-processing, stop words, usually referred to as “stop phrases,” are frequently used keywords that are frequently eliminated from the data. The semantic significance of words like “the,” “and,” and “is” is limited, and they can cause noise in tasks involving natural language processing. Padding, eliminating hashtags, and changing text data to lowercase are additional procedures. Several specific tasks are completed in order to produce information of the appropriate caliber.

Fig. 1
figure 1

Overall flow of WPHDL-SAEPR approach.

Full size image

Normalization in pre-processing

Normalization

In particular, in algorithms sensitive to input magnitudes, normalization guarantees constant scales across features, preventing some characteristics from predominating over others during model training. It frequently speeds up the iterative optimization methods’ convergence, facilitating quicker training and more effective model learning. Normalization improves the model’s capacity to generalize patterns by lessening the effect of different scales, which leads to improved performance on data that has not yet been observed. By maintaining numbers within a same range, it lessens numerical instability and avoids problems like overflow or underflow during computations. Normalized data also makes data easier to interpret and makes it easier to compare various aspects meaningfully, which helps people realize how important each feature is in relation to the others.

Tokenization

Tokenization is a critical step in the NLP pipeline that is necessary for sentiment analysis and semantics in lexical assessment. It is impossible to create a model without first cleaning the text. Word tokenization and phrase tokenization are the two classes under which tokenization falls. You can count the number of words and their frequency in the text using the tokenized form.

Steps for normalization

Normalization decreases the dimensionality of the data and eliminates text variances without changing the meaning. The process involves cutting punctuation, changing the text’s case, lemmatizing or stemming, and eliminating stop words.

Eliminating stop phrases

Stop words are commonly found in text files, but they don’t add anything to the text’s sense. Their existence makes text mining more difficult and produces inconsistent classifier output. At this point, eliminating stop words reduces text volume while increasing the efficacy of the model.

POS labeling

In text pre-processing, extraneous characters, symbols, and noise are removed from text data by loading and cleaning. NLP libraries like NLTK, SpaCy, or Stanford NLP are used to implement POS tagging after the text has been tokenized into discrete words or phrases. Every token has a unique POS tag (noun, verb, adjective, etc.) attached to it. POS tagging gathers diverse opinions and features from product reviews and classifies words in a dataset according to predetermined grammatical forms. Parts of speech are identified by tags such as proper noun (P), noun (N), adjective (ADJ), verb (V), and article (DET). Furthermore, POS labeling marks relevant and significant candidates as issues.

Rooting

In product reviews, stemming is utilized to improve SA performance. Words are reduced to their most basic form so that the model can more accurately classify sentiment and convey the meaning of the review. One can reduce terms like “running,” “runner,” and “ran” to the root “run.”

Lemmatization

In product reviews, lemmatization is used to boost SA performance. By breaking down words into their most basic form, the algorithm improves its capacity to classify sentiment and grasp the meaning of the review. Lemmatization, as opposed to stemming, preserves more context and original word features, which may improve NLP task performance.

Feature extraction and word embedding

The word embedding procedure used the Word2Vec model. The acquired dataset is unsuitable for direct mathematical or statistical computations, hence an appropriate function encoding procedure is required to extract mathematical properties from the text data. To correctly reflect each review and capture the precise semantic meaning of words or sentences, a mathematical model is required. After that, the numerical features that were derived were used for processing and analysis.

N-grams

N-grams, which are tokenized text sequences of ‘n’ words in a sequential manner, are used to extract context. They can be anything from single words, or unigrams, to two-word sequences, or bigrams, and more. N-grams improve natural language processing (NLP) tasks such as entity recognition, machine translation, and text categorization by capturing the local context of words. By taking word context into account, n-grams enhance the model’s performance and enable it to more accurately convey the sentiment and meaning of the text. For example, a bigram like “not impressed” can express feelings more clearly than a single word.

TF-IDF

The TF-IDF approach assesses a word’s significance in relation to the review’s overall emotion. TF-IDF augments the sentiment analysis algorithm by taking into account a term’s rarity across all reviews as well as its frequency in a given review. By weighting informative terms that signal sentiment, TF-IDF is frequently employed in machine learning models for tasks like text categorization and sentiment analysis, improving model performance.

Word-to-Vec

Word2Vec is a vectorization tool that examines and computes word relationships. It employs techniques such as continuous skip-gram approaches and Continuous Bag of Words (CBOW). Word2Vec preserves the semantic connections between words while it transforms them into vectors. It can distinguish between brief phrase semantic similarities and examine settings like nearby bus or subway stops or news headlines. Words are one-hot vectors in Word2Vec, both in the input and output layers. The output layer is a softmax layer that produces a probability distribution for every word in the dictionary, while the first layer, an FC layer, is devoid of activation functions. While CBOW uses context word vectors to create the expression vector for the target word, skip-gram guesses context from input word vectors.

The correlation coefficient of Matthews (MCC)

A metric for binary classification that takes into account true positives, true negatives, false positives, and false negatives is the Matthews Correlation Coefficient. This allows for a fair assessment of the performance of the model.

Sentiment classification

This work uses the WPHDL model, which combines RBM and SVD models, for sentiment classification. Through the utilization of the complementary properties of both Singular Value Decomposition (SVD) and Restricted Boltzmann Machine (RBM), our study introduces a hybrid SVD-RBM sentiment analysis model. The trade-offs between model complexity, interpretability, and performance must be carefully considered, particularly in light of more straightforward models like TF-IDF, which obtains a performance of 96.09%. With a performance of 98.76% utilizing WPHDL-SAEPR, our model shows encouraging results; still, it is crucial to evaluate these trade-offs. The hybrid SVD-RBM outperforms TF-IDF by a margin of about 2.67%, which begs the question of whether switching to a more complicated model is justified. It makes us question if, given the small performance advantage, the hybrid model’s increased complexity is necessary. Less complex models, like TF-IDF, frequently provide a feature space that is easier to parse, which facilitates understanding of the variables affecting sentiment predictions. On the other hand, the hybrid SVD-RBM could add complexity that makes the model harder to understand and less useful in real-world situations.

Complex mechanisms and interactions between SVD and RBM components are involved in the hybrid SVD-RBM. Although increased complexity could help the model capture subtle patterns, it also makes the model harder to understand and might necessitate additional processing power. Consideration should be given to the practical implications, computational efficiency, and practical applicability when selecting a more complex model. Adopting such a model may not be as warranted if the marginal increase in performance is uncorrelated with the added complexity.

A thorough evaluation of the numerous trade-offs involved ought to direct the choice between more complex models, like the hybrid SVD-RBM, and less complex options. This entails assessing the results’ interpretability and the marginal gain in performance in relation to the model’s complexity. Our findings underscore the significance of upholding transparency in the reporting of these matters in order to furnish a comprehensive comprehension of the pragmatic ramifications of model selection. SVD is a useful technique for dimensionality reduction21. SVD’s primary difficulty lies in producing a low-rank approximation. Given a m×n matrix A of rank r, the following is a description of the SVD(A):

$$\:SVD\left(A\right)=U\times\:S\times\:{V}^{T}\:$$
(1)

In Eq. (1), \(\:U\) and \(\:V\) denote corresponding orthogonal matrices with dimensions \(\:m\)×\(\:m\) and \(\:n\)×\(\:n\). Matrix \(\:S\) refers to a diagonal matrix, a singular matrix with dimensions \(\:m\)×\(\:n\), with a nonnegative real number. The set of \(\:r\) initial values of \(\:S\)\(\:({s}_{1},\:{s}_{2},\dots\:\:,\:{s}_{r})\) are positive with \(\:{s}_{1}\ge\:{s}_{2}\ge\:{s}_{3},\dots\:,\ge\:{s}_{r}.\) first \(\:r\) column of \(\:V\) is the eigenvector of \(\:{A}^{T}A\) and signifies the right singular vector of A. Likewise, the first \(\:r\) column of \(\:U\) is AAT’s eigenvector and characterizes A’s left singular vector. \(\:SVD\) provides the optimal low-rank approximation of \(\:A\). This can be attained by removing \(\:r-k\) rows from \(\:V,\) retaining the first \(\:k\) diagonal value, and eradicating \(\:r-k\) columns from \(\:U\:\) as follows:

$$\:{A}_{k}={U}_{k}\times\:{S}_{k}\times\:{V}_{k}^{T}\:\:\:$$
(2)

The regenerated matrix \(\:{A}_{k}\) is the approximation nearest to \(\:A\). The \(\:k\)-rank approximation of matrix \(\:A\) about the Frobenius norm is shown in the following equation:

$$\:\left|\right|A-{A}_{k}|{|}_{F}^{2}={\sum\:}_{i,j}({a}_{ij}-{\sum\:}_{k}{U}_{ik}\times\:{S}_{k}\times\:{V}_{kj}^{T})\:\:\:$$
(3)

Prediction generation based on \(\:SVD\):

In Eq. (4), The prediction can be made by calculating the cosine similarity (dot product) among \(\:m\) pseudo-customer \(\:{U}_{ik}.\sqrt{{S}_{k}^{T}}\) and \(\:n\) pseudo‐product \(\:{S}_{k}.\sqrt{{V}_{k}^{T}},\:\)when \(\:m\)×\(\:n\) rating matrix \(\:R\) is decomposed and minimized into three \(\:SVD\) component matrices with \(\:k\) features \(\:{U}_{k},\) Sk, and \(\:{V}_{k}\). In particular, the prediction score \(\:{R}_{i,j}\) for \(\:{i}^{th}\) customer on \(\:{j}^{th}\) product is obtained by adding the row average \(\:\overline{{r}_{\iota\:}}\) to the similarity.

$$\:{R}_{i,j=}\overline{{r}_{\iota\:}}+{U}_{k}.\sqrt{{S}_{k}^{T}}\left(i\right)\:+{s}_{k}.\sqrt{{V}_{k}^{T}}\left(i\right)\:$$
(4)

A dot product calculation is included in the forecasting process when SVD is implemented. This computation takes 0(1) times to complete since k is constant. A two-layer stochastic neural network (NN) that consists of both visible and hidden units is the RBM. As can be seen in Fig. 2, it is composed of a bias unit, a single layer of visible unit (user preference), and a single layer of hidden unit (the latent factor). The state of the bias unit may often be on, and it is modified to account for the intrinsic popularity of all items. In extra, every visible unit is networked with every hidden unit in an undirected form:-, consequently, every hidden unit is interconnected with each visible unit and the bias unit is interconnected with each hidden and visible unit. The network was restricted in such a way that there was no link between the hidden levels and the visible layers, which made learning more straightforward.

The stochastic, binary visible unit is responsible for encoding the user’s choice for the item coming from the training dataset. It is thus possible to recognize the status of all the visible components22. A stochastic binary variable that is capable of capturing latent properties is sometimes referred to as the Hidden unit. Through the use of the energy function, the network assigns a probability to each and every possible pair of the visible and hidden vector:

$$\:p\left(\text{v},h\right)=\frac{{e}^{-E\left(v,h\right)}}{Z}\:$$
(5)

In Eq. (5), \(\:E\) denotes the system energy, and \(\:Z\) represents the normalizing factor. Hinton proposed a contrastive divergence method for weight training.

Fig. 2
figure 2

Structure of RBM.

Full size image

There are \(\:M\) products, \(\:N\) customers, and integer ratings on a scale of 1 to \(\:K\). First, product ratings effectively deal with many missing ratings. The visible unit corresponds to products the customer does not rate simply do not exist. The visible unit is often turned off; thus, its state is often zero. On the other hand, it is evident that using a unique RBM for each customer-, shares a similar unit; however, each contains the softmax unit for the products rated by the customer. The visible binary unit can be replaced with a softmax unit. For every customer, the RBM includes a softmax unit for the product rated by that customer. Consider user u-rated m products and visible unit V as a K×m matrix so that \(\:{\nu\:}_{i}^{k}=1\) if u rated product i as k and 0 if not. Assume \(\:{h}_{j},j=1,\dots\:\dots\:,\:F\), as the binary value of the hidden unit of the user features. The hidden latent feature h models columns of V and multinomial (softmax) are modeled similarly to Eqs. (4) and (5), respectively.

$$\:p\left({\nu\:}_{i}^{k}=1|h\right)=\frac{\text{exp}\left({b}_{i}^{k}+{\varSigma\:}_{j=1}^{F}{h}_{j}{W}_{ij}^{k}\right)}{{\varSigma\:}_{l=1}^{k}\text{exp}\left({b}_{i}^{l}+{\varSigma\:}_{j=1}^{F}{h}_{j}{W}_{ij}^{l}\right)}\:$$
(6)
$$\:p\left({h}_{j}=1|h\right)=\frac{1}{1+\text{exp}\left(-{b}_{j}-{\varSigma\:}_{i=1}^{m}{\varSigma\:}_{k=1}^{K}{v}_{i}^{k}{W}_{ij}^{k}\right)}\:\:$$
(7)

Here, \(\:{W}_{ij}^{k}\) denotes the weight connecting the rating \(\:k\) of product \(\:i\) and hidden unit \(\:j,\)\(\:{b}_{i}^{k}\) represents the bias of rating \(\:k\) for product \(\:i\) and \(\:{b}_{j}\) bias term of hidden unit \(\:j\):

$$\:p\left(V\right)={\sum\:}_{h}\frac{\text{exp}\left(-E\left(V,h\right)\right)}{{\varSigma\:}_{{V}^{{\prime\:}},h},\text{exp}\left(-E\left({V}^{{\prime\:}}{,}^{k}\right)\right)}\:$$
(8)

Using the energy term given by:

$$\:E\left(\text{V},h\right)=-{\sum\:}_{i=1}^{m}{\sum\:}_{j=1}^{F}{\sum\:}_{k=1}^{K}{W}_{ij}^{k}{h}_{j}{\nu\:}_{i}^{k}-{\sum\:}_{i=1}^{m}{\sum\:}_{k=1}^{K}{\nu\:}_{i}^{k}{b}_{i}^{k}-{\sum\:}_{j=1}^{F}{h}_{j}{b}_{j}$$
(9)

The product with the missing rating doesn’t contribute to the energy function. Hybrid weighting is a hybridization method that calculates the predictive score of each recommendation method by providing weight to every system and summing the weight to generate a novel output recommendation. The innovative fusion of Restricted Boltzmann Machine (RBM) and Singular Value Decomposition (SVD) introduces a powerful hybrid model. This novel approach optimally balances complexity and efficiency by leveraging RBM’s ability to capture intricate patterns and SVD’s dimensionality reduction. The integration promises improved performance in capturing nuanced semantics and reducing computational overhead, making it well-suited for applications demanding accuracy and scalability. Dynamic weighting and empirical bootstrapping are two different ways to define consequences. The RBM and SVD methods are compiled using the weighted method, and the predictive score of user u for item i is calculated using Eq. (10):

$$\:re{c}_{we\text{i}\text{g}\text{h}ted}\left(\text{u},i\right)={\sum\:}_{k=1}^{n}{\beta\:}_{k}\times\:{R}_{k}\left({u}_{t}i\right)\:$$
(10)

Let \(\:{\beta\:}_{k}\) be the weight of algorithm \(\:{R}_{k}\left({u}_{t}i\right)\). In this study, weighted hybridization integrates two approaches; therefore, we set \(\:n=2\). Consequently, if \(\:n=2\), the prediction score is computed as:

$$\:re{c}_{we\text{i}\text{g}\text{h}\text{t}\text{e}\text{d}}\left(u,\:i\right)={\beta\:}_{1}\cdot\:{R}_{1}\left(u,\:i\right)+\left(1-{\beta\:}_{1}\right)\cdot\:{R}_{2}\left(u,\:i\right)$$
(11)

optimized weight is obtained by calculating:

$$\:\beta\:\_1=(\varSigma\:\_u\:\varSigma\:\_i\:(re\_weighted\:(u,i)-R\_2\:(u,i)\left)\right(R\_1\:(u,i)-R\_2\:(u,i)\left)\right)/(\varSigma\:\_u\:\varSigma\:\_i\:(R\_1\:(u,i)-R\_2\:(u,i))^2\:)\:$$
(12)

Finally, the weighted parallel hybrid models provide correct predictions for personalized recommendations. Personalized recommendations enhance user experience by tailoring content to individual preferences, increasing user engagement and satisfaction. Leveraging user behavior data, the model learns and adapts, delivering more relevant suggestions. This approach fosters user loyalty, as individuals feel a stronger connection to content aligned with their unique tastes. The personalization strategy optimizes content discovery, making the user journey more enjoyable and increasing the likelihood of successful interactions. Ultimately, personalized recommendations create a more user-centric platform that caters to diverse preferences and improves overall system performance. The cold-start issue can be mitigated using the weighted hybridization forecasting scores of the two approaches, combining the benefits of RBM and SVD.

Our study initially outlined a sentiment classification task with three classes: positive, neutral, and negative, as mentioned in the abstract. However, upon closer examination, it becomes apparent that our model is configured for binary sentiment classification, focusing on distinguishing between positive and non-positive sentiments. The model outputs sentiment predictions that can be either positive or non-positive. The decision to simplify the sentiment classification task was driven by practical considerations, including dataset characteristics and the focus on distinguishing positive sentiments from others. This clarification is pivotal in aligning the readers’ expectations with the model’s design. In the context of our study, the mention of “weighted parallel hybrid models providing correct predictions for personalized recommendations” refers to the model’s ability to make accurate sentiment predictions tailored to the individual user’s preferences. This personalization is achieved by incorporating personalized features, including historical user interactions, preferences, or other context-specific information. While the core of our sentiment analysis model is designed for binary classification, integrating personalized features enables a level of customization in sentiment predictions. These personalized features are incorporated into the model’s architecture, allowing it to adapt its predictions based on individual user characteristics.

The notion of personalized recommendations in the context of sentiment analysis has practical implications for applications where user-specific sentiment understanding is crucial. It allows the model to surpass generic sentiment predictions and adapt to individual users’ unique preferences and context. This revision addresses the inconsistency in sentiment classification, clarifies the binary nature of the model’s output, and explains how personalized recommendations are incorporated into the model’s predictions. It aims to offer a clear and coherent understanding of the model’s design and its implications.

Results and discussion

Critical gaps in natural language processing tasks are addressed by combining the usage of Word2Vec, Singular Value Decomposition (SVD), and Restricted Boltzmann. Word2Vec’s Semantic Embeddings: By embedding words into vector spaces, Word2Vec captures subtle contextual linkages and improves the model’s knowledge of word semantics. SVD optimizes model efficiency and interpretability, particularly when handling big textual datasets, by shrinking the feature space while preserving pertinent information. This helps to address high-dimensionality problems. RBM’s modeling of hidden representations allows it to capture complex patterns, which enhances the model’s ability to learn complicated hierarchical structures, especially in unsupervised learning scenarios. A comprehensive strategy for effective feature learning is ensured by combining Word2Vec, SVD, and RBM. Together, Word2Vec and SVD reduce dimensionality, RBM captures complicated patterns, and together they handle a variety of representation learning difficulties. Word2Vec, SVD, and RBM integration improves textual understanding by providing a complete solution that fills in the gaps in complex pattern recognition, dimensionality, and semantic representation. The model’s capacity to glean insightful information from textual input for a range of natural language processing applications is improved by this synergistic approach. Two datasets, as shown in Table 1, are used to assess the experimental validation of the WPHDL-SAEPR technique: the television products dataset (dataset-2) and the cell phones and accessories dataset (dataset-1).

Table 1 Details of the dataset.
Full size table
Fig. 3
figure 3

Confusion matrices of WPHDL-SAEPR approach on dataset-1 (a,b) TRP/TSP of 80:20 and (c,d) TRP/TSP of 70:30.

Full size image

The confusion matrices of the WPHDL-SAEPR method for dataset-1 are shown in Fig. 3. The results indicate that the WPHDL-SAEPR technique effectively identifies the positive and negative samples on the applied dataset-1 under different sizes of the training phase (TRP) and testing phase (TSP). Table 2 reports the average sentiment classification results of the WPHDL-SAEPR technique on dataset 1. The experimental values inferred that the WPHDL-SAEPR technique effectively categorized the positive and negative samples. For 80% of TRP, the WPHDL-SAEPR technique reaches an average \(\:{accu}_{bal}\) of 99.51%, \(\:{Prec}_{n}\) of 97.96%, \(\:rec{a}_{l}\) of 98.39%, \(\:{F}_{score}\) of 98.18%, and an MCC of 96.35%. Meanwhile, for 20% TSP, the WPHDL-SAEPR method reaches an average \(\:accuracy\) of 98.76%, \(\:precission\) of 97.67%, \(\:rec{a}_{ll}\) of 98.76%, \(\:{F}_{score}\) of 98.21%, and an MCC of 96.43%.

Table 2 Sentiment classifier outcomes of WPHDL-SAEPR approach on dataset-1.
Full size table
Fig. 4
figure 4

TACY and VACY outcome of WPHDL-SAEPR approach on dataset-1.

Full size image

The TACY and VACY of the WPHDL-SAEPR model on dataset-1 are represented in Fig. 4. The results showed that the WPHDL-SAEPR model exhibited enhanced performance with higher values of TACY and VACY.In particular, the WPHDL-SAEPR model showed maximal TACY outcomes. The TLOS and VLOS of the WPHDL-SAEPR approach on dataset-1 are represented in Fig. 5. The figure indicates that the WPHDL-SAEPR methodology performs better with the lowest values of TLOS and VLOS. It is evidence that the WPHDL-SAEPR model reduced VLOS outcomes.

Fig. 5
figure 5

TLOS and VLOS outcome of WPHDL-SAEPR approach on dataset-1.

Full size image
Fig. 6
figure 6

Confusion matrices of WPHDL-SAEPR approach on dataset-2 (a,b) TRP/TSP of 80:20 and (c,d) TRP/TSP of 70:30.

Full size image

The confusion matrices of the WPHDL-SAEPR method for dataset-2 are shown in Fig. 6. The results indicate that the WPHDL-SAEPR approach effectively identified the positive and negative samples on the applied dataset-2 under different sizes of TRP and TSP. Table 3 shows the average sentiment classification result of the WPHDL-SAEPR technique on dataset 2. The experimental values inferred that the WPHDL-SAEPR technique effectively categorized the positive and negative samples. On 70% of TRP, the WPHDL-SAEPR technique reaches average \(\:acc{u}_{bal}\) of 92.87%, \(\:pre{c}_{n}\) of 95.71%, \(\:rec{a}_{l}\) of 92.87%, \(\:{F}_{score}\) of 94.23%, and an MCC of 88.54%. Meanwhile, for 30% TSP, the WPHDL-SAEPR technique reached an average \(\:acc{u}_{bal}\) of 92.93%, \(\:pre{c}_{n}\) of 95.72%, \(\:rec{a}_{l}\) of 92.93%, \(\:{F}_{score}\) of 94.26%, and MCC of 88.60%.

Table 3 Sentiment classifier outcomes of WPHDL-SAEPR approach on dataset-2.
Full size table
Fig. 7
figure 7

TACY and VACY outcome of WPHDL-SAEPR approach on dataset-2.

Full size image
Fig. 8
figure 8

TLOS and VLOS outcome of WPHDL-SAEPR approach on dataset-2.

Full size image
Fig. 9
figure 9

Precision-recall outcome of (a,b) 80:20 and 70:30 on dataset-1 and (c,d) 80:20 and 70:30 on dataset-2.

Full size image

The TACY and VACY of the WPHDL-SAEPR model in dataset-2 are shown in Fig. 7. The Figure shows that the WPHDL-SAEPR model has improved performance with higher values of TACY and VACY. In particular, the WPHDL-SAEPR model had better TACY outcomes. The TLOS and VLOS of the WPHDL-SAEPR model on dataset 2 are represented in the Fig. 8. The Figure indicates that the WPHDL-SAEPR method performs better with the lowest values of TLOS and VLOS. Specifically, the WPHDL-SAEPR model reduced the VLOS outcomes. A detailed precision-recall investigation of the WPHDL-SAEPR algorithm on the two databases is shown in Fig. 9. The Figure shows that the WPHDL-SAEPR approach has improved precision-recall values in every class label.

Fig. 10
figure 10

ROC outcome of (a,b) 80:20 and 70:30 on dataset-1 and (c,d) 80:20 and 70:30 on dataset-2.

Full size image

A detailed ROC analysis of the WPHDL-SAEPR system using the two databases is shown in Fig. 10. The results show that the WPHDL-SAEPR method can classify distinct classes. Table 4 comprehensively compares the SA results of the WPHDL-SAEPR technique with those of recent models. Figure 11 presents a comparative study of the WPHDL-SAEPR technique regarding \(\:acc{u}_{racy}\) and \(\:{F}_{score}\). Among the various methods, the WPHDL-SAEPR technique increased the values of \(\:acc{u}_{racy}\) and \(\:{F}_{score}\). Based on \(\:the\:acc{u}_{racy}\), the WPHDL-SAEPR technique reaches a higher \(\:acc{u}_{racy}\) of 98.76%, whereas the DNN-LSTM, DNN-RNN, hybrid DNN, BoW, N-gram, and TF-IDF models attained a lower \(\:acc{u}_{racy}\) of 97.05%, 96.88%, 96.83%, 96.16%, 96.18%, and 96.09% respectively. Meanwhile, based on \(\:{F}_{score}\), the WPHDL-SAEPR method achieved a higher \(\:{F}_{score}\) of 98.21%, whereas the DNN-LSTM, DNN-RNN, hybrid DNN, BoW, N-gram, and TF-IDF methods attained lower \(\:{F}_{score}\) of 97.02%, 97.26%, 97.37%, 96.44%, 97.27%, and 97.21% respectively.

Table 4 Comparative outcome of WPHDL-SAEPR approach with recent systems.
Full size table
Fig. 11
figure 11

\(\:\varvec{A}\varvec{c}\varvec{c}{\varvec{u}}_{\varvec{y}}\) and \(\:{\varvec{F}}_{\varvec{s}\varvec{c}\varvec{o}\varvec{r}\varvec{e}}\) outcomes of the WPHDL-SAEPR approach with recent systems.

Full size image
Fig. 12
figure 12

\(\:\varvec{P}\varvec{r}\varvec{e}{\varvec{c}}_{\varvec{n}}\) and \(\:\varvec{r}\varvec{e}\varvec{c}{\varvec{a}}_{\varvec{l}}\) outcomes of the WPHDL-SAEPR approach with recent systems.

Full size image

Figure 12 presents a detailed study of the WPHDL-SAEPR technique regarding \(\:pre{c}_{n}\) and \(\:rec{a}_{l}\). Among the various methods, the WPHDL-SAEPR algorithm increases values of \(\:pre{c}_{n}\) and \(\:rec{a}_{l}\). Based on \(\:pre{c}_{n}\), the WPHDL-SAEPR method reaches a higher \(\:pre{c}_{n}\) of 97.67%, whereas the DNN-LSTM, DNN-RNN, hybrid DNN, BoW, N-gram and, TF-IDF models attain a lower \(\:pre{c}_{n}\) of 96.42%, 96.36%, 96.06%, 97.48%, 96.02%, and 96.20%, respectively. Meanwhile, based on \(\:rec{a}_{l}\), the WPHDL-SAEPR technique reaches higher \(\:rec{a}_{l}\) of 98.76% while the DNN-LSTM, DNN-RNN, hybrid DNN, BoW, N-gram and, TF-IDF models attain lower \(\:rec{a}_{l}\) of 96.74%, 96.25%, 96.43%, 96.72%, 96.23%, and 96.92% respectively. The analysis mentioned above confirmed the superior performance of the WPHDL-SAEPR technique over other recent methods.

Based on the above results and analysis, when it comes to sentiment analysis on e-commerce platforms, the WPHDL-SAEPR model has a number of advantages over conventional machine learning and deep learning models. The model leverages the advantages of several methods, including Singular Value Decomposition (SVD) for dimensionality reduction, Word2Vec for rich semantic word embeddings, and Restricted Boltzmann Machine (RBM) for feature learning, by employing a weighted parallel hybrid architecture. Combining these two makes the model better at capturing the intricate links that exist between words and sentiments, which improves sentiment classification accuracy. Furthermore, the WPHDL-SAEPR model’s parallel structure enhances scalability and minimizes computing time by processing huge datasets effectively. Better sentiment prediction and generalization are achieved by prioritizing more relevant features through the use of weighting algorithms. When it comes to accuracy, efficiency, and resilience, WPHDL-SAEPR performs better than traditional sentiment analysis models overall.

Conclusion

For the purpose of classifying sentiments in e-commerce product reviews, we introduced a novel Weighted Parallel Hybrid Deep Learning-based Sentiment Analysis and Classification Model (WPHDL-SAEPR) in this work. Efficiently identifying a variety of emotions conveyed in online shopping reviews is the main goal of the WPHDL-SAEPR technique. Using the Word2Vec model for word embedding, the suggested system combines multiple discrete pre-processing stages to convert the input into a format that is compatible. SVD and RBM models are combined in the WPHDL model for sentiment categorization. A consumer review dataset was used in the experimental validation of the WPHDL-SAEPR approach. The outcomes show that, for a variety of performance metrics, the WPHDL-SAEPR strategy performs better than other contemporary approaches. This suggests that the suggested method is successful in improving the accuracy of sentiment classification. Future work could lead to the development of multiple deep learning-based voting categorization models, which could further enhance sentiment classification performance. With these upcoming improvements, sentiment analysis and sentiment classification in e-commerce product reviews may be even more robust and accurate.