A novel approach to enhance the quality of health care recommender system using fuzzy-genetic approach

Abstract

In recent generations of the digital world medical data in Recommender Systems. Health Care Recommender System (HCRS) analyses the medical data and then predicts the user’s or patient’s illness. Nowadays, healthcare data is used by various users or patients in recommendation systems which are useful for everyone. Analysing and predicting medical data provides awareness to users and these data predictions may be enriched using various techniques of RS. Machine learning techniques are used to make sure that health data is reliable and of high quality. In every RS the issues are targeted such as scalability, sparsity and cold start problems. In many social networking applications, these issues are resolved using ML algorithms. However, there is a significant gap between IT systems and medical diagnosis. The fuzzy genetic method is used in HCRS in order to bridge the gap between IT and healthcare applications. Through the use of the mutation and crossover operators, a real-value genetic method is used in this to compute similarity. With the user’s extra personalized information, fuzzy rules are later generated for the database. The Hybrid fuzzy-genetic method, also known as this situation, combines both techniques to improve recommendation quality. Utilizing this method will improve the quality of the recommendation process by discovering the most precise similarity measures among different users. Six factors are subjected to fuzzification, including age, gender, employment, height, weight, and region. Genre-interesting measure weights are then used, including Very Light, Light, Average, Heavy, and Very Heavy. Finally, the evaluation metrics used MAE and RMSE to evaluate the recommendation accuracy which showed the best results in comparison with baseline approaches such as Convolutional Neural Networks and Restricted Boltzman Machine.

Keywords

Recommender system confidentiality deep learning convolutional neural networks fuzzy logics

1 Introduction

Recommender systems (RS) [2–4 , 14] are used to manage data and produce knowledge. The user is given recommendations by RS, which is classified as content-based, collaborative, hybrid, and demographic. A growing number of people today are using health care RS, which includes suggesting various hospitals, doctors, and disease prognostication, among other things. The symptom checker is used in all health care recommendation systems (HCRS) to produce some predictions of the users’ or patients’ diseases [39, 40]. HCRS provides suggestions based on the profiles and interests of users. The RS [34 , 51–53] has been Classified into Content-Based, Filtering (CBF), Collaborative Filtering (CF) [7 , 19f] and Hybrid Filtering (HF) [44]. In CBF, Users’ past behaviour is used for recommendation. whereas in CF similar users’ behaviour is considered for recommendation. In many applications, to predict the data both CBF and CF are needed. So, HF is considered when there is a need for past behaviour and similar users’ data. Although many health-related applications are using this RS, still it suffers from various issues.

Privacy. Patient information is crucial for HCRS analysis and illness prediction. The information is dispersed across numerous geographic places, though. It can be difficult to connect data from various regional sources. The privacy of users may have been violated as a result of multiple network sites sharing the data of numerous patients. HCRS is unable to accomplish data confidentiality due to privacy concerns. With regard to privacy, many academics have worked on HCRS. However, the computation cost of confidentiality in offline mode could not be targeted. The homographic encryption method is another name for the process of creating stored data in confidence [25 , 45]. to integrate patient data from online sources In [31] to address privacy concerns when computing the nearest neighbour optimisation. As a result, the privacy and accuracy of the proposed model were reached.

Trust and distrust Cold start user solutions. To improve the accuracy and quality of RS, recent study has concentrated on the same problem. Targeting the cold start user or item issue can be done using a variety of techniques. Deep learning technology is utilized for higher degrees of data representation [37, 43]. In the field of HCRS, social suggestion websites are very important. When predicting suggestions when there are many social relations, trust and distrust are taken into account. These trust-distrust relationships between users were discussed by Yuan, W. (2018) in [59] along with a comparison of current users. You can believe a user if they behave more like an active user than if they don’t. The basis for the confidence and distrust links is the similarity computation threshold. Utilizing SVD clustering, this relationship between confidence and mistrust is achieved. In order to compute the relationships between user trusts, the trust neighbours mining algorithm was suggested in this paper. A sparse rating complement algorithm was also suggested as a means of achieving the effectiveness of suggestions from reliable users, and this could increase prediction accuracy and coverage. Predictions and recommendations in the recommender engine are challenging to make because there is insufficient data about new users or products. While no information was accessible in CCS, only a limited amount of user information was present in ICS. In [56] and [12], Wei, J. et al. (2017) Banda L, Bharadwaj K.K. (2014) presented a framework for Collaborative Filtering (CF) which fix the cold start issues [24].

Sparsity solutions. Collaborative filtering faces a number of challenges, including the management of large amounts of data and the identification of missing records. Sparsity is the term used to describe an issue when there is fewer data in the user-item matrix [27]. This issue needs to be solved in order to accomplish the prediction accuracy or recommendation quality. In order to address the issues of sparsity and scalability, user profiles have been taken into account in this study In order to address the issues of sparsity, scalability, and cold start users associated with collaborative filtering as well as tagging data, the paper presented incremental clustering and trust [14].

Due to the problem of data sparsity, CF ultimately performs poorly and is unable to produce helpful suggestions. In [1], Althbiti et al. (2021) suggested a novel model that addresses the problem of data sparsity in CF by using clustering and artificial neural networks. Four distinct datasets from four well-known fields are used to assess the proposed model CANNBCF, which stands for Clustering and Artificial Neural Network Based Collaborative Filtering. (books, music, jokes, and movies). With greater prediction precision, this model resolves the sparsity issue.

Scalability solutions. HCRS manages a large amount of data, which is a challenging job. Prediction and recommendation become more challenging as the amount of user and object data in the database grows. The issue is also referred to as the scaling issue [13]. Shahabi, C. et al.’s (2001) suggestion model was presented in [47]. Yoda is intended for large-scale data from web-based applications. This combines content-based and collaborative filtering methods to increase accuracy [28]. FLHS improves the precision of recommendations. The algorithm of weighted slope fixes the problems to enhance the accuracy of recommendation [32]. The following sections present the literature review in section 2, the Experiments and Results section, the fuzzy genetic method is used in the proposed framework for HCRS. in section 3, the Discussion of the Experiments and Results in section 4, and lastly the Discussion of the Conclusion and Future Work in section 5.

2 Literature survey

For every user who participates in social recommendation sites, the present HCRS has developed into a crucial system. A component of decision support systems, HCRS analyses and predicts user or patient data [15]. Numerous scholars have worked on HCRS and have faced a variety of problems. For finding valuable information in HCRS, computer algorithms are used, and these algorithms can be used in other applications. In machine learning, there are numerous tools available to manage large amounts of unstructured or hidden data. Future classification and prediction are done for this using intelligent optimization methods. In [42], DL tools were introduced in the context of HCRS applications by Priyadarshini et al. (2018). Internet-of-Things (IoT) is a popular rising technology that is cost-effective in HCRS in addition to machine learning tools [41]. IOT includes wearing a smart device and utilizing numerous software programmes. IOT applications can be used for illness prediction, diagnosis, and healthcare-related data. Another illustration of IOT uses for HCRS is wearable smart devices. It is capable of handling complex and heterogeneous data. The diseases linked to stress, hypertension, and diabetes are the main topics of the paper. In [22], Dai, Y. & Wang, G. (2018) suggested a framework built on a deep learning module and presented a deep neural network to determine a patient’s current state of health. On the foundation of the Bayesian inference graph, the action assessment module was later created.

Fig. 1

Proposed Framework for HCRS using Fuzzy-genetic approach.

Patients can exchange private health data using the HCRS paradigm. A difficult job that might result in data overload is a prediction based on a user’s health information. In terms of the information that a recommender engine has about the healthcare data at its core. Due to a lack of knowledge, analysing patient data and conducting searches is very challenging. The HCRS data collection was used in [57] by Yang, C.C. and Jiang, L. The varied data that takes into account individuals, threads and their connections to the MedHelp database is used. Using machine learning algorithms. The work of some more researcher's Literature Survey shown in Table 1. Unstructured big data implementation cases totalling 26 have been found and mapped using it [54, 55] and these information shown in Table 2. The knowledge gap between healthcare organisations and apps needs to be closed.

Table 1

Literature survey on health care recommender systems

Author	Proposed work	Applied technique	Outcome
In [25], Gorski J. et al. (2022)	showed a system framework for a programme that recommends processing methods for condition monitoring.	components made of fuzzy logic that are integrated into the system	Achieved higher accuracy of 14% other than existing methods
In [17], Bushra Alhijawi et al. (2020)	BLIGA	GA applied	more precise forecasts than the alternatives.
In [50], Symeonidis, P. (2021)	research into suggestion strategies	data review linked to health care	The suggested HRS has numerous advantages for bettering health.
In [38], Ochoa, J.G.D., Csiszár (2021)	Using digital technologies, everyday operations can be carried out.	predicted results based on multi-criteria decision makers and continuous-valued reasoning..	outcomes mimic logical decision processes based on the hierarchy of pertinent physiological factors.
In [48], Sharma (2022)	big data analytics, Machine Learning (ML), Neural Network (NN)	suggested a deep learning-based model.	to improve multilevel decision-making in relation to health risk and the severity of various illnesses.
In [8], Alsalem, M.A (2022)	This investigation will present an overview of the current state of MCDM for COVID-19 to MCDM scholars and the larger community.	Fuzzy weighted zero, inconsistent weighting to address the drawbacks of earlier weighting methods, and the MCDM strategy	the Fuzzy Decision by Opinion Score technique, which assigns a unique ranking to each alternative and prioritises them.
In [9], Arian Hafezalkotob (2019)	To make multi-criteria decision-making technique	is MULTIMOORA.	Achieved Accuracy

Table 2

Personalized information of users

u1	Age	Sex	Occupation	Height	Weight	Region	Rating	Symp1	. . .	Symp19
#12	32	F	Teacher	5’6’’	56	Delhi	4	1	0
#14	45	M	Scientist	5’10”	75	Orissa	3	0	1
#18	67	M	Navy	5’11”	77	Jammu	2	1	1

The model notify users of the need for routine check-ups via email, SMS, and online advertisements.

The user or patient registers with the HCRS and can check in using their identification or profile data. To maintain personalised data, CBF, and CF are used.

CBF: provide the past data of users

CF: real value genetic and fuzzy rule-based approaches are used to calculate similarity [30].

Between the front end, which gives people all the information, and the back end, which keeps all the users’ data, is where Healthcare RS sits [10].

The user can get details from HCRS. In this method, the user is given the most accurate suggestions while also receiving personalised information.

The user’s age, sex, employment, height, weight, region, symptoms, and so forth make up their demographic data.

Information from several users needs to be classified, and weights are assigned using real values. GA.

The above data will be used to create the classes using fuzzy classification with rules. ups or actions depending on their profiles. These alerts cover things like immunizations, medical exams, and diseases associated with advancing age. Ratings for hospitals and physicians will be requested from the users. Users will have access to this rating information, which will help them when making choices. To give users the most recent information relevant to their interests, information from different sources is gathered. Through other processes and video lectures, first-aid knowledge will be made available. By SMS, email, or advertisement post, notifications will be given for free check-ups, age-specific diseases, and vaccinations for childcare workers. Most of the time, the patient is not properly informed about hospitals, physicians, and their specialities. The suggested framework aims to create a model that gives the user the right information about hospitals and physicians based on their profiles and dynamic shifts in the user’s interests. The individual has the option of changing their preferences for a hospital, a physician, a demographic profile, etc. attention drift is another name for this process of shifting a user’s attention [11]. Finding comparable users improves the quality of recommendations for HCRS using fuzzy logic [5] and a genetic algorithm [20].

The current healthcare software only offers data on hospitals and physicians of user profiles. The user receives no customized suggestions from these applications. As a result, rather than providing personalized information, the system generates the chance of diseases when a user logs in using the patient’s symptoms. These comprise suggestions based on patient input parameters such as place, environment, area, economic status, and profile data of user.

The personalized information of the user is shown in Table 2. The user profiles are age, sex, profession, stature, location, rating, and health symptoms. It is challenging to take into account every symptom, per the varying datasets offered on different websites. For a disease, there may be a variety of signs, so some common symptoms have been listed here and the flow of Algorithm shown below:

Proposed Algorithm

Step # 1: Arrange the dataset according to the hybrid approach

Step # 2: Select the data from PIMS

Step # 3: Extract the data from Step 1 with attributes of users, profiles and their symptoms.

Step # 4: Pre-process the data set using clean, removal of missing values and transform the data, timestamp or time sensitivity 28 features.

Step # 5: Neighbourhood generation: Compute similarity using real value GA with mutation and cross-over operators.

Step # 6: Fuzzification of data by intuition

Step # 7: Set the membership function based on parameters of age, sex, occupation, height, weight and region.

Step # 8: Compute the evaluation metrics RMSE and MAE

2.1 Convolutional neural networks (CNN)

In this CNN process, the dataset is to be pre-processed. The number of training examples is then decided after the network has been trained by building a neural network (NN) and a calculation graph using the processed data and time information. After that, the training parameters are saved into a file. And use the Mean Square Error (MSE) number to assess how well the neural network’s model is performing. Continually altering the model’s parameters can lower the MSE number. Finally, the user characteristics and item features can be acquired by the model. We use the model to compute the prediction ratings and select the Top-k items with the highest ratings that were not rated by the target user in order to recommend items to the user. Convolutional Neural Networks (CNN) and Restricted Boltzman Machine, however, are the most effective methods.

2.2 Restricted boltzman machine (RBM)

A symmetric bipartite network in an RBM that has no connections between any two units that belong to the same group. Gradient descent and backpropagation techniques can be used to fine-tune multiple RBM, which would also be layered. The term “deep belief network” also applies to this network. Although RBMs are still occasionally used, the majority of the deep learning community has begun using General Instead of them, use adversarial networks or variable autoencoders. Because RBM is a stochastic neural network, a neuron’s behaviour when it is triggered will appear random. An RBM contains two additional levels of bias units, known as hidden bias and visible bias. This distinguishes RBMs from autoencoders. The visible bias aids RBM in reconstructing the input during a backward pass, and the concealed bias RBM generates the activation on the forward pass. Because there is no means for them to communicate with one another, the reconstructed input is never the same as the original input. The primary benefit of RBM is that selection seems to originate from the data distribution. When there are missing records, pattern completion can be done. Sampling is a trickier process when it comes to the training portion, though. Boltzmannn machines tackle search and Learning problems.

Search Problem. Issue With Search Boltzmann machines have connections with set weights that are used as the cost function in an optimisation process.

Learning Problem. The weights of the connections must be determined in order to optimise the training process given a collection of binary data vectors. By resolving the search issue repeatedly, Boltzmann machines alter the values of the weights.

2.3 HCRS using the fuzzy-genetic approach (FG-HCRS)

Fuzzy rules and genetic algorithms, which are crucial techniques used with large datasets, are the foundation of the suggested model’s construction [36]. This vast quantity of data requires adequate processing time.

2.3.1 Fuzzy rule-based systems

Objective function, defined as the goal of decision-making. For instance, a business wants to raise F, the total profit. Therefore, F’s return is an objective function.

Constraints. The Constraints are limitations or limits placed on the decision variables. The fitness value is another name for the objective function’s value. The fittest chromosomes are chosen from the population for use in subsequent procedures, which is why this crucial stage is called “selection.”

Optimization is a very important concept in any business domain be it retail, finance, automobile or healthcare. In simple words, the purpose of optimization is to find a point or set of points in the search space by minimizing/maximizing the loss/cost function, that gives us the optimal solution for the problem at hand. Here, we tried to minimize the objective function f(x) subject to one/multiple constraints.

Genetic Algorithm (GA) is to select, best individuals as parents from the population, asking them to reproduce to extend the generation. During this reproduction process genes from both parents crossover and in a way, an un-inherited error occurs which is known as mutation. Then the next generations are asked to reproduce their offspring and the process continues. The evolutionary algorithm is inspired by this theory of crossover and mutation, where basically Crossover is used to create new solutions from the population’s genetic information and mutation occurs to bring new information or maintain diversity within the population and prevent premature convergence to make the solution more generic.

With the conception of a fuzzy logic-based recommender system, there have been significant advancements in the applications that make it possible to clearly understand the use of such a system. Fuzzy sets are used to analyse the design of recommender systems, which was a breakthrough in the field of recommendations. Users have expressed concern about the difficulty in making reliable recommendations. The degree of object representation uncertainty was investigated. This representation suggests that both content-based and collaborative systems require a vector of ratings in order to identify the proper user interests. The original methodology was created for private systems.

The application of fuzzy logic with data mining techniques for recommender systems was investigated in 2005. A dynamic fuzzy cluster that utilised a changing degree of membership for users’ preferences.

The idea of fuzzy sets enables a particular kind of if-then reasoning known as fuzzy logic, which can be described as the evaluation of criteria to evaluate arguments. Essentially, fuzzy logic is just fuzzy set-based thinking. The terms fuzzy inference, fuzzy reasoning, and others with a comparable meanings are sometimes used synonymously. Remember the saying, “If the tomato is red, then it is ripe.” It is possible to use imprecise logic or reasoning to infer that a tomato is “slightly red” because it is “somewhat ripe”. Prior to the introduction of fuzzy logic, the digital value is used. The numbers between 0 and 1 are used in fuzzy reasoning. Both hardware and software employ this fuzzy reasoning. Digital systems only take into account Boolean values, whereas fuzzy logic considers all possible values in the spectrum of yes or no. Fuzzy logic thus addresses the mechanism’s degree of ambiguity. For the fuzzification process, variable declaration, Fuzzification, fuzzy rules implementation and conversion to graph are the important steps to be followed.

The fuzzification procedure is the first step, and all membership functions are declared as linguistic variables as shown in Fig. 2. A collection is then used to declare crisp values, which are later converted to fuzzy values. Finally, these fuzzy values are subjected to the rule base method and graphed. Here, six parameters based on user profiles, including age, gender, employment, height, weight, and region, are fuzzified. Because these are the various demographic profiles used for feature weights where 28 genres are mentioned in Fig. 6. Fuzzification of these values is imprecise to get accuracy values in the prediction and recommendation. The description of age, weight and height parameters is as follows.

Fig. 2

Fuzzification process of linguistic variables.

$μ_{V_{Y}} (A) = {\begin{matrix} 1, A \leq 13 \\ \frac{20 - A}{7}, 13 < A \leq 20 \\ 0, A > 20 \end{matrix}$ (1) $μ_{Y} (A) = {\begin{matrix} 0, A \leq 13, A > 31 \\ \frac{A - 13}{7}, 13 < A \leq 20 \\ \begin{matrix} 1, 20 < A \leq 24 \\ \frac{31 - A}{7}, 24 < A \leq 31 \end{matrix} \end{matrix}$ (2) $μ_{M} (A) = {\begin{matrix} 0, A \leq 24, A > 42 \\ \frac{A - 24}{7}, 24 < A \leq 31 \\ \begin{matrix} 1, 31 < A \leq 35 \\ \frac{42 - A}{7}, 35 < A \leq 42 \end{matrix} \end{matrix}$ (3) $μ_{O} A = {\begin{matrix} 0, A \leq 35, A > 53 \\ \frac{A - 35}{7}, 35 < A \leq 42 \\ \begin{matrix} 1, 42 < A \leq 46 \\ \frac{53 - A}{7}, 46 < A \leq 53 \end{matrix} \end{matrix}$ (4) $μ_{V_{O}} (A) = {\begin{matrix} 0, A \leq 46 \\ \frac{A - 46}{7}, 46 < A \leq 53 \\ 1, A > 53 \end{matrix}$ (5)

With membership values of 1, the values for gender and profession are considered as fuzzy points. The genre interest index is then split into five fuzzier groups. According to Fig. 3, the GIM of age is broken down into Very Young (VY), Young (Y), Middle Age (M), Old (O), and Very Old (VO). The five linguistic variables are as follows, and equations 1 –5 describe the corresponding membership functions. In these equations, μ_{V
_Y} (A) is defined for the age group of 0 yr to 20 yrs, μ_Y (A) may range between the age of 13 yrs to 31 yrs, μ_M (A) may range between the age of 24 yrs to 42 yrs, μ_O (A) may range between the age of 35 yrs to 53 yrs and μ_{V
_O} (A) may range from 46 yrs and above.

Fig. 3

Membership function for Age.

GIM weight categories include Very Light (VL), Light (L), Average (A), Heavy (H), and Very Heavy. (VH). The following equations describe the five linguistic variables and their membership functions. Figure 4 displays the weight membership algorithm. These are the five linguistic variables, and equations 6 through 10 describe the corresponding membership functions. In these equations, μ_{V
_L} (W) is defined for the weight range of 0 to 38, μ_L (W) may range between the weight of 66.13 lbs to 108.027 lbs, μ_A (W) may range between the weight of 92.59 lbs to 132.27 lbs, μ_H (W) may range between 116.845 lbs to 156.5 lbs and μ_{V
_H} (W) may range from 141.096 lbs and above

Fig. 4

Membership function for Weight.

$μ_{V_{L}} (W) = {\begin{matrix} 1, W \leq 66.13 \\ \frac{83.7 - W}{8}, 66.13 < W \leq 83.7 \\ 0, W > 83.7 \end{matrix}$ (6) $μ_{L} (W) = {\begin{matrix} 0, W \leq 66.13, W > 88.184 \\ \frac{W - 66.13}{8}, 66.13 < W \leq 83.7 \\ \begin{matrix} 1, 83.7 < W \leq 92.59 \\ \frac{108.027 - W}{8}, 92.59 < W \leq 108.027 \end{matrix} \end{matrix}$ (7) $μ_{A} (W) = {\begin{matrix} 0, W \leq 92.59, W > 132.27 \\ \frac{W - 92.59}{8}, 92.59 < W \leq 108.027 \\ \begin{matrix} 1, 108.027 < W \leq 116.845 \\ \frac{132.27 - W}{8}, 116.845 < W \leq 132.27 \end{matrix} \end{matrix}$ (8) $μ_{H} (W) = {\begin{matrix} 0, W \leq 116.845, W > 156.5 \\ \frac{W - 116.845}{8}, 116.845 < W \leq 132.27 \\ \begin{matrix} 1, 132.27 < W \leq 141.096 \\ \frac{156.5 - W}{8}, 141.096 < W \leq 156.5 \end{matrix} \end{matrix}$ (9) $μ_{V_{H}} = {\begin{matrix} 0, W \leq 141.096 \\ \frac{W - 141.096}{8}, 141.096 < W \leq 156.5 \\ 1, W > 156.5 \end{matrix}$ (10) $μ_{V_{S}} (H) = {\begin{matrix} 1, H \leq 76.2 \\ \frac{96.2 - H}{30}, 76.12 < H \leq 96.12 \\ 0, H > 96.2 \end{matrix}$ (11) $μ_{S} (H) = {\begin{matrix} 0, H \leq 76.2, H > 126.2 \\ \frac{H - 76.2}{20}, 76.2 < H \leq 96.2 \\ \begin{matrix} 1, 96.2 < H \leq 106.2 \\ \frac{126.2 - H}{20}, 106.2 < H \leq 126.2 \end{matrix} \end{matrix}$ (12) $μ_{A} (H) = {\begin{matrix} 0, H \leq 106.2, H > 156.2 \\ \frac{H - 106.2}{20}, 106.2 < H \leq 126.2 \\ \begin{matrix} 1, 126.2 < H \leq 136.2 \\ \frac{156.2 - H}{20}, 136.2 < H \leq 156.2 \end{matrix} \end{matrix}$ (13) $μ_{T} (H) = {\begin{matrix} 0, H \leq 136.2, H > 186.2 \\ \frac{H - 136.2}{20}, 136.2 < H \leq 156.2 \\ \begin{matrix} 1, 156.2 < H \leq 166.2 \\ \frac{186.2 - H}{20}, 166.2 < H \leq 186.2 \end{matrix} \end{matrix}$ (14) $μ_{vr} (H) = {\begin{matrix} 0, & H \leq 166.2 \\ \frac{H - 166.2}{20}, & 166.2 < H \leq 18602 \\ 1, & H > 186.2 \end{matrix}$ (15)

The height categories in the GIM (Genre Interestingness Measure) are Very Small (VS), Light (S), Average (A), Tall (T), and Very Tall (VT). (VT). The following equations describe the five linguistic variables and membership functions. the membership function for height is shown in Fig. 5. These five linguistic variables are defined in equations 11 through 15, along with their associated membership functions. In these equations, μ_{V
_S} (H)is defined for the height range of 0 cm to 96.2 cm, μ_S (H)may range between the weight of 76.2 cm to 126.2 cm, μ_A (H)may range between the weight of 106.2 cm to 156.2 cm, μ_T (H)may range between 136.2 cm to 186.2 cm and μ_{V
_T} (H)may range from 166.2 cm and above.

Fig. 5

Membership function for Height.

2.3.2 Hybrid fuzzy-genetic approach

An evolutionary process-based approach to issue solving is offered by the Genetic Algorithm (GA) [23]. An optimisation problem’s potential solution is decoded by the chromosome from a collection of tuned parameters. The fundamental principles of GA are crossing and mutation. Using these two operators, the GA’s efficiency is calculated. The term “feature weight” refers to the unique weight that each user assigns to each feature. These feature weights must be recorded and adjusted to represent the preferences of each user [33]. Each user’s preferences must be taken into account when capturing and fine-tuning these characteristic weights [33]. Weight (u_a)=[w_i] _{i - 1, . . n}, where n is the number of features, represents the feature weights of user u_a as a set of weights. GA was used to give weights in this situation. Good newly generated individuals are substituted for bad ones in each stage by mutation and crossover operator [33]. The collection of weights that best meets the selection criteria ultimately prevails. The weights collection W₁, W₂ . . . W₂₈, where each weight ranges between 0 to 1 and it is shown in Fig. 6.

Fig. 6

Chromosome of full features.

The user id is feature 1, and features 2 through 7 are the user profile data set. The most important demographic characteristics that define a user’s background are age, gender, employment, height, weight, region, and rating. The extra profile characteristics, or symptoms (symptom_0 to symptom_19), are represented by features 9–28. Here the symptoms are based on a single disease and these may vary for various diseases. Here in this, the real values may be considered of chromosomes using GA. These all features are collected from datasets of cms.gov.

The task of GA, which prepares the necessary features, comes first in the HCRS proposal. There are 28 features in the GA built on collaborative filtering. User id, age, gender, employment, height, weight, region, rating, and any symptoms of 4 feature weights (symptom_0 to symptom_19) may be used in this case to create a minimum of 12 features out of a total of 28 to determine the quality of the suggestion.

Fever: 1. Sweating, 2. Chills and shivering, 3. Headache, 4. Muscle aches, 5. Loss of appetite, 6. Irritability, 7. Dehydration, 8. General weakness, Allergies: 9. Allergies, 10. Anaphylaxis, 11. Dermatitis, 12. Burns, Cellulitis, Bone or Joint disease: 13. Arthritis, 14. back pain, Respiratory problems: 15. Dry Cough, Digestive complications: 16. Acidity, 17. Bloating, 18. Dyspepsia, 19. food poisoning, 20. gastritis

To train the model here the basic symptoms were considered. When the patient enters the symptoms and demographic features, then HCRS provide the information of doctors, hospitals and predicted illness with the percentage of possibility will be generated and the recommender engine will recommend the proactive steps to the patient/user.

The second stage is feature selection, in which a feature weight w in real value (or genre) is used to represent a chromosome’s genre [49]. The significance of the higher weight value is increased so that the value may represent how important the feature is to the user.

The algorithm usually comes to an end when either the maximum number of generations has been produced or the population has reached a satisfactory level of fitness. Here for the data set of 20 users may create the possible combinations of 400 using a cross-over operator. So, the mutation and crossover are applied to the dataset, to generate the maximum population so that the fitness function value may get high. After many possible combinations, the GA may stop. A fitness function and a genetic representation of the solution area are necessary for a standard genetic algorithm.

The next step is neighbourhood generation and finally, prediction and recommendation are done. Here the fitness function can be calculated by using either equations 17. $\sum_{i = 1}^{28} W_{i} X {(f_{a} (x_{i}, y_{i}))}^{2}$ (16) $\begin{matrix} Fitness (a, i) = ({Predic}_{v} ote (a, i) * 100 %) \\ / {actual}_{v} ote (a, i) \end{matrix}$ (17)

Equation 16 is used to compute similarity for the 28 feature weights in the suggested model. In HCRS systems, the weights of these characteristics are dynamically changing over time. Fuzzy principles are then applied to the learned weights using a Genetic Algorithm (GA), creating a hybrid fuzzy-genetic HCRS. When the best point in the current population’s fitness function has a value that is less than or equivalent to the Fitness Limit, the Genetic Algorithm stops.

3 Experiments and results

From cms.gov, several experiments using healthcare datasets have been performed. These datasets replicate the HCRS user data 60% for training and 40% for testing. Data on 12,000 patients were gathered from 600 institutions, with user ratings. The ten-fold cross-validation method, which is implemented in Python, is used in this paper’s training for outcomes evaluation. To achieve accuracy in the proposed model, the evaluation metrics MAE and RMSE are used which are shown in Tables 3 4.

Table 3
RMSE comparison of different schemes based on healthcare dataset

No. of users CNN RBM FG-HCRS

5 2.98 2.46 2.45

10 2.92 2.49 2.51

15 2.81 2.60 2.51

20 2.51 2.61 2.52

25 2.61 2.60 2.42

30 2.43 2.61 2.21

No. of users	CNN	RBM	FG-HCRS
5	2.98	2.46	2.45
10	2.92	2.49	2.51
15	2.81	2.60	2.51
20	2.51	2.61	2.52
25	2.61	2.60	2.42
30	2.43	2.61	2.21

Table 4

MAE comparison of various approaches based on the healthcare dataset

Training set	Schemes	A.U (5)	A.U (10)	A.U (20)
100	CNN	.951	.900	.778
	RBM	.903	.854	.723
	FG-HCRS	.701	.648	.716
200	CNN	.856	.846	.786
	RBM	.678	.778	.681
	FG-HCRS	.703	.703	.659
300	CNN	.903	.756	.867
	RBM	.867	.678	.761
	FG-HCRS	.698	.656	.732

Based on Tables 3 4, MAE is computed for a variety of users, and the outcomes of CNN (Convolution Neural Network), RBM (Restricted Boltzman Machine), and FG-HCRS (Fuzzy Genetic Health Care Recommender Systems) are contrasted.

When the number of users is less the MAE result is high.

MAE decreases as the number of users rises and the highest scores are taken into account for comparison. In this case, FG-HCRS has a lower MAE (Mean Absolute Error) than the conventional approaches of CNN and RBM.

When used in relation to machine learning, the term “absolute error” describes the size of the discrepancy between the prediction of a measurement and its actual value. One of the methods for assessing the accuracy of forecasts is the root mean square error or root mean square deviation. It displays the Euclidean distance between predicted values and observed true values.

The degree to which positive class predictions truly fall under that category is known as precision.

Precision = TP/TP+FP

Recall measures the number of effective class predictions that were made using all of the dataset’s valid examples.

Recall = TP/TP+FN

The current users are taken into account in each of the three training sets— 100, 200, and 300. MAE (Mean Absolute Error), RMSE (Root Mean Square Error), precision, and prediction accuracy are the measures used in this case. These metrics show us the degree of agreement between our predictions and the real values, as well as how accurate our predictions are. The split of 100 users is used in the Table 3 instances mentioned above. If we take into account active users 5, the MAE is large due to the complexity of the comparisons that could be made between the total number of users and active users. If there are 10 and 20 busy users, comparisons would be simpler, and predictions and suggestions could be made very quickly.

The more user sample comparisons are the cause. According to the tests we conducted, when splits and active users rise, MAE falls. In Figs. 7 8, the outcomes of these three techniques are displayed. As the number of active users grew, the MAE and RMSE in the previously mentioned Figs. 8 9 decreased.

Fig. 7

RMSE comparison of CNN, RBM and FG-HCRS.

Fig. 8

MAE comparison of CNN, RBM and FG-HCRS.

Fig. 9

Precision comparison of CNN, RBM and FG-HCRS.

The decreasing MAE and RMSE show that the accuracy of prediction is high. In light of the findings, it can be said that a system with fewer suggestions and higher precision will operate more effectively. Additionally discernible is the relationship between recall and the quantity of suggestions. The FG-HCRS has a smaller range of MAE than the other two techniques, CNN and RBM, which was undoubtedly shown in a table. Figures 9 10 below demonstrate how the HCRS outperforms CNN and RBM methods using the hybrid fuzzy-genetic strategy in terms of MAE, RMSE, Precision, and Recall.

Fig. 10

Recall comparison of CNN, RBM and FG-HCRS.

The dataset size is fixed with a crossover rate of.5, the paper employs multi-point crossover, which would require real value GA which runs 1000 times of trials. There have been numerous runs to finite-tune the parameters, over 1000 experiments at each mutation of 0.0023 to 0.021 in steps of 0.0021, and many executions at the combination of rate of mutation.

The best-weighted run for each active user of the FG-HCRS was chosen out of 10 runs, plotted against the results of other schemes, and 100 users were divided up for each split to determine MAE values, prediction percentages, precision, and recall. In contrast to other existing approaches, Split 1 (split is of 100 users) in Fig. 11 shows better results for FG-HCRS prediction accuracy.

Fig. 11

Prediction percentage of CNN, RBM and FG-HCRS.

4 Conclusion

Since they enable users to propose products they are interested in, recommendation systems are essential in online E-commerce. The main weakness that recommendation systems face as the number of users and goods increases is data sparsity and data scalability problems, which draw attention to poor prediction quality and inefficient time usage.

The suggested work presented a fuzzy Genetic HCRS to enhance predictive accuracy and recommendation quality while overcoming present limitations. It focuses on various user demographic features that can compute their similarity and predict the most comparable products for users. The experimental findings show that, when compared to traditional CNN and RBM, our approach substantially increases prediction accuracy, significantly lowers MAE, and improves the quality of the recommendation system.

This study suggests and puts into practice a special and effective framework for healthcare recommendation systems to generate high-quality tips. In the suggested method, suitable weights for the similarity measures are learned using a hybrid fuzzy-genetic method enhancing the efficacy and precision of suggestions.

The observed results show that, when compared to existing methods, in terms of MAE and prediction accuracy, with a target of a minimum MAE of 5.789 (the threshold value ranges between 0 and 1) and prediction accuracy of 96.78%. Results from experiments demonstrate the mechanism at action, which can significantly improve the calibre of recommendations. Scalability, sparsity, and cold start users/items are three problems that the suggested architecture can address.

We want to use machine learning methods to create personalised healthcare in the future while taking geographic information into account.

Authorship statement

All persons who meet authorship criteria are listed as authors, and all authors certify that they have participated sufficiently in the work to take public responsibility for the content, including participation in the concept, design, analysis, writing, or revision of the manuscript. Furthermore, each author certifies that this material or similar material has not been and will not be submitted to or published in any other publication.

Authorship contributions

Category 1

Conception and design of the study: Devendra Gautam, Latha Banda, Anurag Dixit, S.B. Goyal.

Acquisition of data: Latha Banda analysis and/or interpretation of data: Devendra Gautam, Latha Banda.

Category 2

Drafting the manuscript: Latha Banda, Devendra Gautam, Anurag Dixit, S.B. Goyal.

Revising the manuscript critically for important intellectual content: Devendra Gautam, Latha Banda, S.B. Goyal, Chaman Verma, Manoj Kumar.

Category 3

Approval of the version of the manuscript to be published.

Latha Banda, Devendra Gautam, Anurag Dixit, S.B. Goyal, Chaman Verma, Manoj Kumar.

Footnotes

Acknowledgments

The work of Chaman Verma was supported by the Department of Media and Educational Informatics, Faculty of Informatics, ELTE, Budapest, Hungary.

We are also grateful to the Noida International University for providing support for this research. Without their help, it would not have been possible for us to complete this manuscript.

References

Althbiti

, Alshamrani

, Alghamdi

, Lee

and Ma

, Addressing Data Sparsity in Collaborative Filtering Based Recommender Systems Using Clustering and Artificial Neural Network, 2021 IEEE 11th Annual Computing and Communication Workshop and Conference (CCWC), NV, USA, 2021, pp. 0218–0227. doi: 10.1109/CCWC51732.2021.9376008

Adomavicius

and Tuzhilin

, Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, Journal of IEEE Transaction on Knowledge and Data Engineering 6(17) (2005), 734–749.

Adomavicius

and Kwon

, New Recommendation Techniques for Multicriteria Rating Systems, IEEE Intelligent Systems 22(3) (2007), 48–55.

Agrawal

and Jain

, An improved approach for movie recommendation system, International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC), Palladam: (2017), 336–342. doi: 10.1109/I-SMAC.2017.8058367

Al-Shamri

and Bharadwaj

, Fuzzy-genetic approach to recommender system based on a novel hybrid user model, Expert Syst Applications 35 (2008), 1386–1399.

Alsalem

M.A.

, Alamoodi

A.H.

, Albahri

O.S.

, et al., Multi-criteria decision-making for coronavirus disease applications: a theoretical analysis review, ArtifIntell Rev 55 (2022), 4979–5062.

Antonopoulus

and Salter

, Cinema screen recommender agent: combining collaborative and content-based filtering, IEEE Intelligent Systems (2006), 35–41.

Archenaa

and Mary

E.A.

, Anita, Health Recommender System using Big data analytics, J Manag Sci Business Intelligence 2 (2017), 17–24.

Arian Hafezalkotob , Ashkan Hafezalkotob , Huchang Liao and Francisco Herrera , , An overview of MULTIMOORA for multi-criteria decision-making: Theory, developments, applications, and challenges, Information Fusion 51 (2019), 145–177. ISSN 1566-2535.

10.

Baldominos

, De Rada

and Saez

, DataCare: Big Data Analytics Solution for Intelligent Healthcare Management, International Journal of Interactive Multimedia Artificial Intelligence 4 (2018), 13–20.

11.

Banda

and Bharadwaj

K.K.

, Improving Scalability Issues in Collaborative Filtering based on Collaborative Tagging using Genre Interestingness Measure, ARTCom 2012, LNEE: (2012), 403–409.

12.

Banda

and Bharadwaj

K.K.

, Evaluation of Collaborative Filtering based on Tagging with Diffusion Similarity using Gradual Decay Approach, Smart Innovation, Systems and Technologies 27 (2014), 421–428. ISBN: 978-3-319-11933-5.

13.

Banda

and Singh

, Improving Scalability, Sparsity and Cold Start user Issues in Collaborative Tagging with Incremental Clustering and Trust, Recent Patents on Computer Science 12 (2019), 1. https://doi.org/10.2174/2213275912666190617163253. ISSN: 2213-2759.

14.

Banda

, Singh

, Son

L.H.

, Abdel-Basset

, Thong

P.H.

, Huynh

H.X.

and Taniar

, Recommender Systems using Collaborative Tagging in-, International Journal of Data Warehousing and Mining (IJDWM): 1548-3924 16(3) (2020), 183–203. 10.4018/IJDWM. (http://www.igi-global.com/journal/international-journal-data-warehousing-mining/1085)

15.

Behera

R.K.

, Sahoo

A.K.

and Pradhan

C.R.

, Big Data Analytics in Real Time—Technical Challenges and Its Solutions, In Proceedings of the 2017 International Conference on Information Technology (ICIT), Bhubaneswar, India, 21–23 December 2017 (2017), 30–35.

16.

Bell

R.M.

and Koren

, Climbable CF by Joint extractedarea Interpolation Weights, Processing in 2007, 7thIEEE International Conference on Information Extraction, (2007), 43–52.

17.

Bushra Alhijawi and Yousef Kilani , A collaborative filtering recommender system using genetic algorithm, Information Processing & Management, 57(6) (1023), 102310. ISSN 0306-4573.

18.

Canny

, Collaborative filtering with privacy, In IEEE Symposium on Security and Privacy, (2002), 45–57, Oakland, CA.

19.

Chakraborthy

P.S.

, A scalable collaborative filtering-based recommender system using incremental clustering, IACC’09 (2009). 978-1-4244-1888-6/08/IEEE.

20.

Chen

, Li

, Rong

, Bilal

, Yang

and Li

, A diagnosis and treatment recommender system based on big data mining and Cloud computing, Inf Sci 435 (2018), 124–149.

21.

Corcoran

A.L.

and Sen

, Usage of True valued genetic approach to mix rule sets for categorization in Proceedings of 1st IEEE world congress Conference on Evolutionary Computation Intelligence, (1994).

22.

Dai

and Wang

, A Deep inference learning framework for healthcare,139 (2018), 17–25. B. Gedik and L. Liu, Protecting location privacy with personalized k-anonymity: architecture and algorithms, IEEE Trans Mobile Comput 7(1) (2008), 1–18.

23.

Goldberg

, Genetic algorithms in search, optimization, and machine learning, Pearson Education, Upper Saddle River (1989).

24.

Gope

and Jain

S.K.

, A survey on solving cold start problem in recommender systems, In Proceedings of the 2017 International Conference on Computing, Communication, and Automation (ICCCA): (2017), 133–138.

25.

Gorski

, Heesch

, Dziendzikowski

and Dworakowski

, Fuzzy-Logic-Based Recommendation System for Processing in Condition Monitoring, Sensors (Basel) 22(10) (2022), 3695. doi: 10.3390/s22103695. PMID: 35632104; PMCID: PMC9146414

26.

Gruteser

and Liu

, Protecting privacy in continuous location-tracking applications, Secur Priv IEEE 2(2) (2004), 28–34.

27.

and Chua

T.S.

, Neural factorization machines for sparse predictive analytics, In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval: (2017), 355–364.

28.

, Liao

, Zhang

, Nie

, Hu

and Chua

T.S.

, Neural collaborative filtering, In Proceedings of the 26th International Conference on World Wide Web (WWW ’17), International World Wide Web Conferences Steering Committee: (2017), 173–182.

29.

Hoh

and Gruteser

, Protecting location privacy through path confusion, In: Security and privacy for emerging areas in communications networks: (2005), 194–205. IEEE Computer Society, Washington.

30.

Jiang

and Yang

C.C.

, User recommendation in healthcare social media by assessing user similarity in heterogeneous network, Artif Intel Med 81 (2017), 63–77.

31.

Kaur

, Kumar

and Batra

, An efficient multi-party scheme for privacy preserving collaborative filtering for healthcare recommender system, Future Gener Comput Syst 86 (2018), 297–307.

32.

Kumar

and Sharma

, Alleviating Sparsity and Scalability Issues in Collaborative Filtering Based Recommender Systems, In Proceedings of the International Conference on Frontiers of Intelligent Computing: Theory and Applications (FICTA). Advances in Intelligent Systems and Computing, vol. 199. (2013). Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-35314-7_13

33.

, King

and Lyu

M.R.

, Learning to Recommend with Social Trust Ensemble: Proceedings of the SIGIR’09: (2009), 39–46.

34.

, Lu

, Gan

and Zeng

, An explicit trust and distrust clustering based collaborative filtering recommender approach, Electron Commer Res Appl (2017), 29–39. M. Mitchell, An introduction to genetic algorithms, MIT Press, Cambridge, (1998).

35.

Mohamed

M.H.

, Khafagy

M.H.

and Ibrahim

M.H.

, Recommender Systems Challenges and Solutions Survey International Conference on Innovative Trends in Computer Engineering (ITCE), Aswan, Egypt: (2019), 149–155. doi: 10.1109/ITCE.2019.8646645

36.

Mohammad Yahya Al-Shamri

and Kamal Bharadwaj

, Fuzzy-genetic approach to recommender sytems based on a novel hybrid user model, Expert Systems with Applications 35(3) (2008), 1386–1399. ISSN 0957-4174.

37.

, Zeng

and Han

, A Survey of Recommender Systems Based on Deep Learning, IEEE Access 6 (2016), 69009–69022.

38.

Ochoa

J.G.D.

, Csiszár

and Schimper

, Medical recommender systems based on continuous-valued logic and multi-criteria decision operators, using interpretable neural networks, BMC Med Inform Decis Mak 21 (2021), 186.

39.

Palanisamy

and Thirunavukarasu

, Implications of big data analytics in developing healthcare frameworks— A review, J King Saud Univ-Comput Inf Sci 31(4) (2017), 415–425.

40.

Paul

P.K.

and Dey

J.L.

, Data Science Vis-à-Vis efficient healthcare and medical systems: A techno-managerial perspective, In Proceedings of the 2017 Innovations in Power and Advanced Computing Technologies (i-PACT), (2017), 1–8.

41.

Portugal

, Alencar

and Cowan

, The use of machine learning algorithms in recommender systems: A systematic review, Expert Syst Appl 97 (2018), 205–227.

42.

Priyadarshini

, Barik

and Dubey

, DeepFog: Fog Computing-Based Deep Neural Architecture for Prediction of Stress Types, Diabetes and Hypertension Attacks, Computation 6(4) (2018), 62. https://doi.org/10.3390/computation6040062

43.

Priyadarshini

, Barik

R.K.

, Panigrahi

, Dubey

and Mishra

B.K.

, An investigation into the efficacy of deep learning tools for big data analysis in health care, Int J Grid High Perform Comput 10 (2018), 1–13.

44.

Sahoo

A.K.

and Pradhan

C.R.

, A Novel Approach to Optimized Hybrid Item-based Collaborative Filtering Recommender Model using R. In Proceedings of the 2017 9th International Conference on Advanced Computing (ICoAC): (2017), 468–472.

45.

Sahoo

A.K.

, Pradhan

and Mishra

B.S.P.

, SVD based Privacy Preserving Recommendation Model using Optimized Hybrid Item-based Collaborative Filtering, In Proceedings of the 2019 International Conference on Communication and Signal Processing (ICCSP): (2019), 294–298.

46.

Shah

, Salunke

, Dongare

and Antala

, Recommender systems: An overview of different approaches to recommendations in International Conference on Innovations in Information, Embedded and Communication Systems (ICIIECS): (2017), 1–4.

47.

Shahabi

, Banaei-Kashani

, Chen

Y.S.

and McLeod

, Yoda: An Accurate and ScalableWeb-Based Recommendation System, In: C. Batini, F. Giunchiglia, P. Giorgini and M. Mecella, (eds) Cooperative Information Systems, CoopIS 2001. Lecture Notes in Computer Science, vol. 2172. (2001). Springer, Berlin, Heidelberg.

48.

Vedna Sharma and Surender Singh Samant , Health Recommendation System by Using Deep Learning and Fuzzy Technique (July 14, 2022), Proceedings of the Advancement in Electronics & Communication Engineering, 2022.

49.

Simon Fong , Yvonne Ho and Yang Hang , Using Genetic Algorithm for Hybrid Modes of Collaborative Filtering, Online Recommenders hybrid intelligent systems (HIS) (2008), 174–179.

50.

Symeonidis

, Nanopoulos

and Manolopoulos

, A Unified Framework for Providing Recommendations in Social Tagging Systems Based on Ternary Semantic Analysis, IEEE Transactions on Knowledge and Data Engineering 22 (2010), 179–192.

51.

Thomas and Sujatha

A.K.

, Comparative study of recommender systems, International Conference on Circuit, Power and Computing Technologies (ICCPCT): (2016), 1–6. doi: 10.1109/ICCPCT.2016.7530304

52.

Tran

T.N.T.

, Felfernig

, Trattner

, et al., Recommender systems in the healthcare domain: state-of-the-art and research issues, J Intell Inf Syst 57 (2021), 171–201. https://doi.org/10.1007/s10844-020-00633-6

53.

Wang

, He

, Nie

and Chua

T.S.

, Item silk road: Recommending items from information domains to social users, In Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval: (2017), 185–194.

54.

Wang

and Hajli

, Exploring the path to big data analytics success in healthcare, Journal of Bus Res 70 (2017), 287–299.

55.

Wang

, Kung

and Byrd

T.A.

, Big data analytics: Understanding its capabilities and potential benefits for healthcare organizations, Technol Forecast Soc 126 (2018), 3–13.

56.

Wei

, He

, Chen

, Zhou

and Tang

, Collaborative filtering and deep learning Based recommender system for cold start items, Expert Syst Appl 69 (2017), 29–39.

57.

Yang

C.C.

and Jiang

, Enriching User Experience in Online Health Communities Through Thread Recommendations and Heterogeneous Information Network Mining, IEEE Trans Comput Soc Syst 5 (2018), 1049–1060.

58.

Yedder

H.B.

, Zakia

, Ahmed

and Trajkovic

, Modeling prediction in recommender systems using restricted boltzmann machine, In Proceedings of the 2017 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Banff, AB, Canada, 5–8 October 2017.

59.

Yuan

, Li

, Guan

, Han

and Khattak

A.M.

, Socialized healthcare service recommendation using deep learning, Neural Comput Appl 30 (2018), 2071–2082.

A novel approach to enhance the quality of health care recommender system using fuzzy-genetic approach

Abstract

Keywords

1 Introduction

2 Literature survey

2.2 Restricted boltzman machine (RBM)

2.3 HCRS using the fuzzy-genetic approach (FG-HCRS)

2.3.1 Fuzzy rule-based systems

Table 3 RMSE comparison of different schemes based on healthcare dataset No. of users CNN RBM FG-HCRS 5 2.98 2.46 2.45 10 2.92 2.49 2.51 15 2.81 2.60 2.51 20 2.51 2.61 2.52 25 2.61 2.60 2.42 30 2.43 2.61 2.21

Authorship statement

Authorship contributions

Footnotes

Acknowledgments

References

Table 3
RMSE comparison of different schemes based on healthcare dataset

No. of users CNN RBM FG-HCRS

5 2.98 2.46 2.45

10 2.92 2.49 2.51

15 2.81 2.60 2.51

20 2.51 2.61 2.52

25 2.61 2.60 2.42

30 2.43 2.61 2.21