Skip to main content

Table 6 Features and challenges of different recommender systems

From: A systematic review and research perspective on recommender systems

References

Methodology

Features

Challenges

Castellano et al. [1]

Neuro-fuzzy

High accuracy

Model is not efficient for real-time applications

Crespo et al. [2]

K-means clustering

Solves the problems related to electronic commerce and provides a good user experience

Model is not applicable on huge datasets

Lin et al. [3]

GA, k-means, DT and BN

Increases profit in real-time applications along with product sale volumes

Lacks in operation efficiency

Wang and Wu [4]

Association rule mining

Useful for real-time ubiquitous-learning applications

The model does not consider the learner’s behaviour

García-Crespo et al. [5]

Fuzzy logic

High recommendation accuracy and scalability

Not applicable on large datasets

Dong et al. [6]

Semantic similarity model

Solves the issue of service supply ecosystems in a software environment

Low overall performance

Li et al. [7]

Fuzzy linguistic modelling

Good expert recommendations for knowledge management systems

Does not consider the user’s past preferences

Lorenzi et al. [8]

Content-based filtering

This model can alleviate the sparsity issue of recommender systems

Low accuracy

Huang et al. [9]

Rough set model

It enhances the accuracy of recommender systems

The model is complex for real-time applications

Chen et al. [10]

Semantic web rule language

The model can efficiently extract the features of the dataset

The model is dependent on domain ontology data

Mohanraj et al. [11]

Foraging Bees algorithm

A self-adaptive model with low computational time

Does not consider any change in user interests

Hsu et al. [12]

Artificial bee colony algorithm

It has attained better accuracy and improved the execution time

The system is limited to the specific application only

Gemmell et al. [13]

Linear-weighted hybrid approach

The model is highly effective and flexible for capitalizing on a robust relationship in different data dimensions

The model is complex and requires high memory

Choi et al. [14]

Sequential pattern analysis

High accuracy and performance

This model cannot be used for large datasets

Garibaldi et al. [15]

Fuzzy rule system

The model integrates variability with a conventional fuzzy inference system

Not suitable for sparse datasets

Salehi [16]

Content-based filtering

Good F1 score, recall and precision values

Precision decreases with the increase of the size of the recommendation sets

Aher and Lobo [17]

K-means and Association rule mining (ARM)

The model is capable of efficiently recommending online courses to new students

Does not consider specific choices of individual users

Kardan and Ebrahimi [18]

ARM

This hybrid model delivers better efficiency when compared with individual models. It also considers implicit user information

The algorithm does not rank the outcome of the mining algorithm

Chang et al. [19]

HFC, K-means clustering and KNN

Improved resource utilization in a cloud environment

This model suffers from hardware resource complications

Lucas et al. [20]

Associative classification and Fuzzy logic

The model successfully alleviates problems such as gray sheep, cold-start, sparsity, and scalability issues

The model is not suitable for web mining applications

Niu et al. [21]

Spectral clustering

The system attains better performance in terms of recall, precision and F1-score, and delivers a good user experience

Not applicable for low-level features

Liu et al. [22]

Genetic Algorithm

A unique model was developed to save the travelling time of self-driven tourists

The model is platform-dependent, and the recommendations are not user-customizable

Bakshi et al. [23]

Unsupervised learning, PSO

The model addresses the issues of scalability and sparse datasets

The selection of clusters is done arbitrarily

Kim and Shim [24]

Probabilistic model

Good precision, recall and average hit rank

The model is complex

Wang et al. [25]

GA and K-means

High prediction accuracy

The model is not suitable for real-time applications and sparse data

Kolomvatsos et al. [26]

Stochastic decision making

It can efficiently handle multiple requests and save resources and time

The model is not suitable for real-time applications

Gottschlich et al. [27]

Portfolio optimization algorithm

A unique model which considers the crowd’s wisdom to maximize a stock investor’s portfolio performance

It is a general model which does not consider specific user needs

Torshizi et al. [28]

Fuzzy rule-based systems

A novel model useful for prostate treatment recommendations

The model strongly relies on medical data which lacks certainty

Zahálka et al. [29]

Linear SVM model

An information-rich and interactive venue recommender model capable of connecting tourists

The model has low precision values for some datasets

Sankar et al. [30]

Content-based filtering approach

A simple and accurate model for investment guidance

Hidden relationships among features may result in low accuracy

Chen et al. [31]

Artificial immune system

It employs a new prediction formula to improve the recall, precision, and F1-scores

Suffers from data scalability, cold-start and memory-based issues

Wu et al. [32]

Expectation maximization (EM) -based

algorithm

Highly robust in fusing item recommendations for collaborative tagging systems

Low recommendation quality and not suitable for large datasets

Yeh and  Cheng [33]

Delphi panel and repertory grid techniques

The model is easy to use and solves the dimensionality problem and data sparsity issue

Not applicable for real-time applications

Liao et al. [34]

Rough set association rule

The model considers dynamic user behaviour

High computation complexity and not fit for other applications

Li et al. [35]

Association

rule mining

The model can be applied to many different applications

Performance varies for different applications

Wu et al. [36]

CTR and SMF

This model is simple and robust having high accuracy

Does not consider the social trust among users

Adeniyi et al. [37]

KNN algorithm

A scalable model with a good precision rate and low computing load

Not applicable for real-time applications

Rawat and Kankanhalli [38]

Image processing

A scalable system capable of learning offline and working in real-time

The system lacks personalization

Yang et al. [39]

Hybrid gradient boosting

Suitable for real-time and large-scale recommender systems. It efficiently reduces false positive predictions

The model suffers from the cold-start problem and requires pre-clustering and feature engineering

Lee et al. [40]

Collaborative ensemble learning

The model can handle dynamic user behaviour captured from mobile signals

Suffers from data sparsity and cold start issues

Wei et al. [41]

SDAE

The model efficiently solves the cold start problem

Requires high computational resources and storage

Li et al. [42]

Weighted linear regression

Low computational complexity and low prediction and modelling time for real-time applications

The model is complex to implement in real large-scale applications

Mezei and Nikou [43]

Fuzzy optimization

Better accuracy and F1-score

Not scalable and cannot adopt other variants of fuzzy sets

Ayata et al. [44]

SVM, RF, DT and KNN

The model can capture real-time user emotion in a dynamic environment

The model is susceptible to the failure of sensors

Zhao et al. [45]

Multimodal network learning

This high-performance model is capable of handling real-world, large-scale datasets

It suffers from high computational costs and premature convergence

Hammou et al. [46]

Matrix factorization and RF

The model reduced the cost functions by efficient learning of parameters to overcome the data sparsity problem

High computational cost and does not consider social contexts of users

Zhao et al. [47]

ANNInit

A scalable model with high prediction accuracy and convergence efficiency

Not suitable for multi-dimension applications

Bhaskaran and Santhi [48]

FA and K-mean clustering

High performance and accuracy with low error and computation time

Cannot be applied to other optimization methods

Afolabi and Toivanen [59]

User-based approaches

A unique system where patients can share recommendations in real-time resulting in better self-care and decision making

Requires a smooth connection between different healthcare services

He et al. [60]

Knowledge representation learning

A high-performance model which alleviates the data sparsity issue and cold start problem

The model is complex

Han et al. [49]

BAS

The model is simple, portable and has high prediction accuracy with a fast convergence rate

BAS algorithm suffers from low search accuracy and limited exploration scope

Kang et al. [50]

Hashmap

Solves the problem of data sparsity and overspecialization

High time complexity

Ullah et al. [51]

RF and CNN

High efficacy and suitable for real-time product recommendations

It suffers from high time complexity

Cai et al. [52]

MaoEA

High accuracy and applicability in diverse applications

The model lacks stability

Esteban et al. [53]

GA

High performance and reliability for real-time course recommendations

The model lacks user personalization

Mondal et al. [54]

Multilayer graph data model

It provides trust-based results with no requirement of external meta-data for calculating the trust factor

The model is only applicable to limited scenarios

Dhelim et al. [55]

User-based approaches

The system can efficiently detect the meta-interest of the user and has a good recall and precision score for cold-start environments

It requires high computational power to decrease response time

Bhalse et al. [56]

SVD, CS and collaborative filtering

It reduces the complexity by reducing the multi-dimensionality issue

Suffers from sparsity problem and low accuracy

Ke et al. [57]

Reinforcement learning

High prediction accuracy and low entropy loss error

Not suitable for real-world problems

Chen et al. [58]

FNN

The system delivers personalized suggestions by aggregating multiple user interest representations

Suffers from cold-start problem