1 Introduction

Service Oriented Architecture (SOA) has emerged as a promising paradigm in system engineering, where systems are constructed by integrating services as their fundamental building blocks. The proliferation of service providers has led to a vast number of services, making the selection of the most suitable one among many offering equal functionalities a big challenge. To optimize, one approach, services are selected based on their Quality of Service (QoS) attributes (i.e. non-functional properties, such as response time, throughput, etc.). However, certain QoS attributes, being provider-declared, are not inherently stable. For instance, in the widely recognized WSDREAM dataset [1], the response time attribute exhibits fluctuations within a range of [0 s–20 s]. Consequently, the prediction of dynamic QoS attributes has become a critical area of research interest over the past decade.

In dynamic environments, users may encounter varying QoS values from the same service due to fluctuations in service load (number of clients) and network conditions (e.g., congestion) over time. Therefore, time emerges as a critical factor influencing prediction accuracy. Addressing this challenge, time-aware Collaborative Filtering (CF) methods have been proposed to predict QoS in such environments. Recent research has prominently favored these methods for QoS prediction, driven by several reasons. Firstly, they have demonstrated significant enhancements in prediction accuracy by incorporating diverse contextual information about users and services [2,3,4]. Secondly, their versatility has been evidenced across various applications in the service computing domain, including service selection, composition, adaptation, and fault tolerance [5]. Thirdly, they possess the capability to leverage large historical data for predicting current or future QoS values. Lastly, they exhibit adaptability to the dynamic environment, accommodating changes such as the introduction of new QoS values, new users, or services. Given the considerations mentioned above, it is evident that time-aware CF is emerging as a new trend for achieving accurate QoS predictions. In light of this trend, we present a comprehensive literature review that emphasizes various methods employed in this type of prediction.

Our review focused on time-aware QoS prediction using CF methods. The primary studies included in this review were collected from four known digital scientific libraries, namely IEEExplore, Springer, ScienceDirect, and ACM, spanning the years from 2011 to 2022. The named libraries were chosen as they encompass well-known journals and conferences in this field such as SOCA, TSC, ICWS, ICSOC, and SCC. During our search, the following keywords were used in the search queries: Time aware, temporal, Collaborative Filtering, CF, QoS, service, predict, recommend, assess. To limit the scope of our review, two inclusion criteria were applied: first, studies had to propose a time-aware QoS prediction method, predicting either current or future QoS; second, the proposed method had to be a CF method, utilizing data from other users and services in making QoS prediction.

Ultimately, we identified 40 notable studies that represent the current state-of-the-art. These studies were thematically categorized into three groups: (1) time-aware neighborhood CF, (2) time-aware model-based CF, and (3) time-aware hybrid approaches. Remarkably, our literature review is the first dedicated exploration of time-aware CF methods, as previous CF-related work discussed time-aware methods within a broader context. The specialization in our review proves valuable for researchers seeking a comprehensive understanding of state-of-the-art time-aware methods. Additionally, for each primary study, we analyzed its key strengths and weaknesses. We then conducted a comprehensive comparison of time-aware methods within our thematic categorization, offering readers insight into the evolution of research trends over the years. Lastly, we pinpointed key research challenges in time-aware CF and offered potential research directions to guide further exploration in this domain.

The rest of the review is organized as follows: Sect. 2 introduces the background, and Sect. 3 presents the related work. Section 4 describes our classification and the approaches under each category. In Sect. 5 we discuss our findings. Section 6 presents research challenges and directions. Finally, we conclude our work in Sect. 7.

2 Background

Utilizing collaborative filtering for QoS prediction draws inspiration from commercial recommendation systems, such as those employed by Netflix, Amazon, and eBay. The term "collaborative filtering" is defined as the process of filtering information or patterns through techniques involving collaboration among multiple users, agents, and data sources [6]. In their work [7], the authors take the initiative of applying collaborative filtering in QoS prediction. Their approach involved predicting the missing QoS values for a target user by leveraging the existing values from other similar users utilizing the same invoked services. To provide a more precise understanding, we will explain the issue using the formal definition of collaborative filtering methods:

  • Let \(U=\{\) \(u_1\),\(u_2\) , ... , \(u_m\) \(\}\) is set of users for Web services, \(u_i\) denotes a user, where \((1 \le i \le m)\).

  • Let \(S=\{\) \(s_1\),\(s_2\) , ... , \(s_n\) \(\}\) is set of Web services, \(s_j\) denotes a service, where \((1 \le j \le n)\).

  • Let \(Q_{m\times n}\) is user-service matrix, where \(q_{ij}\) represents the QoS value of the \(i\) user when he invoked the \(j\) service.

Fig. 1
figure 1

User-service QoS matrix(U:user, S:service)

Full size image

Figure 1 shows a toy example of the user-service matrix, which holds values for QoS (it can be any attribute, such as response time, and throughput). The blank cells indicate missing (unknown) QoS values since users have not invoked these services yet. In essence, the CF methods consist of three main types, which are neighborhood CF, model-based CF, and hybrid CF. Neighborhood CF, also recognized as memory-based CF, relies on similarity calculations, typically employing the Pearson Correlation Coefficient (PCC). The user PCC (UPCC) and item IPCC (IPCC) are commonly utilized methods for calculating user and service similarity, respectively. The model-based CF is introduced to address scalability and QoS sparsity problems, employing a pre-trained model to predict missing QoS values. Lastly, the hybrid methods aim to leverage the advantages of both neighborhood and model-based approaches to improve prediction accuracy.

Nonetheless, traditional CF types can be expanded by integrating the time of service invocation as an additional contextual factor. This integration, as previously discussed, often leads to improved accuracy in QoS predictions. Methods that incorporate time factors are commonly referred to as time-aware CF. Figure 2, illustrates the QoS matrix used in time-aware methods, which leverage QoS data gathered across various time intervals. This historical data can be augmented in the prediction process through various techniques, such as adjusting time similarity computations in neighborhood methods and training models using specialized time-aware datasets in model-based methods.

Fig. 2
figure 2

User-service QoS matrix(U:user, S:service, T:time)

Full size image

3 Related work

3.1 Time-aware methods for QoS prediction

QoS prediction using CF methods has received the attention of researchers in the last decade. Several studies reviewed and summarized these methods. These studies are general and not dedicated to any specific type, however, this literature review is dedicated only to the time-awareness CF methods in QoS prediction, and to the best of our knowledge there are no reviews that have been conducted in this field. However, we will discuss these general studies according to their relevance to the topic.

In [8], the authors provided a survey about Web service QoS prediction via CF, they categorized these methods in two levels: at the first level, they used the general categorization as neighborhood, model-based, and hybrid, at the second level the methods under each general category were further categorized according to what type of contextual data they incorporated, such as location, time or other. In addition, they discussed the forefront research issues like adaptability, credibility, and privacy-preserving. The work in [9] also provided a survey about QoS Web service prediction methods, the authors categorized methods into the known general categories: neighborhood, model-based, and hybrid. They also dedicated a specific section to time-aware collaborative methods, briefly discussing several popular techniques.

In literature, time-aware CF finds applications in various domains beyond the QoS prediction domain, including service recommendation. For example, in [10] the authors reviewed the time-aware recommender systems (TARS) that are used in various types of services, different types were reviewed such as time-aware CF, time-aware content-based, and time-aware knowledge-based, in their work QoS was used an important criteria for evaluating and recommending a service. In the same line of research, the authors in [11] provided an overview of the Web service recommendation system, they differentiated between recommendations and predictions. They also provided explanations of different types of CF, like user-based, item-based, model-based, personalized, and location-aware.

An alternative strategy that prioritizes time awareness in selecting services is the time series forecasting methodology. This method enables the statistical prediction of QoS values. Prominent methods here are the Moving Average (MA) method, Auto Regressive (AR), and Auto-Regressive Integrated with Moving Average (ARIMA). However, it’s important to note that this approach diverges from collaborative filtering-based methods as it operates on a per-user-service basis, placing it beyond the scope of this review. Despite this limitation, we found that some CF methods integrated time series in their prediction, which prompted us to include a recent relevant study in this domain. In [12], a comprehensive survey on QoS time series modeling and forecasting is presented. The authors selected a collection of studies and examined four key aspects in each: the identified problem, the proposed methodology, the performance metrics considered, and the QoS time series dataset used. Additionally, they highlighted the shortcomings observed in these studies.

3.2 General time-aware CF methods

In this section, we focus on reviewing time-aware CF methods applied in areas not directly related to services. The study in [13] highlighted the significance of incorporating time factors to enhance the accuracy of CF recommendation systems. They discussed traditional CF methods and elaborated on how these techniques can be expanded to incorporate time factors using various techniques. In [14] the authors conducted an analysis of time-aware recommendation systems, highlighting the limitations of the evaluation methods used in these recommenders. They proposed a methodological framework aimed at ensuring a fair evaluation process. Additionally, the work in [15] presented a recent systematic review of neural network-based recommender systems. Within this review, they classified recommender systems into different categories, including CF. The authors specifically emphasized the growing trend of employing temporal (sequential) models to enhance the accuracy of recommenders, addressing this as a separate section within their work.

4 Time-aware CF methods: review

Time-aware CF methods are thematically categorized into three categories: time-aware neighborhood methods, time-aware model-based methods, and time-aware hybrid methods. This categorization aligns with the various aspects of time-awareness in QoS prediction. The subsequent subsections provide a literature review of the diverse methods within each category.

Table 1 Time-aware neighbourhood collaborative filtering
Full size table

4.1 Time-aware neighbourhood collaborative filtering

The methods under this section used the traditional CF computation in both similarity and prediction measurements, however, to be time-aware methods they have to capture the dynamic change of QoS similarity over time. The time-aware similarity can be computed using one of two methods: first, using the time decay function as a weighting major for the effectiveness of QoS values, and second, using the time interval slots method. Next, we provide more details about the studies under each method. Additionally Table 1 highlights the strengths and limitations of each study individually, along with general information such as their publication year and type.

4.1.1 Time decay method 

The authors in [16] used an exponential decay function whose value decreases as the time span between two related QoS increases or as the time span between the current time and two related QoS increases. They alleviated the data sparsity problem by using the random walk algorithm, which discovered the indirect user and service similarities. However, authors in [17] argued that using non-linear decay functions alone is not sufficient for evaluating the effectiveness of QoS values, so they designed a hybrid decay function, of both linear and non-linear. Similarly, the study in [23] has used the exponential time decay function, but a novel idea is added, which aims to increase weights for QoS values that seemed to be too small or too large in user similarity calculation. They also modeled the correlation between user and service locations before calculating the similarity in order to increase prediction accuracy.

4.1.2 Time interval method

The time interval method was used with average similarity computation. This method divides the historical QoS data into time slots and creates a matrix of users and services in each slot. It computes similarity in each time slot and the final value of similarity at the current time is the average of similarities in all time slots. In a study done in [18], the authors calculated user and service similarity in a static number of time slots determined by a variable named \(d\), which was a parameter used to reduce the searching space. The same authors extended their work and introduced a time and location-aware method in [19]. Their new method used location-based clusters of users and services in order to alleviate scalability problems. In [20], the authors tried to improve the work done in [18]. They used a clustering approach that determined dynamically the size of time slots instead of being static.

Another work in [21] introduced a novel approach named CluCF. This work extended the studies [18] and [19]. The authors alleviated the data sparsity problem. They converted the sparse user, service, and time tensor into a high-density user-service matrix, this matrix was converted into userCluster-service matrix and user-serviceCluster matrix. The clustering was based on location data. In the end, a hybrid prediction with weighted parameters is computed from both user and service predictions. In this method, the clusters can be updated when new users or services are introduced, however, it had a trade-off between scalability and prediction accuracy.

Later on, [22] and [2] improved the final similarity measure by using weighting functions and this achieved a better improvement over the average similarity measure used in the aforementioned studies. So first, in [22], a new approach was used to calculate the service similarity in the historical data. They used CANDECOMP/PARAFAC (CP) tensor decomposition to alleviate the data sparsity problem, and they assigned weights to global and temporal neighborhood services. Second, in [2], the user and service similarities were measured in a set of time slots, to compute the final similarity, the authors used a weighted decay function, which emphasized the similarity effect of recent time slots. In addition, they introduced a novel approach that searched for the most similar user in each time slot.

4.2 Time-aware model-based collaborative filtering

Time-aware model-based methods represent a large number of studies in CF methods. They depend on training a model with a large set of historical QoS data. The trained model can be used later for predicting QoS. They are further classified into three subcategories: latent factors methods, clustering, machine learning methods, and deep learning methods.

4.2.1 Latent factors methods

Latent factors methods are based on the assumption that the user-service matrix can be factorized into low-rank latent factor matrices, by utilizing these matrices the missing QoS can be predicted. It’s important to note that while latent factorization is the central focus across all studies in this section, some studies exhibit an overlap with other mentioned approaches. The strengths and limitations of each study are delineated individually in Table 2.

In the year 2011, Zhang et al. [1] introduced the first time-aware CF method which was named WSPred. This method created a tensor of three dimensions: user, services, and time. In order to predict missing QoS data, it performed a tensor factorization that learned the latent factors of users, and services in specific time intervals. The main contribution of their work was the data used in the tensor, which was real data that had been collected and used for the first time. It is known now as WSDREAM dataset2 [24] and it has become a well-known benchmark in the research community. Later on, similar work was introduced in [25]. The authors used a Non-negative Tensor Factorization (NTF) approach. The approach used CANDECOMP/PARAFAC (CP) factorization with consideration to the non-negativity property of QoS data. It decomposed the user, service, and time tensor into three non-negative latent matrices to get an approximation for the temporal QoS values. Moreover, the approach was evaluated using their own collected dataset, which was a tensor of size \(343\times 5817\times 32\) user-service-time.

The same authors introduced another work in [26]. They used a triadic factorization approach on a user, service, and time tensor. The novelty in their approach was providing a mechanism to reduce the memory space needed to store the sparse data in the high dimensional tensor. To do so, they proposed two methods: Tucker Decomposition (TD) and the coordinate approach, the former achieved a remarkable memory space reduction. They evaluated their approach using a tensor of size \(408\times 5473\times 56\) user-service-time.

One of the main limitations in studies [1, 25, 26] was making predictions offline, which means once the models are trained, they are unable to deal with new incoming QoS data. To overcome this limitation, the study in [27] proposed an Incremental Tensor Factorization (ITF) method, the ITF is based on the incremental approach of Singular Value Decomposition (SVD) and Tucker Decomposition (TD). The new approach could update prediction when new QoS data arrives while preserving the scalability and space efficiency properties. It was evaluated on a tensor of size: 408 users and 5473 Web services at 240 time periods, and it achieved higher accuracy than the offline methods.

In [5], the authors used the Adaptive Matrix Factorization (AMF) method that made QoS prediction for candidate services in run-time service adaptation. A set of well-designed steps were followed to achieve the requirements of accuracy, efficiency, and robustness. The method performed matrix factorization for each time slot, with the ability to learn online and to update its parameters using adaptive weights as new QoS data arrives or as new users and services come.

In [28], the authors used a hybrid method of both traditional neighborhood CF and latent factors in order to increase prediction accuracy. In the traditional neighborhood CF part, they used a service-based similarity measure that distinguished between static and temporal QoS attributes. In the latent factor part, they used CANDECOMP/PARAFAC (CP) decomposition on the user, service, and time tenor. The final prediction was a weighted addition of the two parts. In [29], the study used the CP factorization of a user-service-time tensor by applying non-negativity constraint on QoS data. The important contribution of this study was improving the prediction accuracy by several steps including a linear bias for both user, service, and time to model the temporal changes in data, using multiplicative learning rule for parameter optimization, and using of altering direction method in the training process.

In [30], authors provided an outlier resilient prediction method that used Cauchy loss for measuring the prediction errors. However, they extended their method by providing time-aware prediction by using CP factorization approach. Also, they added the non-negativity constraint on QoS data, which caused them to use the Multiplicative Updating (MU) algorithm to optimize the parameters.

In [31], the authors modeled the effect of temporal changes on service recommendation at three levels: users, services, and preferences. They used a latent factor decomposition that had a bias shifting for each one of the mentioned levels. They used the implicit feedback from users, which was collected on their own dataset. In [32], an adaptive matrix factorization approach was used to model the interactions between users and services in a specific time slot. The enhancement, in this approach, was the addition of temporal smoothing of the prediction, which accounted for the dependency between QoS in adjacent time slots. In [3], a model named CARP was proposed, the model can be used for offline and online predictions. The method used K-means clustering to cluster the invocation records, where each cluster represented a specific context and a cluster may contain a set of time slots. In order to alleviate the data sparsity problem, they aggregated invocation records from different time slots in the same cluster. Lastly, a matrix factorization approach was used to predict the final reliability value.

To improve the prediction accuracy, other studies incorporated context data like the location of users and services. Incorporating such context data to cluster users and services may help in alleviating the data sparsity problem. Moreover, it can help in improving the final prediction accuracy due to the implicit correlation between time and location that must be considered when making predictions. An example of these studies is the study in [33], where the authors created a tensor of multi-dimensions(user, service, time, location, and QoS property) and used a tensor decomposition method to predict missing QoS values. Another study is [34], which created local clusters of users and services based on location information, it performed a hierarchical tensor decomposition in two types of tensors: the location-based local tensors and the general global tensors. Finally, in [35], a unified and generalized approach was contributed. The approach created a tensor of five dimensions(user, service, time, location, and QoS property). It used tensor decomposition to predict QoS. The prediction loss was minimized using iRPROP+ optimization method, which produced accurate prediction results.

Table 2 Time-aware latent factors collaborative filtering
Full size table

4.2.2 Clustering and machine learning methods

Several studies have exploited clustering and machine learning approaches in QoS prediction. Clustering is usually used as a data pre-processing step to alleviate the scalability and data sparsity problems. It is not sufficient alone to perform QoS prediction, so other methods like linear regression and QoS averaging are merged with the approaches in this section. Below is a summary of these studies and Table 3 emphasizes the strengths and limitations of each study.

In [36], a method named CLUS was proposed, it predicted reliability attributes for ongoing services. The method performed a K-means clustering of invocation records into three steps: environmental variable (network load) clustering, user-specific clustering, and service-specific clustering. The final prediction was done by cluster-based computations that used the averaging of the reliability values. In addition, the authors used a linear regression model for making predictions.

In [37], the authors provided a novel method that first predicted the QoS at the current time by calculating the average of the historical QoS data in a pre-determined time interval, then a K-means clustering approach was used to make clusters of similar users and services. The authors used the average value of the resulting clusters to make user and service-based predictions, lastly, a linear weighted addition of the two predictions was used.

In [38], the authors proposed a method of two steps: first, it filled in missing QoS values in the historical QoS time slots. This was done by employing clustering to compute user and service similarity, the missing QoS was then calculated by averaging the weighted similarity for both users and services. Second, it predicted QoS in the current time slot by using the averaging of the calculated historical QoS data. The method in [39] generated temporal patterns that represented a series of user invocations for each service, after smoothing the pattern, a clustering approach was used to cluster the generated temporal patterns. The final prediction of missing QoS was done using a polynomial fitting function.

In [40], a novel approach called lasso was proposed, this method treats the QoS as a general regression problem. It used lasso regularization to overcome the sparsity of the QoS data. In addition, it used the location of users and services to improve prediction accuracy. This model also can accommodate newly incoming QoS and provide up-to-date predictions. In [41], a Weighted Support Vector Machine (WSVM) was used. This approach treated the problem of QoS prediction as a linear regression problem but in a high dimensional space. It used an exponential weighting function to give high weights for recent data. A sliding window approach was used to generate data for training.

Table 3 Time-aware clustering and machine learning methods
Full size table

4.2.3 Deep learning methods

To distinguish it from traditional machine learning methods, this section describes methods that use deep learning approaches, including neural networks and their derivations. Table 4 shows more details about the strengths and limitations of each study in this section.

In [4], the authors proposed a novel method called PLMF. The method improved the prediction accuracy by employing Long Short-Term Memory (LSTM), which is a type of Recurrent Neural Network (RNN). It performed online learning and continuously trained with newly coming QoS data by using a moving sliding window. The model used matrix factorization, where the latent factors of both users and services were learned using a personalized LSTM.

The study in [42], proposed a method that used a matrix called QI, which was generated from integrating invocation records with the QoS observation matrix. The method captured the user preferences and service features matrices by using matrix factorization. An LSTM was used to predict the QoS values at each time slice of 64 time intervals, from which, the top \(N\) Web services were recommended to the user. Despite that LSTM can model long-term dependency between QoS data, it has the problem of vanishing gradient, which may stop the learning process in the neural network. In order to overcome this limitation, the study in [43] proposed a method that used a Projected Factorization Machine (PFM) and Gated Recurrent Unit (GRU). The PFM was used to capture the non-linear interaction in a user, service, and time tensor, and the GRU was used to model the long-term dependency between sequential historical QoS records. A combination of the two predictions was adopted. A similar method was proposed in [44] which used Generalized Tensor Factorization (GTF) to model the static relationship between user, services, and time. Indeed, it used a Personalized Recurrent Gated Unit (PRGU) to model the long-term dependency. A maximum activation function was used to combine the two predictions.

Other studies utilized the ability of deep learning in inferring the complex relationships between different input features, they used neural networks to model the correlation between time and location as two important context data in the prediction. In the study [45] two methods named STCA-1 and STCA-2 were proposed. In these methods, the spatial and temporal features of services and users were extracted and entered into hierarchical neural networks. The networks were composed of multiple important layers, for example, an interaction layer was used to identify the first and second-order features. Attention layers were used to assign more weights to spatial features which made this model more interpretable than other models. In [46], a method named QSPC was introduced. It utilized two inputs: the request context and the temporal information. These inputs were fed to a multi-layer neural network. One of the important layers in this network was the LSTM layer, which captured the temporal information into a set of service requests using a static time window. The final output consisted of the prediction of multiple QoS attributes, in their case, response time and throughput. In [47], another method MtforSRec was proposed which accounted for static and dynamic QoS data. It used a factorization machine to model the static feature of QoS and a bi-directional LSTM to model the dynamic features. A softmax layer was used to give the final recommendations from the combined predictions. In [48], a method named DeepTSQP was proposed. It integrated features computed from the traditional similarity measures with binary features. For QoS prediction, it used the GRU model which helped in modeling the temporal dependency and in mining the implicit features in user-service interactions. This method achieved good prediction accuracy compared with the methods covered in this review.

Table 4 Time-aware deep learning methods
Full size table

4.3 Time-aware hybrid collaborative filtering methods

Several recent studies combined the CF methods with other methods, such as time series models and their derivations. Usually, this hybridization is done to improve prediction accuracy, these studies can be summarized as follows, indeed Table 5 highlights the strengths and limitations of these studies.

In [49], the authors proposed a hybrid method that combined autoregressive integrated with moving average (ARIMA) model and traditional CF. ARIMA was used to generate a time series for each Web service, however, ARIMA can’t correct itself timely by taking new observations as feedback. To overcome this limitation, KALMAN filtering was used. The authors employed CF to capture user side effects by using user-based similarity. Lastly, they added two predictions for the final output. In [50], a method was presented that also combined CF with ARIMA. The method first applied traditional CF to predict missing QoS for the past and current Point In Time (PIT). This method used two types of user similarities: global similarity with attenuation function, and user invocation similarity with edit distance measure. In the second step of prediction, the ARIMA method was used to forecast QoS for the future PIT. The final Web services recommendation was done using Multi-Criteria Decision Making (MCDM). In [51], the authors proposed a method that combined time series analysis with cloud model theory based on the CF approach to predict unknown QoS. The QoS data was transformed into a time series that represented different cloud models for different time periods. The similarity between models was measured using two novel methods namely, orientation and dimension similarity, which improved the final similarity computation. This method also used weights for every period using the fuzzy analytic hierarchy method.

Table 5 Time-aware time-aware hybrid methods
Full size table

5 Discussion

Each of the CF time-aware methods presented above possesses its own set of strengths and weaknesses. Neighborhood CF methods are generally characterized by their simplicity, making them easy to comprehend, implement, and increment new data. Moreover, they are more interpretable compared to other methods. In terms of their time-awareness capability, these methods have successfully captured temporal changes in user and service similarities using uncomplicated techniques like time decay and time intervals. However, their effectiveness is constrained to capturing temporal changes within limited time intervals and they cannot account for long-term time dependency between QoS values as a whole. To address this limitation, these methods have been combined with other approaches, such as time series methods [49] and [51], or they have employed tensor decomposition as demonstrated in [28]. Another significant challenge in this category is the scalability issue, wherein computational complexity increases with the growing number of users and services over time.

The methods in latent factor model-based CF models provide better scalability along with better prediction accuracy. They modeled the time factor by embedding it in a tensor and employing tensor decomposition methods like TD or CP. Other simpler methods use matrix factorization in a specific time slot, however, the tensor structure still suffers from data sparsity and scalability problems [3]. However, these methods can only model the interactions between users and services in short-term (limited time) intervals and do not account for long-term dependency [44]. Additionally, to provide up-to-date predictions, they require re-training of the entire model to accommodate the new coming QoS data, which is considered an expensive process. Alternatively, they utilize online adaptive learning through methods like Stochastic Gradient Descent (SGD) [5].

The clustering model-based CF methods are recognized for their simplicity and efficiency. They have been employed to address scalability issues by reducing the search space, consequently decreasing computational complexity [19]. Simultaneously, this space reduction has also mitigated the sparsity problem, leading to improvements in final prediction accuracy [21]. However, it is known that the clustering method is less effective when dealing with noisy and outlier data, particularly in cases of highly fluctuating QoS values. To tackle this, data smoothing is necessary before applying clustering, as demonstrated in [39]. In terms of providing up-to-date predictions, these models face the challenge of decreasing prediction accuracy with the accumulation of new data. Consequently, clusters need to be rebuilt periodically to address this issue, as discussed in [21, 36], and [3].

However, with rapid evolution in deep learning methods; a lot of the previously mentioned problems have been mitigated. These methods can accommodate a large number of inputs (users, services, and their context data) without a noticeable decline in computation speed. Also, they succeeded in assembling different methods to gain additional advantages, for example, the PLMF method, in [4], has integrated matrix factorization which captures user, and service interactions with LSTM which captures long-term dependency. This integration has achieved better accuracy results compared with CARP, in [3], and CLUS, in [36]. Regarding their ability to update predictions, two of the approaches, discussed in this review: [4] and [46], have provided up-to-date predictions by using incremental training. They employed algorithms like Stochastic Gradient Descent (SGD) and Adam optimizer to dynamically update their parameters online. Nevertheless, a primary drawback of deep learning methods lies in their lack of interpretability, as they often overlook the reasoning aspect when inferring correlation among input features such as users, services, time, and location.

The hybrid methods, presented in this review, have integrated CF with times series methods, like ARIMA. Specifically, CF utilized the time series in two ways: first, by modeling long-term temporal dependency between QoS data, as done in [49] and [51], resulting in improved prediction accuracy. Second, by forecasting the future values of QoS, as in [50]. However, time series models cannot capture personalized features of users or services. Indeed, generating time series for each Web service is considered a costly and difficult process. For providing up-to-date predictions, these hybrid models require novel methods for updating with new QoS data; however, none of the studies reviewed have contributed to solving this specific challenge.

In Figure 3, the distribution of studies across different types over time is presented. It is evident that despite their conventional nature, neighborhood methods (NH) continue to be prominent in the literature, particularly when integrated with other types. Likewise, latent factor (LF) methods have maintained consistent usage over the years, serving as fundamental techniques in model-based approaches. However, there has been a notable shift in the trend of QoS prediction towards the adoption of machine learning (ML), with increasing attention towards deep learning (DL) methods in recent years, as illustrated in the aforementioned figure. The reasons behind this shift are illustrated in the preceding discussion.

Fig. 3
figure 3

Distribution of studies over years

Full size image

6 Research challenges and directions

In this section, we explore the challenges that face researchers in this field. Furthermore, we examine potential avenues for new research directions that researchers can seek in future work.

6.1 Research challenges

6.1.1 Data sparsity

In reality, a user usually invokes a limited number of services, so the QoS values of the un-invoked services remain unknown forming what is called the data sparsity problem. This problem becomes more critical when building time-aware methods since it will occur in multi-time slots during user-service interactions. Several studies, in the literature, came up with several sparse-tolerant solutions such as using random walk algorithm [16], using data aggregation [3], or using clustering [21]. However, this challenge is still unsolved and there is room for more innovative ideas to mitigate it.

6.1.2 Deficiency in incorporating other context data correctly

Time is one of the factors that affect prediction accuracy, however, other contextual factors such as the location of users or services, and environmental factors also play a role in prediction accuracy. The important point here is the understanding of the correlation between the time factor and other factors. This is considered a kind of context reasoning that can be inferred by observing and analyzing the historical QoS values in the datasets. Several studies’ attempts can help in investigating datasets, on this issue, such as [24] and [52]. In fact, models must be built based on observations and evidence that would interpret context data correlation. This will help in generating true context-aware models that have high prediction accuracy.

6.1.3 The deficiency in providing up-to-date predictions

It is very important for time-aware to be updated continuously as new QoS data is coming. The majority of methods discussed in this review are offline methods (i.e. all QoS data are collected before the training phase). The accuracy of the offline methods deteriorates as time advances since they ignore new QoS observations that may carry changes in users, service similarities, or changes in context. Another important point, here, is that in a dynamic environment, the number of users and services also change over time. In reality, new users or services may appear, or current users and services may be disconnected. However, to address this challenge two solutions exist: first re-training the offline model periodically, re-training is required to accommodate new real-time QoS observations and new users or services. The limitation of this solution is the expensive time spent in re-training and testing the models as in [3, 21, 36]. Second building adaptive online models, these models can adapt to changes timely and can provide accurate up-to-date predictions. The premise of these models is that no need to train the whole model, however, there are limitations to this solution, for example, in online clustering models, there is always a trade-off between accuracy and scalability. Also, the online latent factor and deep learning models need special techniques that use moving sliding window and Adam or SGD optimizer to enable the online incremental training [4, 46]. However, this incremental training is a modern trend that needs further exploration of many issues such as computational complexity, resource consumption, stability, and maintainability.

6.2 Optional research directions

There are several research directions that researchers may work on in order to increase the accuracy of time-aware methods, from these we mention the following:

6.2.1 Creating generalised methods

Most of the research methods attempted to increase their accuracy concerning a limited number of known datasets commonly used in the experiments. However, this may result in creating data-biased methods that produce inaccurate results when they are evaluated on large-scale datasets [53]. Hence, there is room for enhancements here, for example: testing these methods using other different real datasets, applying them in real environments in the industry, or integrating them with real applications that need QoS prediction.

6.2.2 Creating unified methods

The majority of the current methods incorporated one or two contextual data, like being location-aware, time-aware, or both. The more contextual data used by the prediction method the higher accuracy it provides [54]. To this end, some methods are oriented toward building a unified framework, which can be extended to include new contextual data without changing the model’s internal structure. In fact, this will release researchers from updating or creating models to support new types of contextual data. In these models, contextual data, like service semantic or load, environmental conditions, user-specific context, etc. can be combined into one unified model. In addition, these unified models may be extended to support multi-QoS factor predictions, such as predicting response time, throughput, and reliability at the same time, which is expected to increase prediction accuracy [35].

6.2.3 Creating new datasets

The majority of studies in this review utilized the WSDREAM dataset [24]. Although this dataset is a real dataset, it has several limitations. First, the used Web services are SOAP-based, so it would be helpful to include other recent types of Web service, such as Restful API, providing other types may bring other research challenges in QoS prediction for cloud, mobile, and IoT fields. Second, the size of this dataset is considered small, so creating a larger dataset is an important need to keep up with the huge increase in the number of Web services in the real world. Third, this dataset records QoS values such as response time and throughput independently in different datasets, this forms a limitation to research that attempts to conduct multi-predictions. Including QoS attributes in a synchronous manner will bring about new research issues.

6.2.4 Performing empirical studies

Performing empirical studies in the field of time-aware CF methods is considered an important need. However, until the time of writing this review, there are no empirical studies in this field. In fact, most of the studies, in this review, have deficiencies in selecting the baseline methods for comparison, they may compare their methods with non-time-aware methods or with a small number of time-aware methods. So an empirical study is needed to provide a clear picture of the performance of these methods at computational and prediction accuracy levels. Moreover, most of the studies, in this review, discussed the accuracy of their approach without reporting any information about their computational complexity. Researchers who are interested in this direction can benefit from empirical studies that have been conducted in the time series field, like [55], where the authors compared (23) methods and proved that Genetic Programming (GP) had better accuracy than ARIMA. Similarly, [56] compared time series methods with some machine learning methods. Another less comprehensive one is in [57] where authors compared less complicated time series methods. However, for CF, one may compare several well-known deep learning methods, several online methods, or any other combinations. Surely, this comparison will help in selecting the right method either in academic or industrial fields.

7 Conclusion

This paper presents a comprehensive review of time-aware Web service QoS prediction using CF methods, encompassing a total of forty studies. The reviews are thematically classified into time-aware neighborhood CF, time-aware model-based CF, and time-aware hybrid approaches. Each classification is thoroughly analyzed, revealing research challenges such as data sparsity, inadequate exploration of crucial service prediction context data, and the absence of up-to-date service predictions. Identified limitations include the challenges in providing up-to-date predictions, high computation costs for offline re-training methods, limited details on computational complexity, interpretability issues, particularly in deep learning methods, and insufficient exploration of service and user context along with QoS factors. The review proposes potential research directions aimed at addressing these challenges and advancing the field.