A time-aware hyperlink-induced topic search-based reputation evaluation method for optimal manufacturing service recommendation in distributed peer-to-peer networks

Abstract

With the explosive growth of manufacturing services on a global scale, it has become an enormous challenge to recommend optimal manufacturing service in distributed manufacturing environments. Reputation evaluation is a key step to make optimal manufacturing service recommendation. However, the reputation values calculated through conventional reputation evaluation methods are static and local, and have a severe impact on the effectiveness of manufacturing service recommendation. This article presents a novel and effective reputation evaluation method for optimal manufacturing service recommendation in distributed peer-to-peer networks. The main contribution of the proposed approach is that it makes it possible to calculate the global and dynamic reputation values of enterprises by employing a time-aware hyperlink-induced topic search algorithm. This algorithm not only disseminates the reputation value of transaction enterprises, but also takes account of the temporal dimension in the process of computing reputation values. A concrete example demonstrates the feasibility of calculating reputation values of the proposed method, and an evaluation experiment reveals that both of recall rate and precision rate are improved, employing our proposed method for recommending manufacturing service.

Introduction

The rapid development of the Internet and overwhelming growth of global manufacturing resources have given enterprises increasing opportunities to cooperate with each other to enhance their respective levels of core competitiveness. Service-oriented architectures (SOAs) ¹ and web services offer promising methods of overcoming the barriers to the seamless interoperability of distributed resources by eliminating the technical discrepancies between the different software, platforms, and infrastructures of enterprises. ² However, enterprises also have faced the tremendous challenge of rapidly and effectively determining the optimal service within distributed manufacturing environments.

In our previous works, we have adopted peer-to-peer (P2P) architecture to overcome the conventional difficulties of ineffective and stiff manufacturing service discovery, including a self-organized P2P framework, ² a reputation-based P2P architecture, ³ and an agent-based P2P architecture. ⁴ We have presented enterprises as peers that can not only offer manufacturing services for sharing through their private service registries, but can also consume manufacturing services provided by other peers in distributed manufacturing environments. Manufacturing services that satisfy the fundamental requirements of manufacturing enterprises can be efficiently and effectively discovered and recommended. In distributed P2P networks, there are two main differences between conventional service recommendation and manufacturing service recommendation. On the one hand, quality of service (QoS) values of traditional services are usually easily obtained through the ratings from a large amount of consumers, while there usually existed scarce QoS values of manufacturing services for different consumer enterprises because of the expensive or time-consuming real-world execution of manufacturing services. ⁵ On the other hand, manufacturing service providers usually pay more attention on their business reputation than the traditional service provider, because it is more difficult for a reputable business to escape its social responsibility.

Reputation evaluation is a key step to make manufacturing service recommendation, because each enterprise peer hopes to employ the optimal manufacturing services provided by other trustworthy peers. However, the reputation values calculated through conventional reputation evaluation technologies are usually local, since the global characteristics of reputation propagation throughout manufacturing service supply chain networks are often ignored during calculation. As is well-known, the hyperlink-induced topic search (HITS) algorithm ⁶ is a popular link analysis algorithm for calculating the importance of a webpage, which is based on the following two basic ideas: that a high-quality authority page can be linked to by a great number of high-quality hub pages, and that a valuable hub page can point to a variety of valuable authority pages. Through comparative analysis, we have found that the manufacturing service supply chain network has similar characteristics to a webpage network, which has inspired us to introduce the HITS algorithm as a means of calculating global reputation value in a distributed P2P environment. With the aid of the HITS algorithm, the reputation value of each peer, which is allocated both an authority value and a hub value, can be disseminated successfully along the topology of manufacturing service supply chains in distributed P2P networks. Beyond that, the reputation values calculated through conventional reputation evaluation technologies are usually static, due to the fact that they rarely consider the timeliness of the reputation values of peers. As time goes by, the reputation values of enterprises also change dynamically. Therefore, the timeliness of reputation value is an indispensable factor in the process of computing reputation values.

In order to overcome the above defects, we present a novel and feasible method, called the time-aware HITS algorithm, to calculate the global and dynamic reputation values of enterprises in a distributed P2P environment. Based on these global and dynamic reputation values, the optimal manufacturing service can be effectively and reliably recommended to the enterprise that needs it.

The remainder of the paper is organized as follows. Related works are summarized in “Related works” section. “A novel method for optimal manufacturing service recommendation” section describes a time-aware HITS-based reputation evaluation method for evaluating the reasonable reputation values of enterprises in distributed P2P networks. In “Case study and evaluation experiment” section, an example and an evaluation experiment are proposed to illustrate the feasibility and effectiveness of our proposed method. The last section presents our conclusion with regard to the main contributions of this paper, as well as suggestions for further work.

Related works

P2P-based service discovery

Thus far, a great number of researchers have paid close attention to service discoveries in a distributed environment. P2P is a promising and successful technology for both sharing and discovering a variety of distributed resources flexibly and scalably.

For example, Mirtaheri and Sharifi ⁷ have presented a promising and useful framework, called discovery of heterogeneous multiple compute resources framework (DHMCF), for more efficiently discovering suitable resources and more reliably deploying responding structures in totally unstructured P2P networks. Mashayekhi and Habibi ⁸ have proposed an efficient and feasible framework for routing the most reputable peers to discover a suitable service, by combining a P2P search algorithm with a trust system. Deng et al. ⁹ have employed an ant colony optimization (ACO) algorithm in a large-scale P2P grid system, which uses a multi-attribute range query to efficiently discover the desired resource. Xiang et al. ¹⁰ have proposed the VPeers framework, which involves P2P service discoveries of virtual manufacturing organizations, and in which each peer can provide services for sharing and maintain a list of transaction friend VPeers. Fan et al. ¹¹ have proposed a novel trust method that considers both recommending resource service behaviors and recommended behaviors to compute eigenvectors of trust ratings for enterprises in P2P file-sharing networks. Our previous works^2–4 have presented three unique service-oriented P2P architectures for the efficient and effective semantic discovery of manufacturing services in heterogeneous and distributed environments. In our reputation-based P2P architecture, ³ we designed five related trust parameters and a multi-criteria decision-making mathematical (MCDMM) method for discovering required manufacturing services.

However, the aforementioned studies of P2P-based service discovery have a common drawback: that the characteristic of global reputation propagation among transaction peers is not taken into account. Therefore, in this paper, we take full advantage of the HITS algorithm to calculate the global reputation values in distributed P2P networks.

HITS algorithm

In 1999, Jon Kleinberg first presented the HITS, ⁶ which is a famous and significant ranking algorithm. Ding et al. ¹² elaborated on the HITS algorithm, ranking webpages by combining probabilistic analysis with matrix algebra, and mutually reinforcing the in-links and out-links of webpages. The algorithm has since then been constantly extended and widely used across a variety of domains, including engineering and economics.

For example, Häme and Hakula ¹³ have presented a modified HITS algorithm for effectively discovering feasible measures to the dial-a-ride problem (DARP). In order to overcome the DARP, the values of all hubs that are characterized as nodes with numerous out-links are computed and ranked to offer guidance to a backtracking algorithm. Deguchi et al. ¹⁴ adopted a weighted HITS algorithm to investigate the economic relationships of the world trade network (WTN) from 1992 to 2012. The hub value and authority value of each country illustrate the temporal evolution of world trade, and some typical behaviors can be explained by analyzing changes in the relationships of countries. Yang and Sun ¹⁵ have built a complex model for a knowledge map in an academic web space by combining the HITS algorithm, which is leveraged to discover web resources from an open web space, with social network analysis (SNA). Chen et al. ¹⁶ have presented a hybrid mechanism for calculating the corresponding values of tags regarding web service description language documents and service tag network information by using semantic computation and HITS, respectively.

Applying the HITS algorithm is a remarkable and promising method for carrying out the aforementioned types of research, since hub value and authority value are constantly mutually reinforcing in the calculation process. To the best of our knowledge, HITS has rarely been applied in the domain of manufacturing service recommendation. Furthermore, the temporal dimension of reputation value has rarely been taken into account during the calculation process, in which deviations of hub value and authority value are easily produced. Yu et al. ¹⁷ have introduced Timed PageRank for publication search, which adds a temporal dimension to the PageRank algorithm. ¹⁸ However, Timed PageRank needs more iterations to calculate a result, and the convergent rate is slower, making HITS a more appropriate means of recommending manufacturing services in distributed P2P networks.

Therefore, we propose a unique method, called time-aware HITS algorithm, which makes the reputation values of enterprises more seasonable and the recommendation of manufacturing services more reliable.

A novel method for optimal manufacturing service recommendation

In this section, we employ a time-aware HITS-based reputation evaluation approach to make trustworthy manufacturing service recommendation in distributed P2P environments. Our approach can be divided into four parts: manufacturing service discovery in distributed P2P networks, reputation value dissemination between discovered peers by employing the HITS algorithm, calculation of the reputation value of each peer by using a time-aware HITS algorithm, and manufacturing service recommendation through our method.

Manufacturing service discovery in distributed P2P networks

With the intense growth of global manufacturing services, a lack of flexibility, scalability, and effectiveness in traditional service discovery architecture have made the distributed manufacturing service discovery process undesirable. ¹⁰ We have taken full advantage of the P2P framework to facilitate manufacturing service discovery in a manner that is more scalable and efficient than in previous works, including the use of a self-organized P2P approach, ² a reputation-based P2P architecture, ³ and an agent-based P2P architecture. ⁴ Manufacturing service discovery can be divided into four main categories, leveraging the P2P architecture: structured, unstructured, super-peer, and hybrid. ¹⁹ The aforementioned works suggest that each enterprise acts as a peer that not only can provide manufacturing services through a private service registry and gain relevant trust ratings, but can also consume suitable services and offer a corresponding trust rating. ¹¹ If a peer requires a manufacturing service, it can first search its local private service registry to find a semantically similar service. If that search fails, the peer has to make a further query to its neighbor peers for the semantic discovery of manufacturing services. If the neighbor peers still cannot find matching services, they in turn transfer the query to their neighbor peers until some matching manufacturing services feed back to the original peer in question.

These P2P technologies help to discover a set of manufacturing services. However, the services discovered through these means will merely satisfy the fundamental demands of enterprises. Instead, optimal and highly trustworthy manufacturing services should be recommended to consumer enterprises, as chosen from among the discovered manufacturing services in distributed P2P networks. With this in mind, we employ the HITS algorithm, which enables the dissemination of the reputation values of transaction enterprises and be used to calculate the global reputation values of mutual transaction peers, making it possible to recommend the optimal manufacturing services to the consumer peers in distributed P2P networks.

Reputation dissemination by employing the HITS algorithm

The HITS algorithm assumes that a valuable authority page is pointed to by a large number of valuable hub pages, and that a high-quality hub page points to many high-quality authority pages. According to the above assumptions, the reputation values of both an authority page and a hub page can be computed through iterative computation until result convergence occurs. The process involved in the HITS algorithm means that the reputation value of each webpage, which originates from predecessors' reputations, can be propagated to successors.

It is obvious that the HITS link structure is similar to manufacturing service discovery and trust rating in distributed P2P networks. The authority value and hub value of each webpage represent the rated reputation value and the rating reputation value of each peer, respectively. Due to the great number of manufacturing services in distributed P2P networks, a lot of transaction data exist regarding service publication and service consumption. The authority value and hub value of each enterprise peer can be calculated based on the historical trust ratings of recommended enterprise peers. An example of manufacturing service transaction trust ratings in distributed P2P networks is illustrated in Figure 1. Each trust rating link is assigned a corresponding weight.

Figure 1.

Trust ratings of eight enterprises in distributed P2P networks.

Reputation calculation by using a time-aware HITS method

Traditional HITS algorithm

In the HITS algorithm, each enterprise peer is assigned an authority value a _i and hub value h _i , where i denotes peer i, and each trust rating link is assigned a corresponding weight. The authority value and hub value of peer i can be represented by the following formula:

$$a_i=\sum _{j:e_{\mathrm{ij}}\in E}{w}_{\mathrm{ij}}*h_j,h_i=\sum _{j:{e_j}_i\in E}{w}_{\mathrm{ji}}*a_j$$

(1)

We illustrate an example of the authority value and hub value of enterprise peer 6 in Figure 2.

Figure 2.

Authority value and hub value of enterprise peer 6.

Manufacturing services recommendation in distributed P2P networks can be denoted by a directed graph G = (V, E), where V = |V| represents the number of peers and E = {i| Transa(i), e _ij } denotes the trust rating regarding service transactions. Transa(i) is the set of enterprise peers from which peer i has obtained its required services, while e _ij is the corresponding trust rating offered by peer i.

In order to depict the iterative calculation of the authority value and hub value of each peer, we build a |V|*|V| adjacency matrix M, where originally m _ij  = 1 if there is a link or recommendation from peer i to peer j, and otherwise m _ij  = 0. Each link carries a relevant weight, so that m _ij  = 1 × w_ij, if a trust rating is from peer i to peer j. As an example, Table 1 demonstrates the trust ratings of manufacturing enterprises in P2P architecture in Figure 1 with the corresponding adjacency matrix.

Table 1.

Adjacency matrix M.

	E1	E2	E3	E4	E5	E6	E7	E8
E1	0	0.2	0.2	0	0	0.3	0.2	0.1
E2	0	0	0.4	0	0	0.6	0	0
E3	0	0	0	0	0.7	0	0	0.3
E4	0	0	0.2	0	0.4	0	0	0.4
E5	0	0	0.6	0	0	0	0.4	0
E6	0.8	0	0	0	0	0	0.2	0
E7	0	0	0	1	0	0	0	0
E8	0	0	0	0	0.5	0	0.5	0

If vector x = (x₁, … , x _n ) ^T and vector y =  (y₁, … , y _n ) ^T denote the authority value and the hub value of every peer, respectively, the iterative process of the HITS algorithm can be represented with the following formula:

$$\eta {\mathbf{x}}^{(k+1)}={\mathbf{M}}^T{\mathbf{y}}^{(k)}={\mathbf{M}}^T\mathbf{M}{\mathbf{x}}^{(k)},\eta {\mathbf{y}}^{(k+1)}=\mathbf{M}{\mathbf{x}}^{(k+1)}=\mathbf{M}{\mathbf{M}}^T{\mathbf{y}}^{(k)}$$

(2) where η acts as a normalization constant so that ||x|| = ||y|| = 1, k is represented by the kth iteration and $k\in \{0,1,2\dots \}$ , M^TM, and MM^T denote the authority matrix and hub matrix, respectively.

The initial authority values of the eight enterprises are all assigned as 1/8, and the authority value and hub value of each peer can be calculated with equation (2). The result of the authority values x[8] = [0.050, 0.019, 0.183, 0.002, 0.321, 0.089, 0.192, 0.144] ^T as the reputation values of enterprises. It appears in this case that peer 5 is the highest reputation value of the manufacturing enterprises, but there is a fatal flaw that remains to be improved.

Time-aware HITS method

In reality, the reputation values of enterprises always change over time. For example, an enterprise with a high reputation value may offer unsuitable or even bad services to others, since its previously valuable services may not satisfy the current requirements of manufacturing enterprises. Therefore, timeliness is an indispensable factor in the process of calculating reputation value. We thus propose a modified HITS algorithm, which we call the time-aware HITS algorithm, to calculate the global, dynamic, and reliable reputation values of every enterprise peer, and to recommend optimal manufacturing services.

Inspired by Timed PageRank, ¹⁷ we propose a time-aware HITS algorithm that enhances the accuracy of the reputation values of enterprises, making it possible to recommend the optimal manufacturing service. In order to make the reputation value of each peer more reasonable, we present the time function f(t), which is based on the principle of continual weakening. Function f(t) is represented as follows ¹⁷ :

$$f(t_{\mathrm{ij}})=D{R}^{(c-t_{\mathrm{ij}})/12},\mathrm{\hspace{1em}\hspace{1em}\hspace{1em}}f(t_{\mathrm{ji}})=D{R}^{(c-t_{\mathrm{ji}})/12}$$

(3) where c is the current time, and t _ij and t _ji are the previous service transaction times from peer i to peer j and from peer j to peer i, respectively, so that (c − t _ij ) and (c − t _ji ) are the two gaps in months and are less than 12. DR is a decay parameter and set as 0.3 which is selected after trail test, and will be further explained in “Evaluating the recall rate and precision rate of the optimal manufacturing service recommendation” section.

We modify equation (1) by multiplying the time function f(t) by each reputation value of each peer. The authority value and hub value of peer i can be updated as follows:

$$a_i=\sum _{j:e_{\mathrm{ij}}\in E}f(t_{\mathrm{ij}})*w_{\mathrm{ij}}*h_j,h_i=\sum _{j:{e_j}_i\in E}f(t_{\mathrm{ji}})*w_{\mathrm{ji}}*a_j,$$

(4)

In order to guarantee that the sum of the hub values of each peer is 1, or in other words, that the sum of each row of matrix M is 1, the element m _ij in equation (2) can be calculated as follows:

$$m_{\mathrm{ij}}=\frac{f(t_{\mathrm{ij}})*w_{\mathrm{ij}}}{\sum _{g\in \mathrm{Hub}(j)}f(t_{\mathrm{ig}})*w_{\mathrm{ig}}}$$

(5) where f(t _ij ) can be computed by equation (3), and w _ij is the relevant weight of a trust rating from peer i to peer j. Hub(j) represents the set of peers g that contains all corresponding trust rating from peer i to peer g.

Function f(t) only evaluates the importance of a peer from the past. However, we also need to consider that the potential significance of the peer in the future, such as after-sales service. Hence, we introduce another parameter, direction factor. d _p and d _n represent the number of peers that employ the services of peer i during the past month and following month, respectively. We then define the direction ratio for peer i, as direction(i) = d _n /d _p . ¹⁷ As an extreme example, if the enterprise went bankrupt one month ago, then n _n  = 0, so that direction(i) = 0, which means that peer i departed from the P2P networks. The final authority value and hub value of peer i can then be computed as follows:

$$a_i=\mathrm{direction}(i)*\sum _{j:e_{\mathrm{ij}}\in E}f(t_{\mathrm{ij}})*w_{\mathrm{ij}}*h_j,h_i=\mathrm{direction}(i)*\sum _{j:{e_j}_i\in E}f(t_{\mathrm{ji}})*w_{\mathrm{ji}}*a_j$$

(6)

According to the equations (2) and (6), the ultimate authority value and hub value of peer i can be computed as follow:

$$\mathbf{A}[i]=\mathbf{D}[i]*\mathbf{x}[i],\mathbf{H}[i]=\mathbf{D}[i]*\mathbf{y}[i]$$

(7) where i acts as enterprise peer i in distributed P2P networks, x and y are the vectors of the result of formula (2), and matrix D indicates the potential significance of peer i in the future, called the direction vector.

Although the proposed method above takes into account that the reputation of each peer changes both in the past and in the future, the method is not effective for another special circumstance in which a new peer joins the P2P networks of manufacturing service recommendation, since the new peer neither has manufacturing service recommendation for any peer nor has obtained recommendation from others. For the formula direction(i) = d _n /d _f , d _f is 0, so that direction(i) has no meaning. Therefore, we adopt a simple and rational method for estimating the original authority value and hub value of a new peer. The average authority value and average hub value of all peers in a distributed and dynamic P2P network can be computed as the following original value:

$$h(o)=\sum _{i=1}^n{h}_i/n,\mathrm{\hspace{1em}\hspace{1em}\hspace{1em}}a(o)=\sum _{i=1}^n{a}_i/n$$

(8)

If the new peer provides valuable services, the authority value and hub value will be rated higher by customer peers in the near future. Finally, we gain the global, dynamic, and reasonable reputation value of each peer by implementing equation (2), and recommend the optimal manufacturing service to the consumer enterprises.

Manufacturing service recommendation process leveraging the HITS algorithm

The manufacturing service recommendation process employing the HITS algorithm is shown in Figure 3. Firstly, a consumer enterprise publishes a service query, and candidate supplier peers are discovered from the manufacturing service registry by employing our previously researched P2P architecture.^2–4 Secondly, the relative weight of each trust rating is determined from the historical feedback trust rating registry, and we employ the proposed time-aware HITS algorithm to calculate the authority value and hub value of each candidate enterprise peer through several iterations. Finally, the higher reputation values of the enterprises are recommended to the enterprise that requires them. Meanwhile, the result of new reputation values and transaction dates is preserved in the historical feedback trust rating registry for future reference for manufacturing service recommendation.

Figure 3.

The process of manufacturing service recommendation by leveraging the HITS algorithm.

Case study and evaluation experiment

In this section, a case study is presented to illustrate how our proposed method effectively and efficiently to calculate the reasonable and accurate reputation values of manufacturing enterprises. An evaluation experiment reveals that both of recall rate and precision rate are improved, employing our proposed method for recommending manufacturing service in distributed P2P networks. We employ a Java-based object-oriented software prototype to implement our proposed method.

A case study of our proposed method

The case study aims to calculate the reasonable and accurate reputation values of manufacturing enterprises by employing the proposed method. In it, an enterprise posts a query searching for the optimal service. The candidate enterprises are then discovered from a registry of manufacturing service, through the use of P2P-based manufacturing system architecture we developed earlier.^2–4 From these results of the services, we obtain information on candidate enterprises, as well as corresponding trust ratings for previous transactions of enterprises, from a registry of historical feedback trust rating. It is necessary to sensibly decide which of them have the higher reputation values.

In this paper, we select eight candidate enterprises (CE1 to CE8), and use their relevant information as experimental subjects. The specific process of calculating the reasonable and accurate reputation values of the candidate enterprises involves the following steps:

(1) Preliminary data preparation

The previous services' transactions information of the eight candidate enterprises can be extracted from their historical feedback trust ratings, including mutual trust ratings and rating weights, which have been presented in “Traditional HITS algorithm” section. The initial vector x of equation (2) are also assigned as 1/8. The relative weights of historical services transactions among the eight candidate enterprises have been shown in Table 1. Table 2 shows the month of previous transactions, and the number of enterprises that have used the services of the eight candidate enterprises in September and October is shown in Table 3.

• (2) Reputation calculation

  Table 2.

  Months of previous transactions services for the eight candidate enterprises.

  

  	CE1	CE2	CE3	CE4	CE5	CE6	CE7	CE8
  CE1	Null	August	September	Null	Null	June	October	June
  CE2	Null	Null	July	Null	Null	October	Null	Null
  CE3	Null	Null	Null	Null	October	Null	Null	September
  CE4	Null	Null	October	Null	September	Null	Null	October
  CE5	Null	Null	August	Null	Null	Null	October	Null
  CE6	July	Null	Null	Null	Null	Null	August	Null
  CE7	Null	Null	Null	August	Null	Null	Null	Null
  CE8	Null	Null	Null	Null	June	Null	September	Null

Table 3.

Number of historical consumed services provided by eight candidate enterprises in September and October.

	CE1	CE2	CE3	CE4	CE5	CE6	CE7	CE8
September	3	3	1	4	2	3	2	3
October	6	2	3	6	3	0	4	3

The final objective of this step is to calculate the reputation value of the eight candidate enterprises by leveraging the method above. The time of the case study was November, and thus we obtained the value of function f(t _ij ) by combining Table 2 and equation (3). We compute the corresponding elements m _ij of adjacency matrix M by using equation (5). The M matrix is as follows:

According to the above matrix M in Table 4, we compute vectors x[8] = [0.062, 0.022, 0.177, 0.001, 0.285, 0.09, 0.234, 0.128] ^T and y[8] = [0.107, 0.095, 0.193, 0.158, 0.162, 0.079, 0.001, 0.204] ^T by leveraging equation (2). The direction vector D[8] = [6/3, 2/3, 3/1, 6/4, 3/2, 0/3, 4/2, 3/3] ^T . The authority values and hub values of the eight candidate enterprises can then be calculated with equation (7), which not only considers previous trust ratings, but also the potential significance of the enterprises in the future, making it possible to calculate the result as follows: A[8] = [0.124, 0.014, 0.532, 0.002, 0.428, 0, 0.467, 0.128] ^T ; H[8] = [0.215, 0.063, 0.579, 0.237, 0.243, 0, 0.002, 0.204] ^T . The authority value and hub value of CE6 are all 0, implying that CE6 has departed from the P2P networks of service recommendation. Obviously, the highest reputation value of the eight candidate enterprises is CE 3, since the authority value of CE 3 is 0.532 and the hub value of CE 3 is 0.579.

Table 4.

Adjacency matrix M.

	E1	E2	E3	E4	E5	E6	E7	E8
E1	0	0.201	0.223	0	0	0.247	0.246	0.082
E2	0	0	0.330	0	0	0.670	0	0
E3	0	0	0	0	0.721	0	0	0.279
E4	0	0	0.208	0	0.376	0	0	0.416
E5	0	0	0.551	0	0	0	0.449	0
E6	0.738	0	0	0	0	0	0.217	0
E7	0	0	0	1.0	0	0	0	0
E8	0	0	0	0	0.425	0	0.575	0

Evaluating the recall rate and precision rate of the optimal manufacturing service recommendation

In this section, we employ three different models of calculating QoS value to rank the QoS values of manufacturing services, which have satisfied the functional requirements. The Top-K manufacturing services with higher QoS values are chosen to recommend to consumer enterprises for further selection. We assume that tp manufacturing services are adopted as part of candidate services from Top-K manufacturing services, and fp manufacturing services are adopted as another part of candidate services from outside of the top-K services. The formulas of recall rate and precision rate are represented as follows ²⁰ :

$$\mathrm{precision}=\frac{\mathrm{tp}}{\mathrm{tp}+\mathrm{fp}},\mathrm{recall}=\frac{\mathrm{tp}}{K}$$

(9)

According to the previous model of calculating QoS value, ²¹ the aforementioned three different models of calculating QoS value are illustrated as follows:

(1) The first model of the QoS properties is only considering manufacturing service price, response time and lead time, and the weights of them are assigned as 0.4, 0.3, and 0.3, respectively.

(2) The second model not only contains the three properties of the first model but also considers the reputation value of supplier enterprise as the QoS attribute. The reputation values of the enterprises are calculated through using a MCDMM method which has considered five trust parameters to evaluate reputation value of each enterprise for manufacturing service recommendation in P2P networks. ³ However, the MCDMM method has ignored the influence of the global and dynamic of reputation values in distributed P2P networks. The weights of the price, response time, lead time, and reputation value are allocated as 0.3, 0.2, 0.2, and 0.3, respectively.

(3) The arrangements of third model are the same as second model, including the four attributes and their corresponding weights. However, the reputation values of the supplier enterprises are calculated by employing our proposed method that not only considers the characteristic of global reputation propagation among transaction peers but also takes account of temporal dimensions of reputation values of enterprises in distributed P2P networks. Besides, DR is an important factor of time-aware HITS-based reputation evaluation method to recommend optimal manufacturing service. We make trial tests on three parameters (i.e. 0.3, 0.5 and 0.7) to find a better parameter.

In this experiment, there are 256 adopted peers of P2P service supply chain networks and 1246 trust ratings of historical services transactions, K is arranged as 8. In order to make evaluation more objective, 10 similar experiments are performed to compare the aforementioned three different QoS models regarding their recall rates and precision rates, and the result of comparisons of average recall rates and precision rates is demonstrated in Figure 4.

Figure 4.

Comparisons of average recall rates and precision rates.

The result illustrates that both of average recall rates and precision rates are improved, when our proposed method with DR = 0.3 is employed for recommending manufacturing service. The first reason is because the first model of computing QoS value only considered conventional three aspects, while ignored the reputation value of enterprise that is a significant factor in recommending optimal manufacturing service. Although the second model has considered reputation value of enterprises in calculation of QoS values, the local and static drawbacks of traditional reputation evaluation method have not been well addressed. Therefore, the time-aware HITS method with DR = 0.3 is more reasonable and effective for manufacturing service recommendation in distributed P2P networks.

Conclusion

In this paper, we propose a time-aware HITS-based reputation evaluation method for recommending the optimal manufacturing service in distributed P2P networks. Using the time-aware HITS algorithm is a method that is both novel and feasible, as it not only enables the dissemination of the reputation values of transaction enterprises but also takes account of temporal dimensions in the process of computing reputation values, making it possible to obtain the global and dynamic reputation values of each enterprise peer to recommend reliable service. The primary contributions of this paper can be summarized as follows:

The elaboration of reputation dissemination based on the HITS algorithm in distributed P2P networks, which considers the global reputation value of mutual transaction ratings among manufacturing enterprises in order to evaluate reputation values and make the results more reasonable.

The development of a methodology for manufacturing service recommendation, based on the time-aware HITS algorithm, which adds timeliness into the HITS algorithm to calculate the dynamic and global reputation values of enterprises, and make manufacturing service recommendation more reasonable and scientific in distributed P2P networks.

The proposed approach, however, still has limitations that should be further explored and improved. Social network information also has a great influence on manufacturing service recommendation, which has been ignored in this article. Our future work will establish another feasible method for optimizing manufacturing service recommendation by combining the above approach and SNA.