Location-Aware and Personalized Collaborative Filtering for Web Service Recommendation

Location-Aware and Personalized Collaborative Filtering for Web Service Recommendation

ABSTRACT

           The proposed method leverages both locations of users and Web services when selecting similar neighbors for the target user or service,  Collaborative Filtering (CF) is widely employed for making Web service recommendation. CF-based Web service recommendation aims to predict missing QoS (Quality-of-Service) values of Web services. Although several CF-based Web service QoS prediction methods have been proposed in recent years, the performance still needs significant improvement. Firstly, existing QoS prediction methods seldom consider personalized influence of users and services when measuring the similarity between users and between services. Secondly, Web service QoS factors, such as response time and throughput, usually depends on the locations of Web services and users. However, existing Web service QoS prediction methods seldom took this observation into consideration. In this paper, we propose a location-aware personalized CF method for Web service recommendation. The proposed method leverages both locations of users and Web services when selecting similar neighbors for the target user or service. The method also includes an enhanced similarity measurement for users and Web services, by taking into account the personalized influence of them. To evaluate the performance of our proposed method, we conduct a set of comprehensive experiments using a real-world Web service dataset. The experimental results indicate that our approach improves the QoS prediction accuracy and computational efficiency significantly, compared to previous CF-based methods.

EXISTING SYSTEM

          Location-Aware and Personalized Collaborative Filtering for Web Service Recommendation existing Quos prediction methods seldom consider personalized influence of users and services when measuring the similarity between users and between services However, existing Web service QoS prediction methods seldom took this observation into consideration. We conducted an experiment to evaluate the prediction time of our method, and compare it with some existing

EXISTING SYSTEM ALGORITHMS

Collaborative filtering is one of the most popular recommendation techniques, which has been widely used in many recommender systems. In this section, we give a brief survey of CF algorithms, and summarize recent work on CF-based Web service recommendation.

PROPOSED SYSTEM

v We proposed an enhanced measurement for computing QoS similarity between different users and between different services. The measurement takes into account the personalized deviation of Web services’ QoS and users’ QoS experiences, in order to improve the accuracy of similarity computation.

v Although several CF-based Web service QoS prediction methods have been proposed in recent years, the performance still needs significant improvement

v we propose a location-aware personalized CF method for Web service recommendation.

v The proposed method leverages both locations of users and Web services when selecting similar neighbors for the target user or service

v To evaluate the performance of our proposed method, we conduct a set of comprehensive experiments using a real-world Web service dataset.

v Based on the above enhanced similarity measurement, we proposed a location-aware CF-based Web service QoS prediction method for service recommendation.

v We conducted a set of comprehensive experiments employing a real-world Web service dataset, which demonstrated that the proposed Web service QoS prediction method significantly outperforms previous well-known methods.

PROPOSED SYSTEM ALGORITHMS

Ø We first formally define notations for the convenience of describing our method and algorithms.

Ø The Top-K similar neighbor selection algorithm is often employed

Ø The Top-K similar neighbor selection algorithm can be employed to select K Web services that are most similar to the target Web service

Ø We can see that the algorithm first searches local users for similar users.

Ø This algorithm has a high probability of finding users similar to the active user in his/her local region.

Ø Prediction coverage is also an important metric for evaluating a QoS prediction algorithm

ADVANTAGES

          In addition to the prediction accuracy, another advantage of our method is its high efficiency of QoS prediction. This indicates that our method is more scalable than traditional CF methods when applied to large-scale service recommender systems. This indicates that our method is more scalable than traditional CF methods when applied to large-scale service recommender systems. The reason is that, in most cases we can limit similar neighbor searching to a small subset of users (or Web services), especially when K is small.

System Architecture

Overview of our Web service recommendation method

Influence of user location on QoS prediction

MODULE DESCRIPTION

          Web services

          Collaborative Filtering (CF)

          Web Service Recommendation

Incorporating QoS Variation into User and Service Similarity Measurement

Incorporating Locations of Users and Services into Similar Neighbor Selection

Ø User location information handler

Ø Service location information handler

Ø Find similar users

Ø Find similar services

Ø User-based QoS prediction

Ø Service-based QoS prediction

Ø Hybrid QoS prediction

Ø Recommender

Location Representation

Location Information Acquisition

Location Information Processing

Web services

CF-based Web service recommendation aims to predict missing QoS (Quality-of-Service) values of Web services. With the prevalence of Service-Oriented Architecture (SOA), more and more Internet applications are constructed by composing Web services. As a consequence, number of Web services has increased rapidly over the last decade.

Collaborative Filtering (CF) is widely employed to rec-ommend high quality Web services to service users. Based on the fact that a service user may only have in-voked a small number of Web services, CF-based Web service recommendation technique focuses on predicting missing QoS values of Web services for the user.

Collaborative Filtering (CF)

          Collaborative filtering is a method of making automatic predictions (filtering) about the interests of a user by collecting preferences or taste information from many users (collaborating)

CF techniques can be generally decomposed into two categories: model-based and memory-based [12],[13]. Memory-based CF is also named neighborhood-based CF. Depending on whether user neighborhood or item neighborhood is considered, neighborhood-based CF can further be classified into user-based and item based.

For example, using the temporal context, a travel recommender system would provide a vacation recommendation in winter very different from the one provided in summer. They demonstrated that incorporating contextual information in essence would improve both the effectiveness and the efficiency of a recommender system.

Web Service Recommendation

          Various recommendation techniques have recently been applied to Web service recommendation, such as the content- based  link prediction-based. Their argued that, for every pair of ac-tive user and target Web service, both the QoS experience of the users similar to the active user and the QoS values of the services similar to the target service can be em-ployed for QoS prediction. However, these previous ap-proaches failed to exploit the characteristics of QoS in the similarity computation. Based on the traditional CF approaches, several en-hanced methods have been proposed to improve the pre-diction accuracy. This is probable if the Web services are deployed in a high performance Cloud environment. If the QoS is good enough (as in this instance), a small variation of QoS values over all users is likely to be ob-served. Some Web services may have a very poor QoS for all users.

Incorporating QoS Variation into User and Service Similarity Measurement

          Previous QoS prediction methods assume that the co-invoked Web services have equal contribution weights when computing similarity between two users. We argue that the personalized characteristics (e.g., QoS variation) of both Web services and users should be incorporated into measuring the similarity among users and services. Web service QoS factors, such as response time, avail-ability and reliability, are usually user-dependent. From different Web services, we can derive different personal-ized characteristics, based on their QoS values, as perceived by a variety of users. Some Web services may have a very good QoS for all users.

For example, the availabil-ity is always 100%. This is probable if the Web services are deployed in a high performance Cloud environment. If the QoS is good enough (as in this instance), a small variation of QoS values over all users is likely to be ob-served. Some Web services may have a very poor QoS for all users. For example, the availability is always below 50%. This is probable if the Web services are deployed in a network environment with poor performance and bandwidth. These Web services are also likely to have small variation of QoS values over different users. Many other Web services may have a relatively large variation of QoS over different users. For example, the availability varies from 50% to 100% for different users. These Web services are considered to be user-sensitive. The following example explains why Web services with different QoS variations could contribute differently when computing the similarity between service users.

Incorporating Locations of Users and Services into Similar Neighbor Selection

          Web services are deployed on the Internet. Thus, QoS of Web services (such as response time, reliability and throughput) is highly dependent on the performance of the underlying network [33]. If the network between a target user and a target Web service is of high performance, the probability that the user will observe high QoS on the tar-get service will increase. There are several factors affecting the network performance between the target user and the target service. The most important factors include network distance and network bandwidth, which are highly relevant to locations of the target user and the target service. When the user and the service are located at different networks which are far away from each other on the Internet, network performance is likely to be poor due to both the transfer delay and the limited bandwidth of links between different networks.

In contrast, when the user and the Web service are located in the same network, the user is more likely to observe high network performance.

User location information handler: This module obtains location information of a user including the network and the country according to the user’s IP address. It also provides support for efficient user-querying based on location.

Service location information handler: This handler acquires additional location information of Web services according to either their URLs or IP addresses. The location information includes the network and the country in which the Web service are located. It also provides functionalities for supporting efficient locationbased Web service query.

Find similar users: This module finds users who are similar to the active user by considering both the users’ QoS experiences and locations. For accurate user similarity measurement and scalable similar user selection, we propose a weighted user-based PCC via exploring QoS variation of Web services and incorporate user locations into similar user selection.

Find similar services: In contrast to finding similar users, this module finds similar Web services for a target service, considering both QoS of Web services as well as service locations. A weighted service-based PCC for measuring similarity between services is proposed

User-based QoS prediction: After a certain number of similar users are identified for the active user, this function aggregates the QoS values they perceived on target Web services, and predicts the missing QoS values for the active user.

Service-based QoS prediction: After a certain number of similar services are identified for a target Web service, this function aggregates their QoS values to predict the missing QoS values for the active user

Hybrid QoS prediction: This function combines the userbased QoS prediction and the service-based QoS prediction results, making final QoS predictions. The cold-start problem and data-sparsity problem in QoS predictions are also addressed in this module

Recommender: After predicting missing QoS values for all candidate Web services, this function recommends Web services with optimal QoS to the active user

LOCATION INFORMATION REPRESENTATION, ACQUISITION, AND PROCESSING

This section discusses how to represent, acquire, and pro-cess location information of both Web services and ser-vice users, which lays a necessary foundation for imple-menting our location-aware Web service recommendation method.

Location Representation

We represent a user’s location as a triple (IPu, ASNu, CountryIDu), where IPu denotes the IP address of the user, ASNu denotes the ID of the Autonomous System (AS)1 that IPu belongs to, and CountryIDu denotes the ID of the country that IPu belongs to. Typically, a country has many ASs and an AS is within one country only. The Internet is composed of thousands of ASs that inter-connected with each other.

Generally speaking, intra-AS traffic is much better than inter-AS traffic regarding transmission performance, such as re-sponse time [34]. Also, traffic between neighboring ASs is better than that between distant ASs. Therefore, the Inter-net AS-level topology has been widely used to measure the distance between Internet users [34]. Note that users located in the same AS are not always geographically close, and vice versa. For example, two users located in the same city may be within different ASs. Therefore, even if two users are located in the same city, they may look distant on the Internet if they are within different ASs. This explains why we choose AS instead of other geographic positions, such as latitude and longitude, to represent a user’s location.

Location Information

 Acquisition Acquiring the location information of both Web services and service users can be easily done. Because the users’ IP addresses are already known, to obtain full location in-formation of a user, we only need to identify both the AS and the country in which he is located according to his IP address. A number of services and databases are available for this purpose (e.g. the Whois lookup service2). In this work, we accomplished the IP to AS mapping and IP to country mapping using the GeoLite Autonomous System Number Database3. The database is updated every month, ensuring that neither the IP to AS mapping nor the IP to country mapping will be out-of-date.

SIMILARITY COMPUTATION AND SIMILAR NEIGHBOR SELECTION

          In this section, we first formally define notations for the convenience of describing our method and algorithms. We then present a weighted PCC for computing similarity between both users and Web services, which takes their personal QoS characteristics into consideration. Finally, we discuss incorporating locations of both users and Web services into the similar neighbor selection.

Similar Neighbor Selection

Similar neighbor selection is a very important step of CF. Selecting the neighbors right similar to the active user is necessary for accurate missing value prediction. In conventional user-based CF, the Top-K similar neighbor selection algorithm is often employed [8]. It selects K users that are most similar to the active user as his/her neighbors. Similarly, the Top-K similar neighbor selection algorithm can be employed to select K Web services that are most similar to the target Web service. There are several problems involved, however, when applying the Top-K similar neighbor selection algorithm to Web service recommendation. Firstly, in practice, some service users have either few similar users or no similar users due to the data sparsity. Traditional Top-K algorithms ignore this problem and still choose the top K most ones. Because the resulting neighbors are not actually similar to the target user (service), doing this will impair the prediction accuracy. Therefore, removing those neighbors from the top K similar neighbor set is better if the similarity is no more than 0. Secondly, as previously mentioned, Web service users may happen to perceive similar QoS values on a few Web services. But they are not really similar.

Considering the location-relatedness of Web service QoS, we incorporate the locations of both users and Web services into similar neighbor selection.

User-based QoS Value Prediction

In this subsection, we present a user-based location-aware CF method, named as ULACF. Traditional user-based CF methods usually adopt for missing value predictions. This equation, however, may be inaccurate for Web service QoS value prediction for the following reasons. Web service QoS factors such as response time and throughput, which are objective parameters and their values vary largely. In contrast, user ratings used by traditional recommender systems are subjective and their values are relatively fixed [29]. Therefore, predicting QoS values based on the average QoS values perceived by the active user (i.e., r (u) ) is flawed. Moreover, Eq. (9) does not distinguish local and remote users that are similar to the active user. Intuitively, given two users that have the same estimated similarity degree to the target user, the user closer to the target user should be placed more confidence in QoS prediction than the other.

Item-based QoS Value Prediction

In this subsection, we present an item-based locationaware CF method, named as ILACF. Based on the similar consideration as ULACF’s, we use Eq. to compute the predicted QoS value for a service based on the QoS values of its similar services .

Integrating QoS Predictions

Due to the sparsity of the user-item matrix, to make the missing value prediction as accurate as possible, it’s better to fully explore the information of similar users as well as similar services. Therefore, we develop a hybrid location- aware CF, named as HLACF, which integrated the user-based QoS prediction with the item-based QoS prediction. The following four cases will be considered in integrating QoS predictions