Multi-objective optimization based ranking prediction for cloud service recommendation

Highlights

• Two multi-objective optimization algorithms for service recommendation are proposed.

• The two algorithms both can highly increase the diversity of service recommendation.

• The two algorithms both can hold the similar recommendation accuracy.

• The two algorithms can dock seamlessly with some other effective prediction algorithms.

• The complexity of the two algorithms can be comparable with the traditional service recommendation ones.

Abstract

Performing effective ranking prediction for cloud services can help customers make prompt decisions when they are confronted by a large number of choices. This can also enhance web service user satisfaction levels. Improving ranking prediction of QoS-based services continues to be an active topic of research in cloud service recommendation. Most service recommendation algorithms focus on prediction accuracy, ignoring diversity, which also may be an important consideration. In this paper we view service recommendation as a multi-objective optimization problem, and give two modified ranking prediction and recommendation algorithms that simultaneously consider accuracy and diversity. Existing algorithm recommendations can be made much more diverse by adjusting weights on service origin and substantially reducing the risk of inappropriate recommendations. Our experiments show that the algorithms we propose can yield greater diversity without greatly sacrificing prediction accuracy.

Introduction

With fast technology advances in cloud computing, many information resources have been encapsulated and released as cloud services on public servers such as Amazon and Tmall [1]. Since there exist a large number of available cloud services [2], [3], it becomes increasingly common that a given service request from cloud users can be fulfilled by multiple cloud services — while these services share the same or similar functionality to meet the user's main service requirement, each poses non-functional features that are important for the users, e.g., QoS (Quality of Service) and diversity of the services [4], [5].

To address this issue, a number of service recommendation approaches have been proposed to assist service selection. A service recommendation system often ranks a small number of candidate services that are most likely to match the user's requirements. The ranking cannot be made just based on their functionality metric. For example, two shopping services A and B may be able to satisfy the shopping demand. While the check out speed of A is 10% faster than that of B, we may not always rank A first. If B has a set of additional features, such as production comparison, historic price comparison, that can significantly enhance the clients' shopping experience, it would be natural to rank B over A. However, if B has security concerns in checking out process, these additional features would become less important. The additional non-functional features are often referred to as auxiliary decision support system (DSS) [6], [7].

Due to the complexity in service ranking and selection, it remains one of the most challenging tasks in cloud computing to choose the cloud service that not only fulfills the required functionality but also has the best QoS and diversity match [8]. On the one hand, not all QoS values are available for service recommendation — these values are either too expensive to collect or lost over the time [9]. The service ranking task has to adapt to work with partial data set. On the other hand, publicly available cloud services target at servicing a large number of users and thus are often less flexible to adapt to a particular user's demand. To improve the recommendation accuracy in the system level, the diversity of recommended services is often ignored [10], [11], and vice verse. That is, few recommendation approaches can effectively address both the ranking accuracy and the diversity requirements [12]. While recent advances proposed recommendation systems that consider both in commendation, these systems often adopt greedy approaches, which may result in sub-optimal recommendation [13].

In this paper, starting from the high-performance CloudRank Algorithm presented by Zheng et al. [13], we propose novel service recommendation schemes to address the above challenges. With the goal to improve ranking accuracy and to meet the diversity requirements from cloud users, we focus on improving user satisfaction and formulate it as a multi-object optimization problem. We devise two rank prediction algorithms to find the best tradeoff between the competing optimization sub-goals. Our experimental results show that the proposed schemes effectively boost the spatial diversity of service origins while achieving good recommendation accuracy.

The remainder of this paper is structured as follows. At first, in Section 2, we review the related work on ranking prediction for cloud service recommendation. Then, starting from the utility weighting, we present two novel ranking approaches for the suggested services that commonly take the QoS quality and recommending diversity into consideration in Section 3. After that, extensive experiments on some real-world data sets are carried out to validate the performance of the proposed service ranking prediction methods in Section 4. At last, in Section 5, we end this paper with some concluding remarks and further point out some feasible directions in the future.