基于偏差的邻域模型：云与IoT服务上下文感知QoS预测

101 浏览量更新于2024-08-31 收藏 866KB PDF 举报

"Deviation-based neighborhood model for context-aware QoS prediction of cloud and IoT services" 随着互联网上云服务（Cloud services）和物联网服务（IoT services）的爆发式增长，如何为用户提供个性化、高质量的服务选择成为一个日益重要的问题。服务质量（Quality of Service, QoS）预测是解决这一问题的关键，它借鉴了推荐系统的思想，即协同QoS预测。在这个框架下，本文提出了一种新颖的基于偏差的邻域模型，旨在利用大众智慧进行QoS预测。现有的工作主要关注于传统的邻域模型，而我们的方法引入了一个双层结构，这使得模型能够更有效地适应上下文感知（Context-aware）的QoS预测。上下文感知是指根据用户的环境、时间、位置等多维度信息来提供定制化的服务质量预测。在我们的方法中，"偏差"是关键概念，用于度量用户的历史行为与服务实际表现之间的差异。通过分析这些偏差，模型可以更好地理解用户对服务质量的需求和预期，从而提供更为精确的预测。我们的模型首先通过收集并分析大量用户和服务的交互数据，构建一个全局的服务质量数据库。然后，利用用户的历史行为数据，找到与当前用户具有相似QoS偏好和上下文条件的邻居用户。接下来，通过对这些邻居用户的服务评价进行加权平均，我们得到一个基础预测值。然而，由于每个用户的偏好可能有微小但重要的差异，我们引入偏差调整机制，对基础预测值进行精细化调整，从而提升预测的准确性。实验结果表明，提出的偏差基邻域模型在预测精度和全局优化效率方面优于传统方法。尤其在处理具有细粒度调整需求的场景下，我们的模型表现出更优的性能。此外，由于模型考虑了上下文因素，它能更好地满足不同环境和条件下用户对服务质量的个性化需求。这篇研究论文为云服务和物联网服务的QoS预测提供了新的视角和解决方案，通过结合用户的行为偏差和邻域学习，实现了更准确、更智能的预测，有望为服务提供商和用户在服务选择过程中提供有力的支持。未来的研究可以进一步探索如何动态更新邻域模型，以适应快速变化的用户需求和技术环境。

552 H. Wu et al. / Future Generation Computer Systems 76 (2017) 550–560

2.2. QoS prediction based on matrix factorization

Different from neighborhood-based approaches, matrix factor-

ization approaches have attractive accuracy and scalability thus

recently become popular in recommender systems [16]. A typical

model associates each user u with an user-factors vector A

∈ R

and each item i with an item-factors vector B

∈ R

. The prediction

is done by taking an inner product, i.e.,

= A

. To exploit matrix

factorization for QoS prediction, Zheng et al. [8,12] using proba-

bilistic matrix factorization (PMF [33]) approach to decompose the

QoS matrix. For identifying latent factors, A and B, a least-square’s

problem like (2) are built and solved using gradient descent [16].

min

A,B



(u,i)∈E

− A

)

+ λ

∥A

∥

+ λ

∥B

∥

. (2)

Matrix factorization can partially alleviate sparsity-sensitive

problem of collaborative filtering thus increase the accuracy of

QoS prediction. For the last two years, numerous efforts have been

made on improving MF-based models. These works concentrate

on utilizing additional information, such as spatial and temporal

information associated with users or services. Zhang et al. [22]

use collective matrix factorization that simultaneously factor

the user–service quality matrix, service category and location

context matrices. Zhang et al. [23] factorize user–service–time

matrix of QoS using non-negative tensor factorization with time

information. Yin et al. [25] develop a location-based regularization

framework for PMF prediction model. Lo et al. [24] exploit PMF

prediction model with a location-based pre-filtering stage on QoS

matrix. He et al. [34] develop location-based hierarchical matrix

factorization. Yu et al. [26] experience trace-norm regularized

matrix factorization. Following neighborhood-integrated matrix

factorization(NIMF) [8], Qi et al. [30] propose a MF-based method,

integrating both user network neighborhood information and

service neighborhood information, to predict personalized QoS

values.

Matrix factorization can provide accurate predictions while sac-

rificing explainability, as the learned latent factors are unexplain-

able. Lack of explainability weakens the ability to persuade users

and help users make better decisions in practical systems. Also,

data sparsity has a negative impact on these methods, as data be-

comes extremely sparse, the prediction performance will be not

optimistic.

In addition, neural network based deep learning technology

has been proposed for collaborative filtering, e.g., Restricted

Boltzmann Machines [35]. Similar to matrix factorization, deep

learning methods allow us to discover the latent features

underlying data. It may indicate a new trend in the QoS prediction,

however, deep learning methods always incur high computational

overheads, and suffer from the similar problems in matrix

factorization techniques.

3. Deviation-based neighborhood model for QoS prediction

To cope with existing drawbacks of CF-based prediction

methods, we suggest using machine learning techniques to build

neighborhood model for QoS prediction. New models allow

an efficient global optimization scheme and exploit different

baseline estimate components to improve prediction accuracy. To

distinguish users from services, we take different indexing letters:

for users u, v, and for services i, j. The notation q

indicates

a known quality-score observed by user u on service i and the

notation

represents the predicted value of q

. The (u, i) pairs

for which q

is observed are stored in the set E = {(u, i)|q

=

null}. Usually the vast majority of QoS scores are missing. To

combat overfitting in learning prediction model on the sparse data,

models are regularized so estimates are shrunk towards baseline

estimates [16].

3.1. The framework

User-oriented methods estimate unknown quality based on

recorded QoS of like minded users. Analogously, in service-

oriented methods, a QoS value is estimated using known QoS made

by the same user on comparable services. In cloud/IoT computing

environment, the context with users is more complex and dynamic

than that of services. Prediction leveraged by similar users other

than services is more practical. Therefore, our focus is on user-

oriented approaches, but parallel techniques can be developed in

a service-oriented fashion, by switching the roles of users and

services.

We develop the QoS prediction model on the basis of

neighborhood-based collaborative filtering [16], which allows an

efficient global optimization scheme and offers improved accuracy.

To facilitate global optimization, we would want to abandon

user-specific similarity, s(u, v) in (1), in favor of global weights

independent of a specific user. The weight from user v to user u

is denoted by x

are able to be learned from the data through

optimization. By this, we can overcome the weaknesses with

existing neighborhood-based models. An initial sketch of the

model describes each quality score q

by:

= b



v∈N

− b

), (3)

where N

is the neighbor set of u, b

is the baseline estimate that

we will gradually construct considering different factors.

As for the interpretation of weights, usually they represent

interpolation coefficients relating unknown quality score to the

existing ones in a traditional neighborhood model [16] (recall

s(u, v) in (1)). Here, we adopt them in a different viewpoint that

weights represent offsets to baseline estimates and residual, q

−

, is viewed as the coefficients multiplying those offsets. For two

similar users u and v, x

is always expected to get high, and vice

versa. So, our estimate will not deviate much from the baseline

estimate by a user v that accessed i just as expected (q

− b

around zero), or by a user v that is not known to be predictive on

u (x

is close to zero).

Generally, we can take all users in N

other than u, however, this

would increase the number of weights to be estimated. In order to

reduce complexity of the model, we suggest pruning parameters

corresponding to unlikely user–user relations. Let N

be the set of

k users who are most similar to u, as determined by the similarity

measure s(u, v). Further, we let N

(i;u)

, N

∩N

, where N

is the set

of users have used the service i. Now, when predicting q

according

to formula (3), it is expected that the most influential weights will

be associated with users similar to u. Hence, we replace (3) with:

= b

+ |N

(i;u)

−



v∈N

(i;u)

− b

). (4)

When k = ∞, rule (4) coincides with (3). When k = 0,

= b

However, for other values of k, it offers the potential to significantly

reduce the number of variables involved. This final prediction rule

permits fast online prediction, since more computational works,

such as similarity calculation and parameter estimation, have

been made in the pre-processing stage. Recall that unlike matrix-

factorization, the neighborhood models allow a direct explanation

of their predictions, and do not require re-training the model for

handling new services.

3.2. Components for baseline estimation

Typical QoS data exhibit large user and service effects-

i.e., systematic tendencies for some users to achieve better QoS

than others, and for some services to receive better QoS than

剩余10页未读，继续阅读

weixin_38536841

粉丝: 3
资源: 946

基于偏差的邻域模型：云与IoT服务上下文感知QoS预测

"研究云计算中的频偏估计算法及FPGA实现

"Matlab常用算法程序: 时间序列预测算法

"基于S-PLC的厂用气管网压力控制系统设计(完整版)详解

Fast hybrid fitting energy-based active contour model for target detection

Gestion-Pharmacie-en-cours-deviation-

boilerplate-mean-variance-standard-deviation-calculator

Deviation-Devo7e-v3.0.0 固件

deviation-devo8_12

deviation-devo7e-v4.01cn_by_Mckay

freecodecamp-boilerplate-mean-variance-standard-deviation-calculator:freecodecamp样板均值方差标准偏差计算器

最新资源