TL;DR
This paper tackles the challenge of inaccurate Quality of Service (QoS) prediction for web services in mobile internet environments, where data volatility is a significant issue. They propose a novel approach that preprocesses volatile QoS data, computes user similarity using an enhanced Pearson Correlation Coefficient, and then predicts QoS based on filtered "normal" data from similar users, outperforming existing methods.
What Problem Does This Paper Address?
The proliferation of mobile devices has led to the rise of the "mobile internet," where users increasingly rely on web services. However, unlike traditional internet environments, mobile internet connections are prone to issues like packet loss and delays, leading to highly volatile QoS data (e.g., response times, availability). Existing web service QoS prediction methods, which are accurate in stable internet environments, fail to perform well in this volatile mobile context. This inaccuracy hinders effective web service recommendation, leaving users unable to reliably choose high-quality services. The core problem is how to accurately predict web service QoS when the underlying data is inherently unstable.
Methods
The proposed approach, named WSQP (Web Service QoS Prediction), consists of three main stages:
- QoS Preprocessing Strategy: To mitigate the impact of volatile data, the paper applies a logarithmic transformation to the historical QoS data. This transformation aims to reduce the volatility and bring abnormal data points closer to normal values, making subsequent calculations more reliable.
- User Similarity Computation: The authors employ the Pearson Correlation Coefficient (PCC) to measure similarity between users. To address PCC's tendency to overestimate similarity based on a few common services, they introduce an enhanced PCC that incorporates a "significance weight." This weight considers the number of commonly invoked services relative to the total services invoked by each user, effectively reducing the influence of sparse commonalities.
- QoS Prediction: After identifying the top-K similar users, WSQP filters their historical QoS data to retain only "normal" data points. This filtering is based on defining a "normal data interval" for specific QoS attributes (like response time or throughput) that captures the typical range of values. The final QoS prediction is then computed by taking a weighted average of the normal QoS data from these similar users.
Datasets
The authors implemented their approach and conducted experiments using the WSDream dataset. This dataset is described as containing real-world data with nearly 1 million service response time records. It includes data from 4532 web services and 142 distributed users from Planet-Lab, observed over 64 time intervals. The paper also mentions using synthetic datasets for experiments.
Key Insights and Ideas
- Volatility is the primary culprit: The paper strongly emphasizes that data volatility, not the inherent accuracy of prediction algorithms, is the main reason for poor QoS prediction in mobile internet environments.
- Log transformation is surprisingly effective: A simple logarithmic transformation of QoS data significantly reduces volatility, making it a crucial preprocessing step.
- Enhanced PCC for better user similarity: Simply using PCC can be misleading. Incorporating the number of common services invoked relative to individual usage (the "significance weight") leads to more accurate identification of truly similar users.
- Filtering for "normal" data is key: Identifying and using only the "normal" data points, as opposed to all historical data or randomly selected points, is essential for robust prediction.
Key Results
The paper presents experimental results comparing WSQP with two other approaches: WSRec (a collaborative filtering method) and AVG (likely a simple averaging method).
- Relative Error (RE): WSQP consistently achieved the smallest relative error across the majority of prediction experiments, indicating higher accuracy. Figure 5 shows that WSQP's RE values were maintained within a small range, while WSRec and AVG exhibited larger and more erratic errors.
- Mean Absolute Error (MAE): WSQP demonstrated the smallest MAE, suggesting it significantly improves prediction accuracy in mobile internet environments. Figure 6 visually supports this, showing WSQP's MAE consistently lower than WSRec and AVG.
- Root Mean Square Error (RMSE): Similar to MAE, WSQP also achieved the smallest RMSE, further confirming its superior prediction quality. Figure 7 illustrates this trend.
- Impact of User-Item Matrix Density: Experiments showed that higher density in the user-item matrix leads to more accurate and stable predictions. Densities below 50% resulted in unstable and inaccurate predictions.
- Optimal K Value: The parameter K (number of similar users) was analyzed, with K=2 found to be appropriate for the tested datasets, balancing accuracy and the need to consider sufficient similar users.
Limitations and Future Work
- Data Set Limitations: The authors acknowledge that despite their improvements, errors are still observed among similar users and in QoS prediction, possibly due to limitations in the datasets used.
- Future Work: The paper suggests that further research could explore more advanced methods to address the remaining prediction errors, potentially by investigating more sophisticated similarity metrics or prediction models. The impact of different QoS attributes and their interdependencies could also be a direction for future work.
Summary updated