Transfer Learning Based Diagnosis for Configuration Troubleshooting in Self-Organizing Femtocell Networks Wei Wang, Jin Zhang, Qian Zhang Department of Computer Science and Engineering, Hong Kong University of Science and Technology Email: {gswwang, jinzh, qianzh}@ust.hk Abstract—Diagnosis for configuration troubleshooting in fem- tocell networks is extremely important for end users and net- work operators. However, because the small-size femtocell only serves several users, the historical data are very scarce. The data scarcity makes traditional cellular troubleshooting solutions which require a large amount of historical data not applicable. In this paper, we propose a new framework based on transfer learning technology to address the data scarcity so as to enhance the accuracy of the diagnosis model. The proposed framework extracts additional diagnosis knowledge by transferring data information from other femtocells. Based on this framework, we design a Cell-Aware Transfer scheme (CAT), which splits data for each femtocell to further enhance the diagnosis accuracy. Extensive evaluations show that our approach can achieve higher accuracy than traditional methods in self-organizing femtocell network scenarios. Index Terms—automated diagnosis, femtocell, transfer learn- ing. I. I NTRODUCTION Femtocell is a key technology that provides ubiquitous network coverage to meet the demands for higher data rates for indoor cellular services. Given the importance of the mobile phone to our daily life, high-reliability and high- quality cellular network services are expected. However, unlike the operator-deployed traditional cellular networks, the user- deployed femtocells are not well-planned. Thus, inappropriate configuration problems occur more likely. Such problems can occur within homes due to users’ wrong operations or inappro- priate self-configuration algorithms beyond operator’s control and management, which we refer to as misconfigurations in this paper. Moreover, the number of the femto Access Points (femto AP) is much larger than macro Base Stations (macro BS), making the distributed-manner configurations in user- deployed femtocells more error-prone. Thus, efficient diag- nosis for configuration troubleshooting in femtocell is highly motivated. Existing diagnostic systems for traditional cellular networks fall short for femtocell diagnosis. These diagnostic systems are mainly based on classification reasoning methods, such as Bayesian Network (BN) [1] [2]. The traditional approaches for cellular networks are not applicable in femtocell networks because of the two data scarcity challenges: 1) the indoor femto AP only supports several users, so the data from user end are much less when compared with traditional cellular networks; 2) different from well-planned cellular networks, topology of femtocells is highly dynamic because Femto APs can be re-deployed and turned on/off by end users, so that historical data can be easily outdated. Without enough usable historical data, the accuracy of the classification models in traditional approaches cannot be guaranteed. In this paper, we propose a novel framework to address the data scarcity challenges. Our framework utilizes the existing transfer learning techniques to leverage historical data from other femtocells. However, there are challenges when leverag- ing the transfer learning techniques. Since the wireless envi- ronment (e.g. indoor propagation, neighboring cell layouts) of each femtocell is different, characteristics of femtocells can be very different from each other. How to extract useful information from the massive data needs to be carefully designed. Another challenge lies in that the target femtocell may be just deployed or re-deployed, the misconfiguration instances can be very rare. General transfer learning tech- niques based on historical misconfiguration instances are not accurate. To address these challenges, we design a Cell-Aware Transfer scheme (CAT). In the scheme, we weight the data for each femtocell in order to leverage the data from similar femtocells while eliminate the misleading information from the cells whose scenarios are quite different from the target femtocell. Considering that the misconfiguration instances may not be enough, CAT extracts information from measurement data when femto AP is properly configured. By considering these characteristics of femtocell networks, CAT can diagnose misconfigurations accurately even when the instances are rare in the past. The main contributions of this paper are as follows: 1) We propose a transfer learning framework for femtocell configura- tion troubleshooting. As the best of our knowledge, this is the first diagnosis framework proposed for the femtocell config- uration troubleshooting. 2) We develop a diagnosis scheme based on our transfer learning framework to address the data scarcity challenges in femtocell scenario. 3) Simulation results show that our scheme can achieve higher accuracy than traditional approaches for configuration troubleshooting in self-organizing femtocell networks. The rest of the paper is organized as follows. In Section