Towards Long-Term Coverage and Video Users Satisfaction Prediction in Cellular Networks

Abstract

Network operators are interested in continuously monitoring the satisfaction of their customers to minimise the churn rate: however, collecting user feedbacks through surveys is a cumbersome task. In this work we explore the possibility of predicting the long-term user satisfaction relative to network coverage and video streaming starting from user-side network measurements only. We leverage country-wide datasets to engineer features which are then used to train several machine learning models. The obtained results suggest that, although some correlation is visible and could be exploited by the classifiers, long-term user satisfaction prediction from network measurements is a very challenging task: we therefore point out possible action points to be implemented to improve the prediction results.

Crowdsourcing or Network KPIs? A Twofold Perspective for QoE Prediction in Cellular Networks

“…Many authors during the last decade investigated and evaluated the feasibility of predicting both short-term [7], [3] and long-term [8], [2], [9] customers QoE relative to different network services. Short-term QoE concerns individual and time-limited sessions in which users are instructed to use a service (e.g., watching a video content on YouTube) under controlled network environments and are later asked about the quality of their experience.…”

“…We pre-process the dataset by considering only the users who replied to QoE feedbacks (i.e., whose identifiers are present in the Q c or Q v datasets), resulting in about 2k users for each scenario. According to our previous study [9], we restrict our attention to a subset of features (10 for network coverage and 14 for video streaming), summarized in Table I. Regardless of the scenario, features are computed independently for each user, considering different periods of d days before the date of the survey.…”

“…The variable d relates to the so-called user memory, i.e., for how long a disservice is able to impact the response to a survey. Similarly to what done in [9], we do not use a single value of d, but we set its value to 30, 60 and 90 days before the survey response, each time computing the corresponding subset of features. 1 Finally, we concatenate the three subsets, obtaining a total of 30 and 42 features to be used for each user in network coverage and video streaming scenarios, respectively.…”

“…At the end of this process, we obtain 47 features to be used for both network coverage and video streaming QoE prediction. Similarly to what done in [9], we apply to the computed features a log-like transformation, to make their statistical distributions more similar to Gaussian. Finally, features are standardized to mean and variance, a preprocess which benefits ML methods working with normally distributed inputs.…”

“…The problem we consider has the form of a binary classification problem, such that each user can be classified as either satisfied (class 0, negative case) or unsatisfied (class 1, positive case or alarm). We take therefore a supervised learning approach: among the several available ML classifiers to be trained we select the Random Forest algorithm, which is widely known to perform well in general and has been successfully applied in the past for similar problems [3], [9]. To compare the two approaches to QoE prediction we proceed as it follows: 1) First, we select the subset of users who appear in both datasets, i.e.…”