Multi-task learning is a framework that enforces different tasks to share their knowledge to improve the generalization performance. It is a long-standing active domain that strives to handle several core issues including which tasks are correlated and similar and how to share the knowledge among correlated tasks. Existing works usually do not distinguish the polarity and magnitude of feature weights and commonly rely on linear correlation, due to three major technical challenges in: 1) optimizing the models that regularize feature weight polarity, 2) deciding whether to regularize sign or magnitude, 3) identifying which tasks should share their sign and/or magnitude patterns. To address them, this paper proposes a new multi-task learning framework that can regularize feature weight signs across tasks, beyond the conventional framework for feature weight regularization. We innovatively formulate such sign-regularization problem as a biconvex inequality constrained optimization upon the multiplications among feature weights with slacks. We then propose a new efficient algorithm for the optimization with theoretical guarantees on generalization performance and convergence. Extensive experiments on multiple datasets show the proposed methods’ effectiveness, efficiency, and reasonableness of the regularized feature weighted patterns.

Eradicating poverty in all its forms everywhere remains as the number one Sustainable Development Goal of the 2030 Agenda for Sustainable Development. Developing countries face challenges in measuring the progress of poverty rates at the intra-urban level because they use traditional data collection methods such as censuses that are costly in time and resources. Therefore, local and central governments need ways of producing reliable, accurate, and up-to-date indicators to design effective policies about resource allocation for poverty alleviation programs that prioritize the most vulnerable citizens. For this purpose, we propose to exploit patterns observed in developing countries, where mobile phone usage is pervasive even among the poorest, and the dominant mobile subscription modality is prepaid to purchase airtime credit in advance. Our study analyzes a novel digital source with more than 9M mobile airtime top-up transactions to calculate meaningful indicators of customer economic activity. We aggregate it at the neighborhood spatial resolution to build a regression model to predict the neighborhood socioeconomic status (per capita income). Using a Linear Regression with Regularization L2 (Ridge), we can explain the neighborhood socioeconomic status with a prediction rate of up to 74% for urban neighborhoods of Guayaquil and Quito, Ecuador.

Quantifying income inequalities in developing countries faces challenges regarding data publicly available. Census data, collected every five or ten years, is the only source for socioeconomic indicators. Thus, local authorities need ways of producing more frequently updated indicators. Studies conducted for developed countries (Europe and USA) use Call Detail Records (CDRs) for such a purpose. In our study we propose to exploit patterns observed in developing countries, specifically in Latin America, where mobile phone usage is pervasive even among the poorest and the dominant modality for purchasing mobile airtime is the prepaid scheme (top-ups). We analyze more than 1M top-up transactions together with more than 5K online classified ads for housing sales to predict the socioeconomic status measured at an intra-urban level for 89 neighborhoods. Using a Linear Regression with Regularization L1 (Lasso), we can explain the economic status with a prediction rate up to 71% for urban neighborhoods in Guayaquil, Ecuador. Consequently, we show evidence that top-up transactions provide effective signals to characterize urban neighborhoods socioeconomic status.

Natural disasters have proven that governments, even in developed countries, have difficulties to get up-to-date data about not only affected people but also the location and intensity of infrastructure damage when a country is shook by the nature. Therefore, knowing how mobility patterns are changing, in the post disaster time-frame, is crucial in order to settle rescue centers and send help to the most affected areas. In this scenario, we analyze the relations between human mobility patterns and the effects of an earthquake that shook Ecuador on April 16th, 2016. We do so using more than 11 millions of aggregated call detail records provided by Telefonica. We propose a metric named Reach Score to build timeseries as a way to characterize the residents geographic reach according to their mobile activity. Next, we define the metric Reach Score change, RiSC to capture differences in mobility among two given dates. Our results show that these two metrics calculated on data from the day before and the day after the disaster reflect both the overall change in mobility at the province level and the intensity of infrastructure damage at canton level. In fact, we obtain a Pearson correlation coefficient of r = -0.819 between the metric RiSC and the infrastructure damage score taken from official data.

Customers nowadays are one online post away from their stores, specially when it comes to post-shopping experiences. This translates to large amounts of text messages to evaluate and process for big brands that aim to maintain a good quality of service as well as a digital channel of communication for their customers. Automating the understanding of this text data poses questions such as how large the corpus should be and which are the best algorithms to discriminate whether a social media post is related or not to customer experience (CX). In order to help answering these questions, first, we get hold of posts from three different platforms: Foursquare (77K) , Twitter (153K) and Facebook (2.2M). Such posts are directed to brands ranked in the ForeSee CX Index and the Forrester CX Index rankings. Second, we build a binary classifier using different algorithms to identify customer experience posts on a social platform. The accuracy of the best performing setting is 86.4% for Facebook and 91.2% for Twitter. Third, we explore the effect of increasing the number of training samples, and how a plateau is reached after 5K posts. Finally, we conduct experiments using different combinations of n-grams as features for the text mining process. As a result we observe that uni-grams and bi-grams are the best combination when we need to choose features for a classifier discriminating customer experience social media posts on Twitter and a combination of up to four-grams on Facebook.

Nowadays, microblog platforms provide a medium to share content and interact with other users. With the large-scale data generated on these platforms, the origin and reasons of users engagement in conversations has attracted the attention of the research community. In this paper, we analyze the factors that might spark conversations in Twitter, for the English and Spanish languages. Using a corpus of 2.7 million tweets, we reconstruct existing conversations, then extract several contextual and content features. Based on the features extracted, we train and evaluate several predictive models to identify tweets that will spark a conversation. Our findings show that conversations are more likely to be initiated by users with high activity level and popularity. For less popular users, the type of content generated is a more important factor. Experimental results shows that the best predictive model is able obtain an average score $F1=0.80$. We made available the dataset scripts and code used in this paper to the research community via Github.

Discovering and tracking topic shifts in news constitutes a new challenge for applications nowadays. Topics evolve,emerge and fade, making it more difficult for the journalist -or the press consumer- to decrypt the news. For instance, the current Syrian chemical crisis has been the starting point of the UN Russian initiative and also the revival of the US France alliance. A topical mapping representing how the topics evolve in time would be helpful to contextualize information. As far as we know, few topic tracking systems can provide such temporal topic connections. In this paper, we introduce a novel framework inspired from Collective Factorization for online topic discovery able to connect topics between different time-slots. The framework learns jointly the topics evolution and their time dependencies. It offers the user the ability to control, through one unique hyper-parameter, the tradeoff between the past accumulated knowledge and the current observed data. We show, on semi-synthetic datasets and on Yahoo News articles, that our method is competitive with state-of-the-art techniques while providing a simple way to monitor topics evolution (including emerging and disappearing topics).