Unsupervised Learning of Parsimonious General-Purpose Embeddings for User and Location Modeling

Abstract

Many social network applications depend on robust representations of spatio-temporal data. In this work, we present an embedding model based on feed-forward neural networks which transforms social media check-ins into dense feature vectors encoding geographic, temporal, and functional aspects for modelling places, neighborhoods, and users. We employ the embedding model in a variety of applications including location recommendation, urban functional zone study, and crime prediction. For location recommendation, we propose a Spatio-Temporal Embedding Similarity algorithm (STES) based on the embedding model.In a range of experiments on real life data collected from Foursquare, we demonstrate our model's effectiveness at characterizing places and people and its applicability in aforementioned problem domains. Finally, we select eight major cities around the globe and verify the robustness and generality of our model by porting pre-trained models from one city to another, thereby alleviating the need for costly local training.In this work, we instead represent places by means of embedding vectors in a semantic space. Aiming to annotate places in terms of temporal, geographic, and functional aspects, we extract the time, location, and venue function from check-in records and train our model in the context of check-in sequences which originate from an individual user or neighborhood.In comparison with the traditional discrete method, our approach describes places in a continuous manner and preserves more information about people's real behavior as well as places' day-to-day usage patterns. For instance, in the case of label annotation, three food related places which serve Chinese breakfast, pizza, and sushi, respectively, may all be labeled as Restaurant but their features in food type, active hours, and location may vary dramatically. In the course of this article, we will show how our embedding model represents such within-class variance in a natural way. As embedding vectors are learnt from people's real-time check-ins, we also leverage them in user representation to reflect people's activity patterns and interests.The embedding model is an accurate descriptor of places and users in terms of geographic and functional affinity, activity preferences, and daily schedules. As a consequence, it can be applied in a wide range of settings. In this paper, we consider three practical applications: location recommendation, urban functional zone study, and crime prediction. Our empirical investigation is driven by five research questions: • RQ1. How well does the embedding model differentiate locations and users along temporal, geographic, and functional aspects? • RQ2. How does our location recommendation algorithm STES compare to state-of-the-art methods? • RQ3. How to define and visualize urban functional zones using the proposed model? • RQ4. How well can the model predict typical urban characteristics? • RQ5. With what generalization error can an embedding model trained in one city be transferred to other cities?By answer...

Representing Fine-Grained Co-Occurrences for Behavior-Based Fraud Detection in Online Payment Services

“…By applying the network embedding method to the label-aware network, we get the mapping relationship ϕ I (•). Further, we establish the individual-level models of probability in view of ϕ I (•) [42].…”

Grid-Based Crime Prediction Using Geographical Features

“…However, these methods typically only consider time or space factors separately, and are thus very sensitive to time and space selection, which can result in prediction results that do not outperform a simple linear regression [4]. Many subsequent studies concurrently consider time and space factors [5,6], and gradually explore the integration of additional features, including type of crime [7], footprint and GDP [8], Twitter comments [9][10][11], and explain correlations between features [12]. Some studies [13,14] noted the impact of geographical features on crime, but few studies have attempted to apply geographical features to crime prediction.…”

Utilizing FastText for Venue Recommendation

“…The embedding techniques, Word2Vec [18] and Doc2Vec [10] are frequently used for making recommendations, e.g. [21], [15], [35], [30]. Although these techniques are powerful at learning the semantic relations among venues and users, a newer method named FastText [2] can be more performant to represent the venues.…”

“…While the temporality and spatiality features are frequently used in venue recommendation literature, e.g. [34], [8], [30]; sequentiality of check-ins gained limited attention, e.g. [33], [35].…”