Eunyee Koh

Identifying Types of Misalignments between Promotion Emails and Landing Pages

A key requirement of successful online marketing is to maintain the quality of hyperlinks. However, it is not uncommon for users to get confused or disappointed by a wide range of misalignments between links and their landing pages. This paper presents an online survey that identifies types of such misalignments perceived by recipients.

Personalizable and Interactive Sequence Recommender System

Sequence recommender systems assist people in making decisions, such as which product to purchase and what places to visit on vacation. Despite their ubiquity, most sequence recommender systems are black boxes and do not offer justifications for their recommendations or provide user controls for steering the algorithm. In this paper, we design and develop an interactive sequence recommender system (SeRIES) prototype that uses visualizations to explain and justify the recommendations and provides controls so that users may personalize the recommendations. We conducted a user study comparing SeRIES to a black-box system with 12 participants using real visitor trajectory data in Melbourne and show that SeRIES users are more informed about how the recommendations are generated, more confident in following the recommendations, and more engaged in the decision process.

Detecting Link and Landing Page Misalignment in Marketing Emails

Links and their landing pages in the World Wide Web are oftentimes flawed or irrelevant. We created a data set of 4266 links within 160 marketing emails whose relevance with landing pages have been evaluated by crowd workers. We present a study of common misalignments and propose methods for detecting these misalignments. An F-score of 0.63 can be achieved by a neural network for cases where the misaligned label requires the majority out of 5 crowd worker votes.

Predictive Analysis by Leveraging Temporal User Behavior and User Embeddings

The rapid growth of mobile devices has resulted in the generation of a large number of user behavior logs that contain latent intentions and user interests. However, exploiting such data in real-world applications is still difficult for service providers due to the complexities of user behavior over a sheer number of possible actions that can vary according to time. In this work, a time-aware RNN model, TRNN, is proposed for predictive analysis from user behavior data. First, our approach predicts next user action more accurately than the baselines including the n-gram models as well as two recently introduced time-aware RNN approaches. Second, we use TRNN to learn user embeddings from sequences of user actions and show that overall the TRNN embeddings outperform conventional RNN embeddings. Similar to how word embeddings benefit a wide range of task in natural language processing, the learned user embeddings are general and could be used in a variety of tasks in the digital marketing area. This claim is supported empirically by evaluating their utility in user conversion prediction, and preferred application prediction. According to the evaluation results, TRNN embeddings perform better than the baselines including Bag of Words (BoW), TFIDF and Doc2Vec. We believe that TRNN embeddings provide an effective representation for solving practical tasks such as recommendation, user segmentation and predictive analysis of business metrics.

Perceptual Similarity Ranking of Temporal Heatmaps Using Convolutional Neural Networks

Similarity ranking is central to various analytic tasks. While current approaches work well on low-dimensional datasets, it becomes difficult to define similarity for more complex data types, like event sequences with multidimensional attributes. Often, its definition needs to be manually tuned according to the target domain or dataset. Visualizations are similarly manually tuned by analysts and can contain important clues about relevant features. In this paper, we propose using computer vision techniques on visualizations as a means for similarity ranking. We visualize sequential datasets as temporal heatmaps and show through user studies with 132 participants that humans agree in ranking results to a query based on perceptual similarity. We design and implement Heat2Vec, a convolutional neural network (CNN) to learn latent representations from heatmaps using color, opacity, and position. We evaluate our method against 11 baselines using a wide range of techniques and show that Heat2Vec provides rankings that are most consistently in line with human-annotated similarity ranking.

Higher-order Network Representation Learning.

This paper describes a general framework for learning Higher-Order Network Embeddings (HONE) from graph data based on network motifs. The HONE framework is highly expressive and flexible with many interchangeable components. The experimental results demonstrate the effectiveness of learning higher-order network representations. In all cases, HONE outperforms recent embedding methods that are unable to capture higher-order structures with a mean relative gain in AUC of 19% (and up to 75% gain) across a wide variety of networks and embedding methods.

Continuous-Time Dynamic Network Embeddings.

Networks evolve continuously over time with the addition, deletion, and changing of links and nodes. Although many networks contain this type of temporal information, the majority of research in network representation learning has focused on static snapshots of the graph and has largely ignored the temporal dynamics of the network. In this work, we describe a general framework for incorporating temporal information into network embedding methods. The framework gives rise to methods for learning time-respecting embeddings from continuous-time dynamic networks. Overall, the experiments demonstrate the effectiveness of the proposed framework and dynamic network embedding approach as it achieves an average gain of 11.9% across all methods and graphs. The results indicate that modeling temporal dependencies in graphs is important for learning appropriate and meaningful network representations.

WimNet: Vision Search for Web Logs

With the growing popularity of mobile devices, user web logs are more heterogeneous than ever, across an increased number of devices and websites. As a result, identifying users with similar usage patterns within these large sets of web logs is increasingly challenging and critical for personalization and user experience in many areas, from recommender systems to digital marketing. In this work, we explore the use of visual search for top-k user retrieval based on similar user behavior. We introduce a convolution neural network (WimNet) that learns latent representation from a set of web logs represented as images. Specifically, it contains two convolution layers take row- and column-wise convolutions to capture user behavior across multiple devices and websites and learns latent representation and reconstructs a transition matrix between user activities of given web logs. To evaluate our method, we conduct conventional top-k retrieval task on the simulated dataset, and the preliminary analysis results suggest that our method produces more accurate and robust results regardless of the complexity of query log.

Topical Sequence Profiling

This paper introduces the problem of topical sequence profiling. Given a sequence of text collections such as the annual proceedings of a conference, the topical sequence profile is the most diverse explicit topic embedding for that text collection sequence that is both representative and minimal. Topic embeddings represent a text collection sequence as numerical topic vectors by storing the relevance of each text collection for each topic. Topic embeddings are called explicit if human readable labels are provided for the topics. A topic embedding is representative for a sequence, if for each text collection the percentage of documents that address at least one of the topics exceeds a predefined threshold. If no topic can be removed from the embedding without loosing representativeness, the embedding is minimal. From the set of all minimal representative embeddings, the one with the highest mean topic variance is sought and termed as the topical sequence profile. Topical sequence profiling can be used to highlight significant topical developments, such as raise, decline, or oscillation. The computation of topical sequence profiles is made up of two steps, topic acquisition and topic selection. In the first step, the sequences text collections are mined for representative candidate topics. As a source for semantically meaningful topic labels, we propose the use of Wikipedia article titles, whereas the respective articles are used to build a classifier for the assignment of topics to documents. Within the second step the subset of candidate topics that constitutes the topical sequence profile is determined, for which we present an efficient greedy selection strategy. We demonstrate the potential of topical sequence profiling as an effective data science technology with a case study on a sequence of conference proceedings.