The majority of the experiences and interactions people have on Facebook today are made possible with AI. Well over 1 billion people enjoy unique, personalized experiences on Facebook that are powered by a wealth of AI and machine learning algorithms. AI is an incredibly fast-moving field: engineers and researchers across the company are turning the latest research breakthroughs into tools, platforms, and infrastructure that make it possible for anyone at Facebook to use AI in the experiences and products they build. This talk will look at how Facebook is conducting and applying industry-leading research to help drive advancements in AI disciplines like computer vision, language understanding, speech and video. We will also talk about building an infrastructure that anyone at Facebook can use to easily reuse algorithms in different products, scale to run thousands of simultaneous custom experiments, and give concrete examples of how employees across the company are able to leverage these platforms to build new AI products and services.

A comprehensive understanding of individual customer value is crucial to any successful customer relationship management strategy. It is also the key to building products for long-term value returns. Modeling customer lifetime value (CLTV) can be fraught with technical difficulties, however, due to both the noisy nature of user-level behavior and the potentially large customer base. Here we describe a new CLTV system that solves these problems. This was built at Groupon, a large global e-commerce company, where confronting the unique challenges of local commerce means quickly iterating on new products and the optimal inventory to appeal to a wide and diverse audience. Given current purchaser frequency we need a faster way to determine the health of individual customers, and given finite resources we need to know where to focus our energy. Our CLTV system predicts future value on an individual user basis with a random forest model which includes features that account for nearly all aspects of each customer's relationship with our platform. This feature set includes those quantifying engagement via email and our mobile app, which give us the ability to predict changes in value far more quickly than models based solely on purchase behavior. We further model different customer types, such as one-time buyers and power users, separately so as to allow for different feature weights and to enhance the interpretability of our results. Additionally, we developed an economical scoring framework wherein we re-score a user when any trigger events occur and apply a decay function otherwise, to enable frequent scoring of a large customer base with a complex model. This system is deployed, predicting the value of hundreds of millions of users on a daily cadence, and is actively being used across our products and business initiatives.

Billions of dollars of services and goods are sold through email marketing. Subject lines have a strong influence on open rates of the e-mails, as the consumers often open e-mails based on the subject. Traditionally, the e-mail-subject lines are compiled based on the best assessment of the human editors. We propose a method to help the editors by predicting subject line open rates by learning from past subject lines. The method derives different types of features from subject lines based on keywords, performance of past subject lines and syntax. Furthermore, we evaluate the contribution of individual subject-line keywords to overall open rates based on an iterative method-namely Attribution Scoring - and use this for improved predictions. A random forest based model is trained to combine these features to predict the performance. We use a dataset of more than a hundred thousand different subject lines with many billions of impressions to train and test the method. The proposed method shows significant improvement in prediction accuracy over the baselines for both new as well as already used subject lines.

In this workshop, we have several leading industry experts sharing their knowledge and experiences on how online controlled experiments are used in their applications. The individual talks are followed by a panel discussion.

Online social networks offer opportunities to analyze user behavior and social connectivity and leverage resulting in-sights for effective online advertising. We study the adoption of a paid product by members of a large and well-connected Instant Messenger (IM) network. This product is important to the business and poses unique challenges to advertising due to its low baseline adoption rate. We find that adop-tion by highly connected individuals is correlated with their social connections (friends) adopting after them. However, there is little evidence of social influence by these high degree individuals. Further, the spread of adoption remains mostly local to first-adopters and their immediate friends. We ob-serve strong evidence of peer pressure wherein future adop-tion by an individual is more likely if the product has been

Segmentation based on RFM (Recency, Frequency, and Monetary) has been used for over 50 years by direct marketers to target a subset of their customers, save mailing costs, and improve profits. RFM analysis is commonly performed using the Arthur Hughes method, which bins each of the three RFM attributes independently into five equal frequency bins. The resulting 125 cells are depicted in a tabular format or as bar graphs and analyzed by marketers, who determine the best cells (customer segments) to target. We propose an interactive visualization of RFM that helps marketers visualize and quickly identify important customer segments. Additionally, we show an integrated filtering approach that allows marketers to interactively explore the RFM segments in relation to other customer attributes, such as behavioral or demographic, to identify interesting sub-segments in the customer base. We depict these RFM visualizations on two large real-world data sets and discuss how customers have used these visualizations in practice to glean interesting insights from their data. Given, the widespread use of RFM as a critical, and many times the only, segmentation tool, we believe that the proposed intuitive and interactive visualization will provide significant business value.

Constructive learning algorithms offer an attractive approach for the incremental construction of near-minimal neural-network architectures for pattern classification. They help overcome the need for ad hoc and often inappropriate choices of network topology in algorithms that search for suitable weights in a priori fixed network architectures. Several such algorithms are proposed in the literature and shown to converge to zero classification errors (under certain assumptions) on tasks that involve learning a binary to binary mapping (i.e., classification problems involving binary-valued input attributes and two output categories). We present two constructive learning algorithms MPyramid-real and MTiling-real that extend the pyramid and tiling algorithms, respectively, for learning real to M-ary mappings (i.e., classification problems involving real-valued input attributes and multiple output classes). We prove the convergence of these algorithms and empirically demonstrate their applicability to practical pattern classification problems. Additionally, we show how the incorporation of a local pruning step can eliminate several redundant neurons from MTiling-real networks.

Bioinformatics and related applications call for efficient algorithms for knowledge-intensive learning and data-driven knowledge refinement. Knowledge based artificial neural networks offer an attractive approach to extending or modifying incomplete knowledge bases or domain theories. We present results of experiments with several such algorithms for data-driven knowledge discovery and theory refinement in some simple bioinformatics applications. Results of experiments on the ribosome binding site and promoter site identification problems indicate that the performance of KBDistAl and Tiling-Pyramid algorithms compares quite favorably with those of substantially more computationally demanding techniques.

Multi-layer networks of threshold logic units (TLU) offer an attractive framework for the design of pattern classification systems. A new constructive neural network learning algorithm (DistAl) based on inter-pattern distance is introduced. DistAl constructs a single hidden layer of hyperspherical threshold neurons. Each neuron is designed to determine a cluster of training patterns belonging to the same class. The weights and thresholds of the hidden neurons are determined directly by comparing the inter-pattern distances of the training patterns. This offers a significant advantage over other constructive learning algorithms that use an iterative (and often time consuming) weight modification strategy to train individual neurons. The individual clusters (represented by the hidden neurons) are combined by a single output layer of threshold neurons. The speed of DistAl makes it a good candidate for datamining and knowledge acquisition from large datasets. The paper presents results of experiments using several artificial and real-world datasets. The results demonstrate that DistAl compares favorably with other learning algorithms for pattern classification.