#mlread: The Machine Learning Reading Group in Zillow AI

At Zillow AI we are focused on using AI and various forms of machine learning to help you find your home and to help Zillow succeed in our mission to be the largest, most-trusted, and vibrant home-related marketplace in the world.  So we adapt AI best practices in the unique domain of housing while also innovating in the AI space. Any reader of this blog or user of Zillow may recognize our work in home value estimation via the Zestimate and the Zillow Prize, personalization via home recommendations, 3D home walk-throughs and scene understanding, advertising technology, and the underlying big data work necessary for these data-intensive efforts to succeed.

Helping our engineers and applied scientists grow in their mastery of AI and Machine Learning methods is also a primary goal. With these efforts in mind we come together weekly in a reading group we call #mlread to present and discuss key ideas from the AI literature.  #mlread is an important contributor to helping our teams apply these important ideas.

Over the course of the last year we have delved into topics that are relevant to our work. Members of our technical staff select papers and sign up to present and lead the discussion, so naturally the papers reflect the interests and work of the teams.  Papers are drawn primarily from the scientific literature.  

First off, we care about learning accurate models, and in many cases we also care about human-interpretable models:

• Interpretable Predictions of Tree-based Ensembles via Actionable Feature Tweaking (KDD 2017) by Tolomei et al.
• Model X Knockoffs for High-Dimensional Controlled Variable Selection (2016) by Candes et al.
• Consistent individualized feature attribution for tree ensembles (2018) by Lundberg et al.

Recommender systems are a significant component of our personalization efforts:

• Deep Learning based Recommender System: A Survey and New Perspectives (2017) by Zhang et al.
• Deep Neural Networks for YouTube Recommendations (ACM 2016) by Covington, et al.
• Related Pins at Pinterest: The Evolution of a Real-World Recommender System (WWW 2017) by Liu et al.
• Local Item-Item Models for Top-N Recommendation (RecSys 2016) by Christakopoulou et al.
• Factorization Machines / Fast context-aware recommendations with factorization machines (SIGIR 2011) by Rendle et al.
• Adaptive, Personalized Diversity for Visual Discovery (RecSys 2016) by Teo et al.
• DropoutNet: Addressing Cold Start in Recommender Systems (NIPS 2017) by Volkovs et al.
• E-commerce in Your Inbox: Product Recommendations at Scale (KDD 2015) by Grbovic et al.
• SimRank: A Measure of Structural-Context Similarity (KDD 2002) by Jeh et al.

We learn from our users’ search and browsing behavior.  User clicks help us learn to predict their interests by providing implicit supervision for ranking:

• Beyond Position Bias: Examining Result Attractiveness as a Source of Presentation Bias in Clickthrough Data (WWW 2010) by Yue et al.
• Simple and scalable response prediction for display advertising (2015) by Chapelle et al.
• Position Bias Estimation for Unbiased Learning to Rank in Personal Search (WSDM 2018) by Wang et al.
• Ad Click Prediction: a View from the Trenches (KDD 2013) by McMahan et al.

Regression models are fundamental to value estimation:

• Forest-type Regression with General Losses and Robust Forest (ICML 2017) by Li et al.
• Gaussian Processes Tutorial (2017) by Schulz et al.
• Supervised neighborhoods for distributed nonparametric regression (JMLR 2016) by Bloniarz et al.

Ensemble models are also relevant to a number of the problems we care about:

• A dynamic ensemble approach to robust classification in the presence of missing data (2015) by Conroy et al.

In order to quickly learn from user preferences rather than waiting for statistical significance via A/B experiments, we apply the methods of multi-armed and contextual bandits:

• Taming the Monster: A Fast and Simple Algorithm for Contextual Bandits (ICML 2014) by Agarwal et al.
• Customer Acquisition via Display Advertising Using Multi-Armed Bandit Experiments (2015) by Schwartz et al.
• A Contextual-Bandit Approach to Personalized News Article Recommendation (2010) by Li et al.
• Making Contextual Decisions with Low Technical Debt (2016) by Agerwal et al.

Deep Learning (more generally):

• Dynamic Routing Between Capsules (NIPS 2017) by Sabour et al.
• Mastering the game of Go with deep neural networks and tree search (2016) by Silver et al.

Text is abundant in the form of real estate listing descriptions, chat logs, and structured and unstructured documents, so the methods of NLP and Text understanding are also relevant to our work:

• ABCNN: Attention-Based Convolutional Neural Network for Modeling Sentence Pairs (TACL 2016) by Yin et al.
• Gaussian LDA for Topic Models with Word Embeddings (ACL 2015) by Das et al.
• Deep Keyphrase Generation (ACL 2017) by Meng et al.

Attributing transactions is an important part of the real estate space, hence our interest in record linkage:

• A novel ensemble learning approach to unsupervised record linkage (2017) by Jurek et al.

We track the reliability of our data sources, our services, and our key metrics using the methods of anomaly detection and time-series prediction:

• Isolation-based anomaly detection (KDD 2011) by Liu et al.
• ShortFuse: Biomedical Time Series Representations in the Presence of Structured Information (PMLR 2017) by Fiterau et al.

Many of our problems also involve various resource constraints, so we care about Constrained optimization:

• Putting Data in the Driver's Seat: Optimizing Earnings for On-Demand Ride-Hailing (WSDM 2018) by Chaudhari et al.

If these ideas are compelling to you, if working on a team that builds intelligent systems incorporating machine learning methods is part of your career plan, and if you care about streamlining the process of finding a home, then consider careers in applied science, machine learning, and software engineering on zillow.com/jobs !