Concept to Code: Learning Distributed Representation of Heterogeneous sources for Recommendation (Deep Learning Beginner/Intermediate)
by Omprakash Sonie (Flipkart, India), Surender Kumar (Flipkart, India) and Sudeshna Sarkar (IIT Kharagpur, India)
Deep Learning attempts to learn multiple levels of representations and abstractions from data. Some of the major Deep Learning techniques used in recommender systems are: Embedding methods for embedding different products based on content and transactions, feedforward multi-layer networks and auto-encoders for collaborative filtering, Convolutional Neural Network (CNN) for extracting features from content such as images, sound and text; and session-based recommendation with Recurrent Neural Networks (RNN).
The aim of this tutorial is to provide a conceptual understanding of learning distributed representation techniques by using various data sources including items, users, product images, review texts and ratings for a recommender system. They will be covered with sufficient mathematical background along with actual code.
A case study covering embeddings, RNN, and CNN will be presented along with the code walkthrough and what is going on behind the scenes.
The tutorial is for beginner and intermediate participants who have run a few projects in Python/Jupyter notebooks and would like to understand the high level math behind the techniques.
Material
Date
Tuesday, Oct 2, 2018, 09:00-10:30
Location
Parq B/C
Modularizing Deep Neural Network-Inspired Recommendation Algorithms
by Longqi Yang (Cornell Tech, Cornell University, USA), Hongyi Wen (Cornell Tech, Cornell University, USA) and Eugene Bagdasaryan (Cornell Tech, Cornell University, USA)
Deep Neural Network (DNN)-inspired recommendation algorithms usually leverage numerous heterogeneous sub-models to ingest a wide range of information, such as diverse user feedback signals and auxiliary, contextual and cross-platform, traces. These complexities pose significant challenges to the common recommender development practice which treats each algorithm as monolithic. As a result, in order to experiment with a new method for even a small part of an algorithm, or customize an algorithm for other scenarios, researchers and engineers need to re-implement the model from scratch or extensively patch existing code. For practitioners, this introduces significant barriers for applying state-of-the-art solutions to existing services. To tackle these challenges, we introduce and share hands-on experience with OpenRec, an open-source tool that modularizes recommendation algorithms, so that changes, additions and deletions of components can be explored independently, and development and testing can be readily achieved via plug-and-play manner.
In this tutorial, we will present:
- Recent developments of DNN models for recommendations.
- Why and how to modularize DNN-inspired recommenders.
- Quick introduction to Tensorflow.
- The OpenRec architecture and Python interface.
- Hands-on sessions in developing and adapting state-of-the-art DNN recommenders.
This tutorial targets intermediate and advanced audiences, and requires basic knowledge of deep neural networks (such as MLP, CNN and RNN). We will focus on recommendation-specific developments and applications of DNN. Prior experience with Tensorflow is not required.
Material
Date
Tuesday, Oct 2, 2018, 11:00-12:30
Location
Parq B/C
Emotions and Personality in Recommender Systems
by Marko Tkalčič (Free University of Bozen-Bolzano, Italy)
Psychological aspects of item consumption have been under-explored in the RecSys community. For example, a movie usually contains a roller-coaster of emotions, but the user preference for such a complex experience in recommender systems is usually expressed as one numerical score. Hence, a lot of information is hidden in the psychological characteristics of the user, the emotional characteristics of the item and in the context where the interaction occurs.
This tutorial will consist of 2 parts:
- a lecture, providing a good-enough understanding of the psychological background and
- a hands-on session showing how to build an interface for preference elicitation with emotions and personality acquisition.
The tutorial targets early-stage researchers in recommender systems. They should have a basic understanding of recommender systems. For the hands-on part, they should have a basic command of HTML and JavaScript.
Material
Date
Tuesday, Oct 2, 2018, 16:00-17:30
Location
Parq A
Multimedia Recommender Systems (cancelled)
by Yashar Deldjoo (Politecnico di Milano, Italy), Markus Schedl (Johannes Kepler University, Austria), Balázs Hidasi (Gravity R&D, Hungary) and Peter Knees (TU Vienna, Austria)
This tutorial introduces multimedia recommender systems (MMRS), i.e., recommender systems that leverage multimedia content to recommend different media types. With a particular focus on content-based MMRS and hybrids of collaborative filtering and content-based filtering, we address the target domains of movie and music recommendation. Discussing the topics of traditional content-based multimedia feature extraction (text, audio, visual), deep learned features, recommendation approaches for multimedia items, end-to-end deep models, and evaluation and datasets, we give the audience a profound introduction to serve as starting point to advance the field of recommender systems. We also point to the Grand Challenges in the domain of MMRS.
Given the presenters’ background, we consider both academic and industrial points of view. Particular prerequisite knowledge or skills are not required from the audience, except for a basic understanding of the main concepts in recommender systems. Motivated by previous experiences of a very active RecSys audience, we conclude the tutorial with a hopefully vivid discussion between the audience and the presenters.
Sequence-aware Recommendation
by Massimo Quadrana (Pandora Media, Italy) and Paolo Cremonesi (Politecnico di Milano, Italy)
Most works in the field of recommender systems are focused on the matrix completion problem, where for each user-item-pair only one interaction (e.g., a rating) is considered. In many application domains, however, multiple user-item interactions of different types can be recorded over time, and this information can be used to build richer individual user models and to discover additional behavioral patterns that can be leveraged in the recommendation process.
In this tutorial, we review existing works that consider information from such sequentially-ordered user-item interaction logs in the recommendation process. Based on this review, we describe a categorization of the corresponding recommendation tasks and goals, present the algorithmic solutions in depth, discuss methodological approaches when benchmarking what we call sequence-aware recommender systems, and outline open challenges in the area.
The tutorial will based on a recent survey paper published by the presenters in ACM Computing Surveys and will include a hands-on session. More details will be made available.
Material
Date
Tuesday, Oct 2, 2018, 14:00-17:30
Location
Parq B/C
Mixed Methods for Evaluating User Satisfaction
by Jean Garcia-Gathright (Spotify, USA), Christine Hosey (Spotify, USA), Brian St. Thomas (Spotify, USA), Ben Carterette (Spotify, USA), and Fernando Diaz (Microsoft Research, Canada)
Evaluation is a fundamental aspect of designing and experimenting on recommender systems. Evaluation typically takes one of three forms: (1) smaller lab studies with real users; (2) batch tests with offline collections, judgements, and measures; (3) large-scale controlled experiments (e.g. A/B tests) looking at implicit feedback. But it is rare for the first to inform and influence the latter two; in particular, implicit feedback metrics often have to be continuously revised and updated as assumptions are found to be poorly supported.
Mixed methods research provides an opportunity to develop robust implicit metrics by combining strengths of both qualitative and quantitative approaches and exploring a research area from multiple perspectives. In this tutorial, we will show how qualitative research on user behavior provides insight on the relationship between implicit signals and satisfaction. These insights can inform and augment quantitative modeling and analysis for online and offline metrics and evaluation.
Attendees will learn about:
- how to select implicit signals in user log data important for predicting user satisfaction,
- how to carry out qualitative research to generate hypotheses about user satisfaction,
- how to use these hypotheses to guide and understand quantitative analyses,
- how the resulting insights can be used to measure, evaluate, and optimize performance of recommender systems, and
- how to interpret and understand statistical tests of hypotheses about relative differences in user satisfaction.
Material
Date
Tuesday, Oct 2, 2018, 09:00-15:30
Location
Parq A
