MICES 2021

Dreaming Search

Ángel Maldonado
Empathy.co

Download slides (PDF)

Watch talk on Youtube.

In this talk Angel proposes to open the way we think about users, metrics and objectives to the consideration of affects, emotions and feelings.

Can Commerce Search be developed from theory as well as facts?

The talk entertains a wider view of what Commerce Search can become when theory drives innovation, exemplifying an ideal where the roles of consumers and retailers reverse, a view of Search that serves each household and one where retailers are invited to read and offer under each household terms.

The hope is that the audience comes back to their day to day search battles inspired to freely and genuinely try new ideas, advance new hypothesis while contrasting their attributable value with an open mindset, one that respects and elevates the individual beyond the constrains of data.

Angel Maldonado has spent the last 22 years passionately developing and executing Commerce Search and Discovery solutions. Having studied Computer Information Systems at Liverpool University, Angel started his career working for Autonomy, where he helped clients on pioneering enterprise search projects for seven years before founding Empathy.co.

As the founder and CEO of Empathy.co, Angel drives forward the product vision, innovation and ethos of creating products whose mission is to evoke positive feelings. Empathy's search components and micro-services support brands like Inditex, Carrefour or Kroger in the US.

A keen water sport enthusiast, Angel plays the guitar and enjoys poetry and meditation. He lives in London.

Forget Facets, welcome Refinements (TM)

Lucian Precup & Vincent Bosc
a// & Adelean

Download slides (PDF)

Watch talk on Youtube.

End of 2019 at Haystack Europe (https://haystackconf.com/europe2019/relevant-facets/) we presented tips and tricks on how to get the most out of facets and filters. With the advent of new use cases and technologies (like, among others, voice search, assistants and chat bots - https://2019.berlinbuzzwords.de/19/session/integrate-your-search-engine-voice-assistant.html) the good old facets and filters take other and lots of different forms of expression. Incidentally, they are no longer exclusively implemented with the help of aggregations or similar techniques. Historically, facets had two purposes: give an overview of the data and help filtering towards the right result. The first purpose led to data analysis engines while the second is more important to e-commerce search engines. Coupled to another role of modern e-commerce search engines which is "provide inspiration", we get to a new concept that we call Refinements, and which will probably and progressively replace what we have known as Facets. This talk is an overview and introduction to the next generation of Facets.

Lucian is the CTO of a// - your collaborative search engine - and founded Adelean, which offers, since 2010, consulting services and expertise on search engines.

Vincent is a Senior Search Engines Developer at Adelean and integrates Adelean’s technology and know-how in large-scale e-commerce search engines.

Reinforcement learning in search

Roman Grebennikov
Findify

Download slides (PDF)

Watch talk on Youtube.

As software engineers, we may see a ranking in search results as a solved problem: BM25 and TF/IDF are decades old and proven approaches to it. But in areas like eCommerce multiple stakeholders can have dramatically different objectives to tune search for: apart from NDCG, can you please increase click-through-rate but also make customers prefer high-margin products, and don’t forget about inventory size and bounce rate! And optimizing for multiple objectives at once can quickly become much harder than originally planned.

On one side there is a family of classical learn-to-rank methods using historical data to improve relevancy, on another side - a lot of manual a/b tests exploring how customers and business metrics can react to changes in ranking. Is it possible to bridge these approaches together, so ranking would not only use clicks that happened months ago to improve some abstract metrics like NDCG, but also explore how ranking changes affect more down-to-business things like revenue and CTR? Can we tune a learn-to-rank model in a way so it can become a learn-to-sell?

In this talk, we will discuss an online reinforcement learning approach to LTR when static click data filled with biases is not enough. We will start with reinforcement learning basics like multi-arm bandits and continue on how they can be mixed with search-specific click models, query-item contexts, and overcome a position bias problem. We'll also discuss some advanced topics like CascadeLinUCB, PDGD and CPR ranking algorithms.

We’ll also talk about the challenges faced in building an online Learn-to-Rank system doing hundreds of automated exploratory experiments on real customers, and using the gained knowledge to directly optimize for business metrics.

Roman Grebennikov is a passionate software developer with hands-on experience in machine learning software development and JVM. During last years he has focused on the delivery of functional programming principles and practices to real-world data analysis and information retrieval projects.

Better Search through Query Understanding - Panel

Andrea Schütt (OTTO), Aritra Mandal (eBay), Dominic Bestler (Digitec Galaxus), hosted by Daniel Tunkelang (Independent Consultant)

Andrea Schütt - Download slides (PDF)

Aritra Mandal - Download slides (PDF)

Dominic Bestler - Download slides (PDF)

Watch panel on Youtube.

Much of the work on ecommerce search engines focuses on ranking, using increasingly sophisticated machine learning systems. Ranking can improve relevance and conversion. But focusing on ranking is premature if the search engine does not understand the query.

Query understanding focuses on the place where searchers start their journey: the query. Query understanding is concerned with the searcher’s process of expressing intent and the search engine’s process of determining that intent.

Join host Daniel Tunkelang and search practitioners Andrea Schütt, Aritra Mandal, and Dominic Bestler to learn about techniques for query understanding, as well as how to use those techniques to deliver better outcomes for searchers and businesses.

Dominic is a search product manager at Digitec Galaxus. He loves solving customer problems with a focus on the overall "conversation with the user". He has a degree in computer science from the Zurich University of Applied Sciences.

Aritra is an applied researcher on the search team at eBay. His work focuses on improving eBay’s understanding of search queries. He has degrees in computer science from Birla Institute of Technology and Indiana University–Purdue University Indianapolis.

Andrea is a Data Scientist on OTTO’s search team. Currently she is working on bringing OTTO’s first learning-to-rank model into production. She has a degree in electrical engineering with a focus on automation from Technical University of Munich.

Daniel is an independent consultant specializing in search, discovery, machine learning / AI, and data science -- with a particular passion for query understanding. He was a founding employee of Endeca, a local search tech lead at Google, and a director of data science and search at LinkedIn. His clients have included Apple, eBay, Coupang, Etsy, Flipkart, Gartner, Pinterest, Salesforce, Yelp, and Zoom; as well as some of the largest traditional retailers. He has degrees in computer science and math from MIT and CMU.

Querqy User Group Meeting

Hosted by René Kriegler (Freelancer), Paul M. Bartusch (Freelancer) & Johannes Peter (MediaMarktSaturn)

Dominic Bestler - Download slides (PDF)

Johannes Peter - Download slides #1 (PDF)

Johannes Peter - Download slides #2 (PDF)

Watch user group meeting on Youtube.

Querqy is a query rewriting library for Solr and Elasticsearch. While it can apply all sorts of query manipulations, its main use case lies in e-commerce search, especially in managing and applying business rules to specific search queries. We’ve recently seen a growing interest in Querqy. We receive questions about how synonyms work in Querqy, what’s the difference to using Solr’s Edismax query parser, how can I track if a rewriter was applied to a query in Elasticsearch, and many more.

We think that this is a good time to set up a first Querqy User Group Meeting to learn and discuss about:

What's new in Querqy?

Share your Querqy story: How has it helped you? What would have made your life easier? You're generating thousands of rewriting rules? You've written your own rewriter? - We'd love to hear your story!

What features would you like to see in Querqy? / Ideas for Querqy's roadmap

The broader perspective: SMUI and other components that help us implement e-commerce search more easily

'Ask me anything' - ask the committers, from beginner's questions to that tricky question that they probably cannot answer!

We want to keep the discussion as open and as interactive as possible, but there will also be room for short lightning talks. Please get in touch with us by email via 'info at mices.co' prior to the event, if you want to give a short talk.

The need for an open web search in Europe - The approach of the Open Search Foundation and its implications for E-Commerce-Companies

Alexander Decker
Open Search Foundation

Download the slides (PDF)

Watch talk on Youtube.

In this talk the Open Search initiative (Open Search Foundation, OSF) and its vision will be introduced as a basis. Since the whole project is in progress and there is no such open search index so far, it will be explained how - coming from the strategic framework - services can use such an open search index for their businesses, based on examples. Particularly, e-commerce providers could come in, because they could set up concrete services according to their gusto and needs on the web index.

Since we are in the process of developing ideas on how to design the index as a whole, it is the idea of this talk - based on the initial input given - to explore the views and the requirements of E-commerce providers for such an index in an interactive form in the talk as well.

Alexander Decker is Professor of Consumer Goods Marketing and Digital Media and Head of the Marketing / Sales / Media Master's program at Ingolstadt University of Applied Sciences - THI Business School. There he is also Vice Dean Marketing & Alumni. In addition, he is founder and managing director of the consultancy Seward's Folly. Since September 2018, he is also a founding member and Chief Marketing Officer of the Open Search Foundation.

Previously, Alexander spent more than 15 years as a management consultant at Vectia and held senior marketing positions at Premiere Fernsehen (now Sky Deutschland) and Nestlé Germany. There he was responsible for the Nestlé Marketplace, the first social commerce platform of a food manufacturer. As a social media and digital management expert, he is a speaker at specialist events and the author of various specialist publications, such as the practical and textbook "The Social Media Cycle".

101 hints to improving the customer satisfaction on search engines in the retail industry

Marion Hemery (Carrefour France) & Lucian Precup (a// & Adelean)

Download slides (PDF)

Watch talk on Youtube.

User experience on e-commerce websites is influenced by a lot of factors: the speed, the relevance and the workflow of the search engine are among them. But the customer satisfaction in general and especially in the retail industry is difficult to measure: levels to improve are multiple and the constraint of multiple physical stores backed by the online search engine introduces additional challenges. In this talk we give our hints on first, how to effectively measure the customer satisfaction and second, how to improve it step by step. We give real world examples from our experience in improving our customer satisfaction KPIs in the context of Carrefour, the largest retail group and online grocery store in France.

Marion is the Search Manager at Carrefour France and works notably on the e-commerce search engine behind carrefour.fr. She is also the Product Owner of the Search and Merchandising teams.

Lucian is the CTO of a// - your collaborative search engine - and founded Adelean, which offers, since 2010, consulting services and expertise on search engines.

“An ounce of prevention is worth a pound of cure”: establishing a gold standard-based evaluation in customer projects

Bertram Sändig & Cornelia Werk
Neofonie

Download slides (PDF)

Watch talk on Youtube.

Building an exceptional search or even just optimising standard search solutions up to the point where customers can actually benefit from them, comes with a certain cost. Experts have to work long hours to adjust the search’s performance to the various needs of customers. This process can become very cumbersome because any specific optimisation to one part of the system can turn out to have unforeseen, detrimental effects on other parts. This can be devastating, especially when much work has been put into a specific optimisation or if the damage persists for a long time before it is spotted. If only there was a tool to prevent that from happening!

Of course, there is. This situation can be avoided by a continuous evaluation process. Needless to say, evaluating a search solution and its iterative development, is far from new. However, in our customer projects we have made it a rule to build an evaluation based on a gold standard corpus and in our talk we would like to share our practical experience in establishing such a gold standard corpus together with our customers. Our topics will reach from a) the different perks of having a gold standard-based evaluation through b) finding a suitable strategy to communicate its need to our customers, to c) the actual process of building a corpus. By means of two exemplary customer projects we will present how different and specific to the customer this process can be, but also what a general framework for gold standard corpus creation could look like.

Bertram Sändig is a machine learning and search engineer for the research department of Neofonie GmbH. He is involved in multiple projects including sentiment analysis, toxicity detection, intent recognition and named entity recognition. He also consults enterprises in their efforts to optimize lucene based search systems, for instance through the integration of neural NLU algorithms.

Cornelia Werk is a consultant for NLP, Data Science & Search at Neofonie where she is also part of the research department. Standing at the intersection between customers and developers she is responsible for finding individual solutions based on Artificial Intelligence for Neofonie's customers. She especially focusses on search optimization, but has also gained useful experience as Data Analyst in areas like intelligent data analysis, AI and quality management.

In-session personalization: ranking, type-ahead and query scoping with deep learning - hands-on workshop

Patrick John Chia & Jacopo Tagliabue
Coveo

Download slides (PDF)

Watch workshop on Youtube.

According to industry experts, personalization will drive a 15% profit increase this year; we don't want search results to be just relevant, we want them to be tailored to each shopper's intention. However, personalization is easier said than done, as most shops have high bounce rates and few returning customers. Can machine learning be effective when our models need to work as early as possible in the customer journey? In this hands-on workshop, we show how to build effective personalization models for query scoping, recommendation and type-ahead, leveraging cutting-edge deep learning techniques and a real-world dataset.

The workshop is in two parts: in the first, we present the business problem, explain why in-session inference is so important for digital shops, and frame the problem of personalization in a general way across use cases. In the second part, we dive deep into working code, showing how to represent in-session intent with deep models, and how to inject personalization into several downstream use cases. All the models presented come from our research and product roadmap, and both code and data will be available under an open source license at the end of the workshop.

PREREQUISITES

The first part can be appealing to a broad audience, as the problem framing should be interesting for technical (IR/NLP practitioners) and product folks; as such, there are no real prerequisites if not familiarity with eCommerce as an industry. The second part will be more technical, as we will run actual code, and share tips and tricks on models and implementation at scale: a working knowledge of Python and good grasp of deep architectures will make the workshop easier to follow. Please note that code and data will be released, so participants will be able to review the material at their own pace after the workshop.

Jacopo Tagliabue was co-founder and CTO of Tooso, an A.I. company in San Francisco acquired by Coveo in 2019. Jacopo is currently the Lead A.I. Scientist at Coveo. When not busy building A.I. products, he is exploring research topics at the intersection of language, reasoning and learning, with several publications at major conferences (e.g. WWW, SIGIR, RecSys, NAACL). In previous lives, he managed to get a Ph.D., do scienc-y things for a pro basketball team, and simulate a pre-Columbian civilization.

Patrick John Chia is a machine learning developer at Coveo. Prior to this, he completed his Master's degree at Imperial College London and spent a year at MIT – both experiences greatly influenced his opinions on AI. He is on the organizing committee for the 2021 SIGIR Data Challenge and he is a speaker on topics at the intersection of Machine Learning and eCommerce. His latest interests lie in developing AI that has the ability to learn like infants and applying it to creating solutions at Coveo.