MICES 2018

Registration

Introduction

René Kriegler
Freelancer

(Slides)

Getting started with search tuning and search relevance. Overcoming the Google effect – ‘If Google can get it right, why can’t we?’

Karen Renshaw
Grainger Global Online

Why: Search Tuning and Relevance can seem daunting. Search gets a lot of noise from around the organization. Everyone has an opinion; everyone thinks it’s easy but it's a long term investment.

Overview:

• Getting Started
• Improving search can seem daunting, it’s clear the experience can be improved but is the technology, the content or the experience
• This is not a technology problem, needs a multi-disclipined team to work together
• Set objectives
• Understand the issue: what are you currently, where do you want to be
• Content is important too! Understand how
• Prioritise changes
• Develop a test matrix and approach
• Managing expectations around the organization
• Communicating the plan is important, especially for HIPPO’s and LIPPO’s Search is an ongoing task not a one off piece of work
• Consider the holistic search experience
• Relevancy changes should be understood in the context of the customer journey
• Presentation of the search results and relevancy need to be considered hand in hand
• Measurement
• Understanding results and ongoing plans

Karen has over ten years experience of driving on site search improvement. She is currently Head of Search and Content for Grainger Global Online, where she is responsible for defining and implementing the on site search experience. Previously, Karen was Head of Search for RS Components where she set up a Global Search team and developed a search migration programme.

(Slides)

Break

How building our own e-commerce search solution changed our approach to search quality

Christine Bellstedt & Jens Kürsten
OTTO

Replacing a feature-rich proprietary search system that has been tuned for years comes with many challenges. One key aspect is relevance tuning for business performance. In order to improve our business performance indicators, we faced a number of methodological questions:

• How can we assess the performance of the new solution with respect to search quality in an off-line setting?
• How can we continuously challenge and optimize the current on-line system (configuration)?

Our revised approach to search quality evaluation and optimization is two-fold. We created a light-weight framework for fast off-line experimentation in order to create insights and validate hypotheses. The results allow us to select candidate features or system configurations for future on-site tests in a champion vs. challenger setting.

In this talk, we will share the building blocks of our methodology: how we derived query sets, judgement lists, and search quality metrics, how we implemented an off-line experiment cycle as well as all the hard and fun lessons we learned along the way.

Christine is a specialist in ecommerce search and has been the product owner for search and navigation at OTTO GmbH for four years now. Her impulsion is to find the best solution for search at otto.de by bringing together the user, the business, the market and the product perspective – always embracing change. In the early days of her career she founded a start-up, showing her talent to create an innovative and successful work environment. Christine graduated in Media Management, writing her diploma thesis on how organisational inertia in media companies prevents them from releasing radical innovations.

Jens is a search engine enthusiast working at OTTO GmbH, always pushing hard to find and implement the most relevant metric(s) to quantify search relevance. He graduated from Chemnitz University in Germany while researching methods for organizing and searching audio-visual archives of local TV stations. He is also interested in the inter-connected domains of natural language processing, computer linguistics, and machine learning.

(Slides)

Live search metrics for smart query optimization

Pavel Penchev
SearchHub.io

Self-learning search systems are a hot topic but they can only be as good as the data they get. Optimizing search sits at the core of what we do at SearchHub thus high quality data is a must. We aim to make search a KPI driven experience but finding a tool that can track important metrics such as MRR (mean reciprocal rank) proved to be a challenge. That's why we developed SearchHub Collector - a generic SDK providing search-specific live metrics that can be fed in real time into a self-optimizing system. We are publishing the code as open source. In this talk we want to share the rationale behind it and some of the learnings we gathered along the way.

Pavel is a long time search professional with more than 10 years experience at Fredhopper, a leading European search vendor, and works now at CXP Commerce Experts, a progressive search company from Pforzheim. Having seen the search world through different roles and perspectives, he love digging into the little details that combine to make a great search experience. Pavel has a background in software engineering, cloud systems and product management. He comes from Sofia, Bulgaria.

(Slides)

Break

Measuring and optimising search quality – a practical guide

Andreas Brückner
ATTRAQT Fredhopper

Surprisingly few large online retailers have clearly structured search optimisation processes in place. It is not uncommon to see Search Conversion Rate over time being used as the only indicator for success or failure in on-site search. With search optimisation often only a minor element in already packed job descriptions this should not be entirely surprising. In this talk we will review how successful online retailers approach search optimisation start-to-end. We start by reviewing how online retailers embed search optimisation efforts in their organisational structure. From there we will move on to search quality measurement as the foundation of every successful optimisation project. We will discuss commonly used search quality KPIs and optimisation approaches, with a focus on those that can be implemented easily into existing structures.

Andreas is the subject matter expert for search and merchandising at ATTRAQT Fredhopper, the on-site search technology provider trusted by many of Europe’s largest online shops. Andreas has provided thought leadership and best practice advice to industry heavy-weights such as ASOS, Otto.de, Bonprix and House of Fraser.

(Slides)

Approaches to implement semantic search

Johannes Peter
MediaMarktSaturn Retail Group

Semantic search provides various opportunities to take search, analytics and recommendation of webshops to the next level. The main goal of a semantic processing of user's inputs is to achieve a more detailed understanding of their actual demands. From a technical perspective, the main challenge in implementing semantic search is to identify entities in text. In the context of e-commerce, entities can be described as the technical equivalence of user's demands or parts of user's demands in a more or less abstract manner, e. g. a product, a product with an attribute, or products of a product group with an approximate price. Such entities have to be identified on both the indexing and the querying side of a search application in order to achieve the most reliable matches between the demands of users and the offering of the shop.

Generally, there are two major approaches to extract entities from text: Those based on defining rule collections and those based on machine learning. While rule collections usually cover the deterministic human-understandable parts of semantic structures in text, machine learning is also capable to cover more fuzzy and less deterministic textures. However, as both approaches are associated with benefits, challenges, and risks, discussions on how to implement semantic search should not be limited to methodological considerations. The success of such projects depends on the abilities and specializations of analysts and scientists as well as on how a project is managed and embedded into an organization.

The presentation will focus on contrasting the two major approaches of implementing entity recognition in the context of semantic search. Further, organizational and human factors will be considered. Finally, a specific rule-based approach to semantic search will be discussed.

Johannes works as product manager for search and recommendations at MediaMarktSaturn. His responsibilities include driving the vision and the development of search engine solutions for all saleslines in Europe. Before he moved to MediaMarktSaturn, Johannes was a consultant and architect for search, big data and Natural Language Processing in the finance and insurance industries.

(Slides)

Lunch

Learning to rank (LTR) in an Activity Marketplace

Ashraf Aaref & Felipe Besson
GetYourGuide

Learning to rank (LTR) has been considered the next generation tool to improve relevance of e-commerce search solutions. Motivated by the challenges we have at GetYourGuide, a global marketplace for tours and activities, we started a project to introduce LTR in our Search Engine. In this talk, we would like to share our Logbook from day 1 to the current project status. It covers decisions, lessons learned, good practices and pain points you need to know when applying LTR from scratch such as:

• Know what is relevance means for your business and how diverse are real user queries
• Pick the right tools to collect data, train and run ML models faster
• Challenge the quality of your training set and how you collect judgments after each iteration

Ranking better is an ongoing project. Each of previous topics are still bringing us many good learnings. We would like to share them on this talk and provide good practical contributions to new adopters of LTR.

Ashraf has been working as a software engineer for more than ten years and is currently part of the Search and Recommendation team at GetYourGuide. In the last one and a half years he has been focussing on search and has become very excited about challenges in implementing search, especially about relevance in Marketplaces.

Felipe has a Masters degree in Computer Science by the University of Sao Paulo, Brazil. His research was dedicated to Agile development and Methodologies. Currently, he works as Data Engineer at GetYourGuide. He is part of Search and Recommendation team. Felipe works with big data and information retrieval in the last five years. He can't imagine how people could survive without Search engines.

(Slides)

Next generation product search at Zalando using deep learning

Duncan Blythe
Zalando

Fashion search is quite a challenging problem, as it needs to support all kinds of search types. Sometimes, people just look for white T-shirts, but other times, they are looking for clothing for special occasions, or based on current trends. At Zalando, our in house search engine currently matches full text queries with products through a cascaded architecture. Each step in the cascade processes the input in a specific way, for example, locating mentions of brands, spell checking, and disambiguation. Finally the preprocessed input is used to filter the product attributes (such as color=red, brand=Nike). This architecture has multiple drawbacks, such as fragility, limited scalability and extensibility. In particular, it does not work well on search terms which are not exactly matched by the data on the articles.

We are investigating replacing part or all of this cascade with a single step end-2-end deep learning architecture which involves no textual preprocessing and directly filters image content as well as product meta-data. I will describe the components involved in such a system, as well as potential advantages and disadvantages.

Duncan is a member of the NLP group at Zalando research, where he is applying multimodal deep learning to image and textual data. In the past he has worked on a variety of applications of machine learning, including neuro-imaging, sports science, causal-analysis and genetics. Previously he obtained his PhD from the Technical University of Berlin and his Masters degree from the University of Oxford.

(Slides)

Break

Your business depends on supporting open source search!

Doug Turnbull
OpenSource Connections

Fractured e-commerce firms can never hope to compete with Amazon in search. As an e-commerce search community, we are hesitant to share - often excessively paranoid about competition. Yet sharing means survival. Open source needs to be thought of as a fundamental way of doing business. When it comes to building smarter search - 'relevance' - reinventing common tooling is anchoring us to a lower level of maturity. It means we spend a year stuck in plumbing and infrastructure, rather than innovating for our business problems.

In this talk, we present a business case for why technical and non-technical stakeholders need to risk sharing and collaborating more, not less, through open source.

Doug is author of Relevant Search and the CTO at OpenSource Connections. Doug's team empowers search and discovery teams to reach their potential through consulting and training. He developed a search relevance methodology that guides businesses on making relevance investments. Doug loves taking the mystery out of machine learning and relevance ranking by building tools, blogging, and speaking.

(Slides)

Self-organising sessions / Barcamp

Informal closing session and drinks