Search Technology Meetup XVI: Datascience in search

The sixteenth edition of the Search Technology Meetup Hamburg: Come and join us for food, drinks and another edition of awesome inspiring search related talks.
After getting inspired during the talks, you will also have the opportunity to meet with other tech guys to discuss your search tech stuff and to share knowledge while having a drink as well as some food.

Thanks to Immowelt for sponsoring snacks, drinks and location and Mondia for pizza. All talks will be held in english.

----------------------------------
# Music Information Retrieval mit TensorFlow:
Music information retrieval (MIR) is concerned with extracting information from music, for example by processing an MP3 file or a YouTube video.

This talk will demo a small application that transcribes the melody of an audio recording. Afterwards, we walk through the required
preprocessing of the raw audio signal and a machine learning model, implemented using TensorFlow, to map from the processed audio signal to a sequence of notes. These notes in turn can be used for several things, e.g. searching for similar melodies.

Florian Müller, Senior Data Scientist/Engineer at Risk42 Software

----------------------------------
# Predicting business performance in e-commerce search:
Despite the broad availability of voice-activated technology, a typical e-commerce search user still submits keywords into a search box before buying a product on-line. In response to her query she expects to be presented with product offers that are relevant to her (potentially subjective) intent.
Most of the users start by formulating vague search phrases consisting of only one or two terms.
Result rankings in a typical e-commerce search setting are produced with a combination of topical (or user) relevance and business value. Depending on the user query the two criteria might complement each other for better or worse ranking performance. The goal is to provide the best utility for our users by respecting the business value of the presented items.
In this talk, we will share our approach to estimate the business value of each product on any item listing on otto.de. The goal of a business value prediction model is to provide performance estimation per item that models future business impact better than plain historical data. One of the complex problems in e-commerce search is that the business value of an item might depend on the context (or the query) it is presented in. For the sake of applicability of the business values outside of the search, we modeled our data as being context-independent.
We will share our journey of replacing a logistic regression model with a state of the art gradient boosting regression model. The key challenges were as follows:
• Finding the right key to split the data
• Selecting training and validation data time periods and window sizes
• Applying appropriate data cleansing strategies
• Handling missing data
• Finding the best comparison metrics for model performance
• Offline evaluation of search results with new/old models
We will talk about the major problems we faced during the phases of creating our new models and incorporating them into our search and navigation results.

Andrea Schütt (Data Scientist) & Emmy Le (Productmanager) at OTTOs search team