A large number of empirical studies on applying self-attention models in the domain of recommender systems are based on offline evaluation and metrics computed on standardized datasets. Moreover, many of them do not consider side information such as item and customer metadata although deep-learning recommenders live up to their full potential only when numerous features of heterogeneous type are included. Also, normally the model is used only for a single use case. Due to these shortcomings, even if relevant, previous works are not always representative of their actual effectiveness in real-world industry applications. In this talk, we contribute to bridging this gap by presenting live experimental results demonstrating improvements in user retention of up to 30%. Moreover, we share our learnings and challenges from building a re-usable and configurable recommender system for various applications from the fashion industry. In particular, we focus on fashion inspiration use-cases, such as outfit ranking, outfit recommendation and real-time personalized outfit generation.

Ranked result lists with query-dependent snippets have become state of the art in text search. They are typically implemented by searching, at query time, for occurrences of the query words in the top-ranked documents. This document-based approach has three inherent problems: (i) when a document is indexed by terms which it does not contain literally (e.g., related words or spelling variants), localization of the corresponding snippets becomes problematic; (ii) each query operator (e.g., phrase or proximity search) has to be implemented twice, on the index side in order to compute the correct result set, and on the snippet-generation side to generate the appropriate snippets; and (iii) in a worst case, the whole document needs to be scanned for occurrences of the query words, which could be problematic for very long documents. We present a new index-based method that localizes snippets by information solely computed from the index and that overcomes all three problems. Unlike previous index-based methods, we show how to achieve this at essentially no extra cost in query processing time, by a technique we call operator inversion . We also show how our index-based method allows the caching of individual segments instead of complete documents, which enables a significantly larger cache hit-ratio as compared to the document-based approach. We have fully integrated our implementation with the CompleteSearch engine.

We consider the problem of fuzzy full-text search in large text collections, that is, full-text search which is robust against errors both on the side of the query as well as on the side of the documents. Standard inverted-index techniques work extremely well for ordinary full-text search but fail to achieve interactive query times (below 100 milliseconds) for fuzzy full-text search even on moderately-sized text collections (above 10 GBs of text). We present new preprocessing techniques that achieve interactive query times on large text collections (100 GB of text, served by a single machine). We consider two similarity measures, one where the query terms match similar terms in the collection (e.g., algorithm matches algoritm or vice versa) and one where the query terms match terms with a similar prefix in the collection (e.g., alori matches algorithm). The latter is important when we want to display results instantly after each keystroke (search as you type). All algorithms have been fully integrated into the CompleteSearch engine.

As shown in a series of recent works, the HYB index is an alternative to the inverted index (INV) that enables very fast prefix searches, which in turn is the basis for fast processing of many other types of advanced queries, including autocompletion, faceted search, error-tolerant search, database-style select and join, and semantic search. In this work we show that HYB can be constructed at least as fast as INV, and often up to twice as fast. This is because HYB, by its nature, requires only a half-inversion of the data and allows an efficient in-place instead of the traditional merge-based index construction. We also pay particular attention to the cache efficiency of the in-memory posting accumulation, an issue that has not been addressed in previous work, and show that our simple multilevel posting accumulation scheme yields much fewer cache misses compared to related approaches. Finally, we show that HYB supports fast dynamic index updates more easily than INV.