Deep Learning for Matching in Search and Recommendation - Cap 1 Introduction - Leitura de Artigo

Jun Xu, Xiangnan He and Hang Li (2020), “Deep Learning for Matching in Search and Recommendation”, Foundations and Trends in Information Retrieval: Vol. 14, No. 2–3, pp 102–288. DOI: 10.1561/1500000076

ABSTRACT
Matching is a key problem in both search and recommendation, which is to measure the relevance of a document to a query or the interest of a user to an item. Machine learning has been exploited to address the problem, which learns a matching function based on input representations and from labeled data, also referred to as “learning to match”. In recent years, efforts have been made to develop deep learning techniques for matching tasks in search and recommendation.With the availability of a large amount of data, powerful computational resources, and advanced deep learning techniques, deep learning for matching now becomes the state-of-the-art technology for search and recommendation. The key to the success of the deep learning approach is its strong ability in learning of representations and generalization of matching patterns from data (e.g., queries, documents, users, items, and contexts, particularly in their raw forms).

The goal is to present the user only the information that is of interest and relevance, at the right time, place, and context. Nowadays, two types of information accessing paradigms, search and recommendation, are widely used in a great variety of scenarios.

The fundamental mechanism of search is “pull”, because users first make specific requests (i.e., submit queries) and then receive information. The fundamental mechanisms of recommendation is “push”, because users are provided information which they do not specifically request (e.g., submit queries). Here “beneficiary” means the people whose interests are to be met in the task. In a search engine,
the results are typically created solely based on the user’s needs, and thus the beneficiary is the users. In a recommendation engine, the results usually need to satisfy both the users and providers, and thus the beneficiary is all of them.

As for “serendipity”, it means that conventional search focuses more on information that is clearly relevant. Conventional recommendation, on the other hand, is allowed to offer unexpected but useful
information.

It is also indicated that search (information retrieval) and recommendation (information filtering) are the two sides of the same coin, having strong connections and similarities (Belkin and Croft, 1992).

There is a clear trend that search and recommendation will be integrated into a single system at certain scenarios to meet users’ needs better, where matching plays an essential role.

With the use of deep learning technologies, the matching models for search and recommendation bear even more resemblance in architecture and methodology, as reflected in the techniques: embedding
the inputs (queries, users, documents, and items) as distributed representations, combining neural network components to represent the matching function, and training the model parameters in an end-to-end manner.

A high match degree at the text level does not necessarily mean high relevance at the semantic level, and vice versa. Moreover, queries are issued by users, while documents are compiled by editors. Due
to the ambiguity of natural language, users and editors are likely to use different language styles and expressions for presenting the same concepts or topics. As a result, the search system may suffer from the so called query-document mismatch problem. Specifically, when the users of a search engine and the editors of the documents use different texts to describe the same concept (e.g., “ny times” vs. “new york times”), query-document mismatch may occur.

To address the query-document mismatch challenge, methods have been proposed to perform matching at the semantic level, referred to as semantic matching. ... Both traditional machine learning approaches (Li and Xu, 2014) and deep learning approaches (Guo et al., 2019b; Mitra and Craswell, 2018; Onal et al., 2018) have been developed for semantic matching.

The mismatching problem is even more severe in recommendation. Since the features of users and items are from the spaces of different semantics, the naive approaches based on the matching of superficial features do not work for recommendation. More challengingly, the items can be described by multi-modal features, e.g., images of clothing products and cover images of movies, which could play a pivotal role in affecting the decision-making of users.

To address the mismatching challenge in recommendation, the collaborative filtering principle has been proposed (Shi et al., 2014). Collaborative Filtering (CF), which works as the fundamental basis of
almost all personalized recommender systems, assumes that a user may like (consume) the items that are liked (consumed) by the similar users, for which the similarity is judged from the historical interactions (Sarwar et al., 2001). However, directly evaluating the similarity between users (items) suffers from the sparsity issue, since a user only consumed a few items in the whole item space.

Besides matrix factorization, many other types of CF methods have been developed like neural network-based methods (He et al., 2017c; Liang et al., 2018) and graph-based methods (Wang et al., 2019b; Ying et al., 2018).

The power of deep learning models lies in the ability to learn distributed representations from the raw
data (e.g., text) for the matching problem, to avoid many limitations of hand-crafted features, and to learn the representations and matching networks in an end-to-end fashion.

In search, the mismatch between query and document is more effectively addressed by deep neural networks, including the feed-forward neural networks (FFNs), convolutional neural networks (CNNs), and Recurrent neural networks (RNNs), because they have stronger capabilities in representation learning and matching function learning. Most notably, Bidirectional Encoder Representations from Transformers (BERT) has significantly enhanced the accuracy of matching in search and stands out as the state-of-the-art technique now.

In recommendation, recent focus has shifted from behavior-centric collaborative filtering to information-rich user-item matching as in sequential, context-aware, and knowledge graph enhanced recommendations, which are all practical scenario-driven.

=======================================================

Definições

deep learning

traditional machine learning

neural network 

deep neural network: feed-forward neural networks (FFNs), convolutional neural networks (CNNs), and Recurrent neural networks (RNNs). Deep neural networks are complicated nonlinear functions from input to output.(capítulo 3)

=======================================================

Apresentação -> https://youtu.be/HbnhB-taS-c   &   https://youtu.be/JqkvH7lEfak

Artigo sobre utilização de DL para Search e Recomendation

Como atender a necessidade de informação de um usuário dado um conjunto grande de informações? Recuperando informações de interesse, relevantes e no contexto correto

Search Engines: trabalham no modo pull, beneficiando os usuários e sem "serendipidade" X Recommendation: trabalham no modo push, beneficiando o fornecedor e trazem "serendipidade"

Uma perspectiva integrada em arquiteturas de Busca e Recomendação

 

O problema do mismatch pela falta de semântica: o matching é feito de modo exato com as palavras.

Semantic matching trata tanto a representação semântica da query quanto do corpus assim como as funções de match

ML trouxe avanços mas a DL permitiu superar as limitações da engenharia (manual) de features, do aprendizado de representações e correspondência de redes. 

Avanços da DL (deep neural networds): FFN, CNN, RNN, BERT