Recommender system | EPFL Graph Search

A recommender system, or a recommendation system (sometimes replacing 'system' with a synonym such as platform or engine), is a subclass of information filtering system that provide suggestions for items that are most pertinent to a particular user. Typically, the suggestions refer to various decision-making processes, such as what product to purchase, what music to listen to, or what online news to read. Recommender systems are particularly useful when an individual needs to choose an item from a potentially overwhelming number of items that a service may offer. Recommender systems are used in a variety of areas, with commonly recognised examples taking the form of playlist generators for video and music services, product recommenders for online stores, or content recommenders for social media platforms and open web content recommenders. These systems can operate using a single input, li

Aggregating information from the crowd

Florent Frédéric Garcin

With an ever-growing amount of data generated on the web, aggregating information from the crowd into meaningful knowledge has become crucial to companies in order to create a competitive edge. Not only companies, but also organizations and governments can benefit from this aggregation. In this thesis, we illustrate the benefits of aggregating information from the crowd with three applications. In the first application, we investigate different ways of aggregating users’ ratings on review websites. With the help of gamification techniques, we introduce a new methodology for studying users’ rating behaviour, and show that there is a balance between the amount of private information elicited from the crowd and the accuracy of the aggregated rating. In the second application, we consider the aggregation of implicit feedbacks from the crowd in order to generate personalized recommendations of news articles. We propose a new class of recommender systems based on Context Trees (CT), specifically designed for a dynamic domain like the news. We show that CT recommender systems generate accurate and novel recommendations in an offline setting, but also in real time on the newspaper website swissinfo.ch. In the last application, we address the problem of eliciting private information to predict outcomes of events. We report on an experimental platform called swissnoise that allows users to express their opinions on various topics ranging from sports to politics. It is the first platformto implement a peer prediction mechanism for online opinion polls. We show that peer prediction can be practically implemented, and discuss the design choices and variations made to such mechanism. Finally, we find that peer prediction achieves a performance comparable to that of prediction markets

EPFL2014

Scalable intelligent electronic catalogs

Marc Torrens Arnal

The world today is full of information systems which make huge quantities of information available. This incredible amount of information is clearly overwhelming Internet endusers. As a consequence, intelligent tools to identify worthwhile information are needed, in order to fully assist people in finding the right information. Moreover, most systems are ultimately used, not just to provide information, but also to solve problems. Encouraged by the growing popular success of Internet and the enormous business potential of electronic commerce, e-catalogs have been consolidated as one of the most relevant types of information systems. Nearly all currently available electronic catalogs are offering tools for extracting product information based on key-attribute filtering methods. The most advanced electronic catalogs are implemented as recommender systems using collaborative filtering techniques. This dissertation focuses on strategies for coping with the difficulty of building intelligent catalogs which fully support the user in his purchase decision-making process, while maintaining the scalability of the whole system. The contributions of this thesis lie on a mixed-initiative system which is inspired by observations on traditional commerce activities. Such a conversational model consists basically of a dialog between the customer and the system, where the user criticizes proposed products and the catalog suggests new products accordingly. Constraint satisfaction techniques are analyzed in order to provide a uniform framework for modeling electronic catalogs for configurable products. Within the same framework, user preferences and optimization constraints are also easily modeled. Searching strategies for proposing the adequate products according to criteria are described in detail. Another dimension of this dissertation faces the problem of scalability, i.e., the problem of supporting hundreds, or thousands of users simultaneously using intelligent electronic catalogs. Traditional wisdom would presume that in order to provide full assistance to users in complex tasks, the business logic of the system must be complex, thus preventing scalability. SmartClient is a software architectural model that uses constraint satisfaction problems for representing solution spaces, instead of traditional models which represent solution spaces by collections of single solutions. This main idea is supported by the fact that constraint solvers are extreme in their compactness and simplicity, while providing sophisticated business logic. Different SmartClient architecture configurations are provided for different uses and architectural requirements. In order to illustrate the use of constraint satisfaction techniques for complex electronic catalogs with the SmartClient architecture, a commercial Internet-based application for travel planning, called reality, has been successfully developed. Travel planning is a particularly appropriate domain for validating the results of this research, since travel information is dynamic, travel planning problems are combinatorial, and moreover, complex user preferences and optimization constraints must be taken into consideration.

EPFL2003

Lifelong Machine Learning with Data Efficiency and Knowledge Retention

Fei Mi

Artificial intelligence (AI) and machine learning (ML) have become de facto tools in many real-life applications to offer a wide range of benefits for individuals and our society. A classic ML model is typically trained with a large-scale static dataset in an offline manner. Therefore, it can not quickly capture new knowledge in non-stationary environments, and it is difficult to maintain long-term memory for knowledge learned earlier. In practice, many ML systems often need to learn new knowledge (e.g., domains, tasks, distributions, etc.) as more data and experiences are collected, which is referred to as a lifelong ML paradigm in this thesis. We focus on two fundamental challenges to achieve lifelong learning. The first challenge is to quickly learn new knowledge with a small number of observations, and we refer to it as data efficiency. The second challenge is to prevent an ML system from forgetting the old knowledge it has previously learned, and we refer to this challenge as knowledge retention. These two capabilities are crucial for applying ML to most practical applications. In this thesis, we study three important applications with these two challenges, including recommendation systems, task-oriented dialog systems, and the image classification task. First, we propose two approaches to improve data efficiency for task-oriented dialog systems. The first proposed approach is based on Meta-learning, aiming to learn a better model parameter initialization from training data. It can quickly reach a good parameter region of new domains or tasks with a small number of labeled data. The second proposal takes a semi-supervised self-training approach to iteratively train a better model using sufficient unlabeled data when only a limited number of labeled data are available. We empirically demonstrate that both approaches effectively improve data efficiency to learn new knowledge. The second self-training method even consistently improves state-of-the-art large-scale pre-trained models. Second, we tackle the knowledge retention challenge to mitigate the detrimental catastrophic forgetting issue when neural networks learn new knowledge sequentially. We formulate and investigate the ``continual learning'' setting for task-oriented dialog systems and recommendation systems. Through extensive empirical evaluation and analysis, we demonstrate the importance of (1) exemplar replay: storing representative historical data and replaying them to the model while learning new knowledge; (2) dynamic regularization: applying a dynamic regularization term to put flexible constraints on not forgetting previously learned knowledge in each model update cycle. Lastly, we conduct several initial attempts to achieve both data efficiency and knowledge retention in a unified framework. In the recommendation scenario, we propose two approaches using different non-parametric memory modules to retain long-term knowledge. More importantly, the two proposed non-parametric predictions computed on top of them help learn and memorize new knowledge in a data-efficient manner. Apart from the recommendation scenario, we propose a probabilistic evaluation protocol in the widely studied image classification domain. It is general and versatile to simulate a wide range of realistic lifelong learning scenarios that require both knowledge retention and data efficiency for studying different techniques. Through experiments, we also demonstrate the benefit

EPFL2021