Scalable intelligent electronic catalogs

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.

Learning Explainable User Sentiment and Preferences for Information Filtering

Nikolaos Pappas

In the last decade, online social networks have enabled people to interact in many ways with each other and with content. The digital traces of such actions reveal people's preferences towards online content such as news or products. These traces often result from interactions such as sharing or liking, but also from interactions in natural language. The continuous growth of the amount of content and of digital traces has led to information overload: surrounded by large volumes of information, people are facing difficulties when searching for information relevant to their interests. To improve user experience, information systems must be able to assist users in achieving their search goals, effectively and efficiently. This thesis is concerned with two important challenges that information systems need to address in order to significantly improve search experience and overcome information overload. First, these systems need to model accurately the variety of user traces, and second, they need to meaningfully explain search results and recommendations to users. To address these challenges, this thesis proposes novel methods based on machine learning to model user sentiment and preferences for information filtering systems, which are effective, scalable, and easily interpretable by humans. We focus on two prominent types of user traces in social networks: on the one hand, user comments accompanied by unary preferences such as likes, and on the other hand, user reviews accompanied by numerical preferences such as star ratings. In both cases, we advocate that by better understanding user text through mining its semantics and modeling its structure, we can not only improve information filtering, but also explain predictions to users. Within this context, we aim to answer three main research questions, namely: (i)~how do item semantics help to predict unary preferences; (ii)~how do sentiments of free-form user texts help to predict unary preferences; and (iii)~how to model fine-grained numerical preferences from user review texts. Our goal is to model and extract from user text the knowledge required to answer these questions, and to obtain insights on how to design better information filtering systems that are more effective and improve user experience. To answer the first question, we formulate the recommendation problem based on unary preferences as a top-N retrieval task and we define an appropriate dataset and metrics for measuring performance. Then, we propose and evaluate several content-based methods based on semantic similarities under presence or absence of preferences. To answer the second question, we propose a sentiment-aware neighborhood model which integrates the sentiment of user comments with unary preferences, either through fixed or through learned mapping functions. For the latter type, we propose a learning algorithm which adapts the sentiment of user comments to unary preferences at collective or individual levels. To answer the third question, we cast the problem of modeling user attitude toward aspects of items as a weakly supervised problem, and we propose a weighted multiple-instance learning method for solving it. Lastly, we show that the learned saliency weights, apart from being easily interpretable, are useful indicators for review segmentation and summarization.

EPFL2016

MEM0R1ES: Memory-based Information Systems

Michele Catasta

The advent of the Web reshaped the way in which humans memorize information, arguably unlike any other technological advance of the last decades. Rather than remembering the information itself, people are primed to find the needed information through Web search engines. At the light of this result, psychologists are starting to reconsider the standard framework for which memory is the process by which information is encoded, stored, and retrieved. Content that can be easily retrieved from the Web is encoded to enhance the recall for where to access it; therefore it is only partially stored. As a matter of fact, Internet users are exposed daily to a staggering amount of personal information, especially since Online Social Networks started to play a key role in their lives (i.e., smartphone users spend 20% of their time on the Facebook app). While a great deal of effort has been spent in recommending relevant content to the user (e.g., targeted advertisements, similar news, etc.), there are no comprehensive studies on how much of this information is memorized by the user. In this thesis, we will show the feasibility of memory-based information systems, namely, systems that take into account the idiosyncrasies of human memories. Departing from classical integrated infrastructures providing static views over heterogeneous sources (e.g., through a wrapper-mediator infrastructure or inverted indices), we propose to mimic the way our brain stores and accesses information in order to provide a more natural extension of our cognitive functions when it comes to tap into the chaotic piles of digital information that we generate daily. Building an information system that mimics the way our brain stores and accesses data would dramatically reduce the cognitive burden of the user interactions, while at the same time returning more relevant information. This thesis is divided in two parts. In the first part, we present our effort in redefining 3 fundamental information systems, taking into account the way in which human memories work. Namely, we rethought a search architecture over personal data, we devised a new crowdsourcing paradigm (leveraging the Transactive Memory paradigm), and we envisioned a novel Database Management System. In the second part, we introduce our long-term efforts to gain insights on the human memories in a non-invasive way, by developing appealing software applications that could reach thousands of users. Such work-in-progress has a fundamental importance when it comes to design novel information systems (as we did in the first part of this thesis), because it allows us to rank the information based on its memorability. In what is currently called the ``information era'', showing only what is memorable to the user could become the key feature to tame the deluge of data we are exposed to daily.

EPFL2015

Fast Multi-Step Critiquing for VAE-based Recommender Systems

Diego Matteo Antognini, Boi Faltings

Recent studies have shown that providing personalized explanations alongside recommendations increases trust and perceived quality. Furthermore, it gives users an opportunity to refine the recommendations by critiquing parts of the explanations. On one hand, current recommender systems model the recommendation, explanation, and critiquing objectives jointly, but this creates an inherent trade-off between their respective performance. On the other hand, although recent latent linear critiquing approaches are built upon an existing recommender system, they suffer from computational inefficiency at inference due to the objective optimized at each conversation's turn. We address these deficiencies with M&Ms-VAE, a novel variational autoencoder for recommendation and explanation that is based on multimodal modeling assumptions. We train the model under a weak supervision scheme to simulate both fully and partially observed variables. Then, we leverage the generalization ability of a trained M&Ms-VAE model to embed the user preference and the critique separately. Our work's most important innovation is our critiquing module, which is built upon and trained in a self-supervised manner with a simple ranking objective. Experiments on four real-world datasets demonstrate that among state-of-the-art models, our system is the first to dominate or match the performance in terms of recommendation, explanation, and multi-step critiquing. Moreover, M&Ms-VAE processes the critiques up to 25.6x faster than the best baselines. Finally, we show that our model infers coherent joint and cross generation, even under weak supervision, thanks to our multimodal-based modeling and training scheme.