Pawel Matuszyk

Selective Forgetting for Incremental Matrix Factorization in Recommender Systems

Recommender Systems are used to build models of users’ preferences. Those models should reflect current state of the preferences at any timepoint. The preferences, however, are not static. They are subject to concept drift or even shift, as it is known from e.g. stream mining. They undergo permanent changes as the taste of users and perception of items change over time. Therefore, it is crucial to select the actual data for training models and to forget the outdated ones.

A feature-based personalized recommender system for product-line configuration

Today’s competitive marketplace requires the industry to understand unique and particular needs of their customers. Product line practices enable companies to create individual products for every customer by providing an interdependent set of features. Users configure personalized products by consecutively selecting desired features based on their individual needs. However, as most features are interdependent, users must understand the impact of their gradual selections in order to make valid decisions. Thus, especially when dealing with large feature models, specialized assistance is needed to guide the users in configuring their product. Recently, recommender systems have proved to be an appropriate mean to assist users in finding information and making decisions. In this paper, we propose an advanced feature recommender system that provides personalized recommendations to users. In detail, we offer four main contributions: (i) We provide a recommender system that suggests relevant features to ease the decision-making process. (ii) Based on this system, we provide visual support to users that guides them through the decision-making process and allows them to focus on valid and relevant parts of the configuration space. (iii) We provide an interactive open-source configurator tool encompassing all those features. (iv) In order to demonstrate the performance of our approach, we compare three different recommender algorithms in two real case studies derived from business experience.

Correcting the Usage of the Hoeffding Inequality in Stream Mining

Many stream classification algorithms use the Hoeffding Inequality to identify the best split attribute during tree induction. We show that the prerequisites of the Inequality are violated by these algorithms, and we propose corrective steps. The new stream classification core, correctedVFDT , satisfies the prerequisites of the Hoeffding Inequality and thus provides the expected performance guarantees. The goal of our work is not to improve accuracy, but to guarantee a reliable and interpretable error bound. Nonetheless, we show that our solution achieves lower error rates regarding split attributes and sooner split decisions while maintaining a similar level of accuracy.

Predicting the Performance of Collaborative Filtering Algorithms

Collaborative Filtering algorithms are widely used in recommendation engines, but their performance varies widely. How to predict whether collaborative filtering is appropriate for a specific recommendation environment without running the algorithm on the dataset, nor designing experiments? We propose a method that estimates the expected performance of CF algorithms by analysing only the dataset statistics. In particular, we introduce measures that quantify the dataset properties with respect to user co-ratings, and we show that these measures predict the performance of collaborative filtering on the dataset, when trained on a small number of benchmark datasets.

Hoeffding-CF: Neighbourhood-Based Recommendations on Reliably Similar Users

Neighbourhood-based collaborative filtering recommenders exploit the common ratings among users to identify a user’s most similar neighbours. It is known that decisions made on a naive computation of user similarity are unreliable, because the number of co-ratings varies strongly among users. In this paper, we formalize the notion of reliable similarity between two users and propose a method that constructs a user’s neighbourhood by selecting only those users that are reliably similar to her. Our method combines a statistical test and the notion of a baseline user. We report our results on typical benchmark datasets.

Semi-supervised Learning for Stream Recommender Systems

Recommender systems suffer from an extreme data sparsity that results from a large number of items and only a limited capability of users to perceive them. Only a small fraction of items can be rated by a single user. Consequently, there is plenty of unlabelled information that can be leveraged by semi-supervised methods. We propose the first semi-supervised framework for stream recommender systems that can leverage this information incrementally on a stream of ratings. We design several novel components, such as a sensitivity-based reliability measure, and extend a state-of-the-art matrix factorization algorithm by the capability to extend the dimensions of a matrix incrementally as new users and items occur in a stream. We show that our framework improves the quality of recommendations at nearly all time points in a stream.

Forgetting techniques for stream-based matrix factorization in recommender systems

Forgetting is often considered a malfunction of intelligent agents; however, in a changing world forgetting has an essential advantage. It provides means of adaptation to changes by removing effects of obsolete (not necessarily old) information from models. This also applies to intelligent systems, such as recommender systems, which learn users’ preferences and predict future items of interest. In this work, we present unsupervised forgetting techniques that make recommender systems adapt to changes of users’ preferences over time. We propose eleven techniques that select obsolete information and three algorithms that enforce the forgetting in different ways. In our evaluation on real-world datasets, we show that forgetting obsolete information significantly improves predictive power of recommender systems.

Personalized recommender systems for product-line configuration processes

Abstract Product lines are designed to support the reuse of features across multiple products. Features are product functional requirements that are important to stakeholders. In this context, feature models are used to establish a reuse platform and allow the configuration of multiple products through the interactive selection of a valid combination of features. Although there are many specialized configurator tools that aim to provide configuration support, they only assure that all dependencies from selected features are automatically satisfied. However, no support is provided to help decision makers focus on likely relevant configuration options. Consequently, since decision makers are often unsure about their needs, the configuration of large feature models becomes challenging. To improve the efficiency and quality of the product configuration process, we propose a new approach that provides users with a limited set of permitted, necessary and relevant choices. To this end, we adapt six state-of-the-art recommender algorithms to the product line configuration context. We empirically demonstrate the usability of the implemented algorithms in different domain scenarios, based on two real-world datasets of configurations. The results of our evaluation show that recommender algorithms, such as CF-shrinkage, CF-significance weighting, and BRISMF, when applied in the context of product-line configuration can efficiently support decision makers in a most efficient selection of features.

Scalable Online Top-N Recommender Systems

Given the large volumes and dynamics of data that recommender systems currently have to deal with, we look at online stream based approaches that are able to cope with high throughput observations. In this paper we describe work on incremental neighborhood based and incremental matrix factorization approaches for binary ratings, starting with a general introduction, looking at various approaches and describing existing enhancements. We refer to recent work on forgetting techniques and multidimensional recommendation. We will also focus on adequate procedures for the evaluation of online recommender algorithms.

Framework for Storing and Processing Relational Entities in Stream Mining

Relational stream mining involves learning a model on relational entities, which are enriched with information from further streams that reference them. To incorporate such information into the entities in an efficient incremental way, we propose a multi-threaded framework with a weighting function that prioritizes the entities delivered to the learner for learning and adaption to drift. We further propose a generator for drifting relational streams, and use it to show that our framework reaches substantial reduction of computation time.