Michal Ciesielczyk

Semantically enhanced collaborative filtering based on RSVD

We investigate a hybrid recommendation method that is based on two-stage data processing - first dealing with content features describing items, and then handling user behavioral data. The evaluation of the proposed method is oriented on the so-called find-good-items task, rather than on the low-error-of-ratings prediction. We focus on a case of extreme collaborative data sparsity. Our method is a combination of content features preprocessing performed by means of Random Indexing (RI), a reflective retraining of preliminary reduced item vectors according to collaborative filtering data, and vector space optimization based on Singular Value Decomposition (SVD). We demonstrate that such an approach is appropriate in high data sparsity scenarios, which disqualify the use of widely-referenced collaborative filtering methods, and allows to generate more accurate recommendations than those obtained through a hybrid method based on weighted feature combination. Moreover, the proposed solution allows to improve the recommendation accuracy without increasing the computational complexity.

RSVD-based Dimensionality Reduction for Recommender Systems

We investigate dimensionality reduction methods from the perspective of their ability to produce a low-rank customer-product matrix representation. We analyze the results of using collaborative filtering based on SVD, RI, Reflective Random Indexing (RRI) and Randomized Singular Value Decomposition (RSVD) from the perspective of selected algebraic (i.e. application-independent) properties. We show that the Frobenius-norm optimality of SVD does not correspond to the optimal recommendation accuracy, when measured in terms of F1. On the other hand, a high collaborative filtering quality is achievable when a matrix decomposition - based on a combination of RRI and SVD referred to as RSVD-RRI - leads to increased diversity of low-dimensional eigenvectors. The diversity is observable from the perspective of cosine similarities analyzed in comparison to the analogical case of SVD. Such a feature is more desirable than the fidelity of the input matrix spectrum representation, despite the MSE-optimality of SVD.

Tensor-Based Modeling of Temporal Features for Big Data CTR Estimation

In this paper we propose a simple tensor-based approach to temporal features modeling that is applicable as means for logistic regression (LR) enhancement. We evaluate experimentally the performance of an LR system based on the proposed model in the Click-Through Rate (CTR) estimation scenario involving processing of very large multi-attribute data streams. We compare our approach to the existing approaches to temporal features modeling from the perspective of the Real-Time Bidding (RTB) CTR estimation scenario. On the basis of an extensive experimental evaluation, we demonstrate that the proposed approach enables achieving an improvement of the quality of CTR estimation. We show this improvement in a Big Data application scenario of the Web user feedback prediction realized within an RTB Demand-Side Platform.

Progressive Random Indexing: Dimensionality Reduction Preserving Local Network Dependencies

The vector space model is undoubtedly among the most popular data representation models used in the processing of large networks. Unfortunately, the vector space model suffers from the so-called curse of dimensionality, a phenomenon where data become extremely sparse due to an exponential growth of the data space volume caused by a large number of dimensions. Thus, dimensionality reduction techniques are necessary to make large networks represented in the vector space model available for analysis and processing. Most dimensionality reduction techniques tend to focus on principal components present in the data, effectively disregarding local relationships that may exist between objects. This behavior is a significant drawback of current dimensionality reduction techniques, because these local relationships are crucial for maintaining high accuracy in many network analysis tasks, such as link prediction or community detection. To rectify the aforementioned drawback, we propose Progressive Random Indexing, a new dimensionality reduction technique. Built upon Reflective Random Indexing, our method significantly reduces the dimensionality of the vector space model while retaining all important local relationships between objects. The key element of the Progressive Random Indexing technique is the use of the gain value at each reflection step, which determines how much information about local relationships should be included in the space of reduced dimensionality. Our experiments indicate that when applied to large real-world networks (Facebook social network, MovieLens movie recommendations), Progressive Random Indexing outperforms state-of-the-art methods in link prediction tasks.

Collaborative Filtering Based on Bi-Relational Data Representation

Abstract Widely-referenced approaches to collaborative filtering (CF) are based on the use of an input matrix that represents each user profile as a vector in a space of items and each item as a vector in a space of users. When the behavioral input data have the form of (userX, likes, itemY) and (userX, dislikes, itemY) triples one has to propose a representation of the user feedback data that is more suitable for the use of propositional data than the ordinary user-item ratings matrix. We propose to use an element-fact matrix, in which columns represent RDF-like behavioral data triples and rows represent users, items, and relations. By following such a triple-based approach to the bi-relational behavioral data representation we are able to improve the quality of collaborative filtering. One of the key findings of the research presented in this paper is that the proposed bi-relational behavioral data representation, while combined with reflective matrix processing, significantly outperforms state-of-the-art collaborative filtering methods based on the use of a ‘standard’ user-item matrix.

Logistic Regression Setup for RTB CTR Estimation

In this paper we investigate one of the most interesting problems of Big Data user feedback prediction which is the Real-Time Bidding Click-Through Rate estimation. We evaluate experimentally the impact of the widely-referenced methods for optimization of the logistic regression - the state-of-the art Real-Time Bidding optimization method - on the quality of CTR estimation. From the perspective of this impact, we evaluate different configurations of widely-referenced regularization techniques and compare them with a simple technique of the feature generalization. On the basis of the results of the extensive experimentation, we show that in the context of the application scenario investigated herein, an optimization of the stochastic gradient descent algorithm configuration may be successfully accompanied, or even replaced, by a simple feature generalization.

Evaluation of Tensor-Based Algorithms for Real-Time Bidding Optimization

In this paper we evaluate tensor-based approaches to the Real-Time Bidding (RTB) Click-Through Rate (CTR) estimation problem. We propose two new tensor-based CTR prediction algorithms. We analyze the evaluation results collected from several papers – obtained with the use of the iPinYou contest dataset and the Area Underneath the ROC curve measure. We accompany these results with analogical results of our experiments – conducted with the use of our implementations of tensor-based algorithms and approaches based on the logistic regression. In contrast to the results of other authors, we show that biases – in particular those being low-order expectation value estimates – are at least as useful as outcomes of high-order components’ processing. Moreover, on the basis of Average Precision results, we postulate that ROC curve should not be the only characteristic used to evaluate RTB CTR estimation performance.

On efficient link recommendation in social networks using actor-fact matrices

Link recommendation is a popular research subject in the field of social network analysis and mining. Often, the main emphasis is put on the development of new recommendation algorithms, semantic enhancements to existing solutions, design of new similarity measures, and so forth. However, relatively little scientific attention has been paid to the impact that various data representation models have on the performance of recommendation algorithms. And by performance we do not mean the time or memory efficiency of algorithms, but the precision and recall of recommender systems. Our recent findings unanimously show that the choice of network representation model has an important and measurable impact on the quality of recommendations. In this paper we argue that the computation quality of link recommendation algorithms depends significantly on the social network representation and we advocate the use of actor-fact matrix as the best alternative. We verify our findings using several state-of-the-art link recommendation algorithms, such as SVD, RSVD, and RRI using both single-relation and multirelation dataset.

SPARQL --- Compliant Semantic Search Engine with an Intuitive User Interface

It is crucial to enable users of Linked Data to explore RDF-compliant knowledge bases in an intuitive and effective way. It is not reasonable to assume that a regular user posses any knowledge about the SPARQL nor about the ontology of the given knowledge base. This paper presents the Semantic Focused Crawler SFC system which features a graph-based querying interface that address this issue. As a result of the use of auto-complete recommendations within the SFC query builder interface, the user benefits from using the ontology irrespectively from the degree of the knowledge about semantic technologies he/she possesses. When compared to several widely-referenced alternative solutions in experiments performed with the use of 2011 QALD workshop questions, the presented system appears as achieving high query results accuracy and low complexity of the query formulation process.

Tensor-Based Ontology Data Processing for Semantic Service Matchmaking

In this paper, we present a new application of multilinear data processing to Semantic Web Service matchmaking that is based on the Covariance-Matrix-based Filtering (CMF) algorithm and ontology data representation. We show advisability of integrated algebraic modeling of lexical data derived from web service descriptions and the corresponding ontology-based semantic data. The experimental evaluation results indicate superiority of the covariance-based tensor filtering method over other state-of-the-art tensor processing methods, as well as the advantages of using the proposed ontology data representation.