Stelios Kapetanakis

A Hybrid CBR Approach for the Long Tail Problem in Recommender Systems

Recommender systems is an important tool to help users find relevant items to their interests in a variety of products and services including entertainment, news, research articles, and others. Recommender systems generate lists of recommendations/suggestions based on information from past user interactions, choices, demographic information as well as using machine learning and data mining. The most popular techniques for generating recommendations are through content-based and collaborative filtering with the latter used to provide user to user recommendations. However, collaborative filtering suffers from the long tail problem, i.e., it does not work correctly with items that contain a small number of ratings over large item populations with respectively large numbers of ratings. In this paper, we propose a novel approach towards addressing the long tail recommendation problem by applying Case-based Reasoning on “user history” to predict the rating of newly seen items which seem to belong to the long tail. We present a hybrid approach and a framework implemented with jCOLIBRI to evaluate it using the freely available Movielens dataset [8]. Our results seem promising and they seem to improve the existing prediction outcomes from the available literature.

Reproducibility of Experiments in Recommender Systems Evaluation

Recommender systems evaluation is usually based on predictive accuracy metrics with better scores meaning recommendations of higher quality. However, the comparison of results is becoming increasingly difficult, since there are different recommendation frameworks and different settings in the design and implementation of the experiments. Furthermore, there might be minor differences on algorithm implementation among the different frameworks. In this paper, we compare well known recommendation algorithms, using the same dataset, metrics and overall settings, the results of which point to result differences across frameworks with the exact same settings. Hence, we propose the use of standards that should be followed as guidelines to ensure the replication of experiments and the reproducibility of the results.

A Hybrid Feature Combination Method that Improves Recommendations

Recommender systems help users find relevant items efficiently based on their interests and historical interactions. They can also be beneficial to businesses by promoting the sale of products. Recommender systems can be modelled by applying different approaches, including collaborative filtering (CF), demographic filtering (DF), content-based filtering (CBF) and knowledge-based filtering (KBF). However, large amounts of data can produce recommendations that are limited in accuracy because of diversity and sparsity issues. In this paper, we propose a novel hybrid approach that combines user-user CF with the attributes of DF to indicate the nearest users, and compare the Random Forest classifier against the kNN classifier, developed through an investigation of ways to reduce the errors in rating predictions based on users past interactions. Our combined method leads to improved prediction accuracy in two different classification algorithms. The main goal of this paper is to identify the impact of DF on CF and compare the two classifiers. We apply a feature combination hybrid method that can improve prediction accuracy and achieve lower mean absolute error values compared with the results of CF or DF alone. To test our approach, we ran an offline evaluation using the 1 M MovieLens data set.

Answering with Cases: A CBR Approach to Deep Learning

Every year tenths of thousands of customer support engineers around the world deal with, and proactively solve, complex help-desk tickets. Daily, almost every customer support expert will turn his/her attention to a prioritization strategy, to achieve the best possible result. To assist with this, in this paper we describe a novel case-based reasoning application to address the tasks of: high solution accuracy and shorter prediction resolution time. We describe how appropriate cases can be generated to assist engineers and how our solution can scale over time to produce domain-specific reusable cases for similar problems. Our work is evaluated using data from 5000 cases from the automotive industry.

A Triangle Multi-level Item-Based Collaborative Filtering Method that Improves Recommendations.

One of the most successful approaches that can provide a relevant recommendation in various domains is collaborative filtering. Although this approach has been widely applied, there are still limitations to be overcome in this research area. Accuracy is still one of the areas that need to be improved. In addition, the rapid growth of information available online presents recommender systems with several challenges. More specifically, data sparsity and coverage affect the quality of the recommendations that can be provided. In this paper, we propose an item-based collaborative filtering (IBCF) approach with triangle similarity measures that take into account the length and angle of rating vectors between users and allow positive and negative adjustments using a multi-level recommendation approach. We have improved the predictive accuracy and effectiveness of the proposed method, which outperforms all the compared methods in terms of the mean absolute error (MAE) and the root mean squared error (RMSE). We aimed to evaluate the proposed method by comparing our results with those of some popular similarity measures using k-nearest neighbour (kNN) algorithms. We ran our experiment using three real dataset: MovieLens 100K, MovieLens 1M and Yahoo! Movies.

A Structural Rule-Based Approach for Design Patterns Recovery

Design patterns have a key role in the software development process. They describe both structure, behavior of classes and their relationships. Design patterns can improve software documentation, speed up the development process and enable large-scale reuse of software architectures. This paper presents a Multiple Levels Detection Approach (MLDA) to recover design pattern instances from Java source code. MLDA is able to extract design pattern instances based on a generated class level representation of an investigated system. Specifically, MLDA presents what is the so-called Structural Search Model (SSM) which incrementally builds the structure of each design pattern based on the generated source code model. Moreover, MLDA uses a rule-based approach to match the method signatures of the candidate design instances to that of the subject system. As the experiment results illustrate, MLDA is able to extract 23 design patterns with reasonable detection accuracy.

MLDA: A Multiple Levels Detection Approach for Design Patterns Recovery

Design patterns have a key role in the software development process. They describe both structure, behavior of classes and their relationships. During the maintenance phase, architects can benefit from knowing the underlying software design choices made during the implementation. Moreover, design patterns can improve software documentation, speed up the development process and enable large-scale reuse of software architectures. This paper presents a Multiple Levels Detection Approach (MLDA) to recover design pattern instances from Java source code. The novelty behind MLDA is its ability to extract design pattern instances based on a generated class level representation of an investigated system. Specifically, MLDA presents what is the so-called Structural Search Model (SSM) which incrementally builds the structure of each design pattern based on the generated source code model. As the experiment results illustrate, MLDA is able to extract 22 design patterns with reasonable detection accuracy.

Predicting Fraud in Mobile Money Transfer Using Case-Based Reasoning

This paper proposes an improved CBR approach for the identification of money transfer fraud in Mobile Money Transfer (MMT) environments. Standard CBR capability is augmented by machine learning techniques to assign parameter weights in the sample dataset and automate k-value random selection in k-NN classification to improve CBR performance. The CBR system observes users’ transaction behaviour within the MMT service and tries to detect abnormal patterns in the transaction flows. To capture user behaviour effectively, the CBR system classifies the log information into five contexts and then combines them into a single dimension, instead of using the conventional approach where the transaction amount, time dimensions or features dimension are used individually. The applicability of the proposed augmented CBR system is evaluated using simulation data. From the results, both dimensions show good performance with the context of information weighted CBR system outperforming the individual features approach.

A Blackboard Based Hybrid Multi-Agent System for Improving Classification Accuracy Using Reinforcement Learning Techniques

In this paper, a general purpose multi-agent classifier system based on the blackboard architecture using reinforcement Learning techniques is proposed for tackling complex data classification problems. A trust metric for evaluating agent’s performance and expertise based on Q-learning and employing different voting processes is formulated. Specifically, multiple heterogeneous machine learning agents, are devised to form the expertise group for the proposed Coordinated Heterogeneous Intelligent Multi-Agent Classifier System (CHIMACS). To evaluate the effectiveness of CHIMACS, a variety of benchmark problems are used, including small and high dimensional datasets with and without noise. The results from CHIMACS are compared with those of individual ML models and ensemble methods. The results indicate that CHIMACS is effective in identifying classifier agent expertise and can combine their knowledge to improve the overall prediction performance.

Predicting Dust Storms Using Hybrid Intelligence System

Global dust storm events seem to increase and become more severe year over year. Thus, dust storm event understanding in terms of causes, pre-ignition signals, generation processes, and procedures can be of great significance due to the impact they can have to the society. Dust storm behaviours is usually based on five attributes mainly. These are wind speed, pressure, temperature, humidity and surface condition. Dust storm may affect both rural and urban life conditions since they can cause significant difficulties to outdoor activities in low visibility – high degree of danger weather. However, dust storm predictions using historical storm data has not been used yet effectively. This study examines the process of predicting and identifying dust storms using past storm events through a novel combination of Bayesian networks (BNs), case-based reasoning (CBR) approach and rule based system (RBS) techniques.