Recommender systems

A distributed recommender system architecture

Panagiotis Giannikopoulos and Costas Vassilakis
International Journal of Web Engineering and Technology, vol 7(3), 2012

Abstract:
In contemporary internet architectures, including server farms and blog aggregators, web log data may be scattered among multiple cooperating peers. In order to perform content personalization through provision of recommendations on such architectures, it is necessary to employ a recommendation algorithm; however the majority of such algorithms are centralized, necessitating excessive data transfers and exhibiting performance issues when the number of users or the volume of data increase. In this paper we propose an approach where the clickstream information is distributed to a number of peers, which cooperate for discovering frequent patterns and for generating recommendations, introducing (a) architectures that allow the distribution of both the content and the clickstream database to the participating peers and (b) algorithms that allow collaborative decisions on the recommendations to the users, in the presence of scattered log information. The proposed approach may be employed in various domains, including digital libraries, social data, server farms and content distribution networks.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.
AttachmentSize
PDF icon dist-recom-sys.pdf306.54 KB
Year: 

Improving museum visitors' Quality of Experience through intelligent recommendations: A visiting style-based approach

Ioanna Lykourentzou, Xavier Claude, Yannick Naudet, Eric Tobias, Angeliki Antoniou, George Lepouras and Costas Vasilakis
Proceedings of MASIE 2013 Workshop, co-located with the 9th International Conference on Intelligent Environments IE'13

Abstract:
This paper investigates the effect that smart routing and recommendations can have on improving the Quality of Experience of museum visitors. The novelty of our approach consists of taking into account not only user interests but also their visiting styles, as well as modeling the museum not as a sterile space but as a location where crowds meet and interact, impacting each visitor’s Quality of Experience. The investigation is done by an empirical study on data gathered by a custom-made simulator tailored for the museum user routing problem. Results are promising and future potential and directions are discussed.

Note:This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

AttachmentSize
PDF icon masie2013_submission_2.pdf520.75 KB
Year: 

Knowledge-Based Leisure Time Recommendations in Social Networks

Dionisis Margaris, Costas Vassilakis, and Panayiotis Georgiadis
chapter in: Current Trends on Knowledge-Based Systems: Theory and Applications, to be published at January 2017
Abstract:

We introduce a novel knowledge-based recommendation algorithm for leisure time information to be used in social networks, which enhances the state-of-the-art in this algorithm category by taking into account (a) qualitative aspects of the recommended places (restaurants, museums, tourist attractions etc.), such as price, service and atmosphere, (b) influencing factors between social network users, (c) the semantic and geographical distance between locations and (d) the semantic categorization of the places to be recommended. The combination of these features leads to more accurate and better user-targeted leisure time recommendations.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 

Enhancing Rating Prediction Quality through Improving the Accuracy of Detection of Shifts in Rating Practices

Dionisis Margaris and Costas Vassilakis
Transactions on Large-scale Data and Knowledge-Centered Systems (to appear)

Abstract:
The most widely used similarity metrics in collaborative filtering, namely the Pearson Correlation and the Adjusted Cosine Similarity, adjust each individual rating by the mean of the ratings entered by the specific user, when computing similarities, due to the fact that users follow different rating practices, in the sense that some are stricter when rating items, while others are more lenient. However, a user’s rating practices change over time, i.e. a user could start as lenient and subsequently become stricter or vice versa; hence by relying on a single mean value per user, we fail to follow such shifts in users’ rating practices, leading to decreased rating prediction accuracy. In this work, we present a novel algorithm for calculating dynamic user averages, i.e. time-in-point averages that follow shifts in users’ rating practices, and exploit them in both user-user and item-item collaborative filtering implementations. The proposed algorithm has been found to introduce significant gains in rating prediction accuracy, and outperforms other dynamic average computation approaches that are presented in the literature.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 
Research area: 

Enhancing Rating Prediction Quality through Improving the Accuracy of Detection of Shifts in Rating Practices

Dionisis Margaris and Costas Vassilakis
Transactions on Large-scale Data and Knowledge-Centered Systems, to appear

Abstract:
The most widely used similarity metrics in collaborative filtering, namely the Pearson Correlation and the Adjusted Cosine Similarity, adjust each individual rating by the mean of the ratings entered by the specific user, when computing similarities, due to the fact that users follow different rating practices, in the sense that some are stricter when rating items, while others are more lenient. However, a user’s rating practices change over time, i.e. a user could start as lenient and subsequently become stricter or vice versa; hence by relying on a single mean value per user, we fail to follow such shifts in users’ rating practices, leading to decreased rating prediction accuracy. In this work, we present a novel algorithm for calculating dynamic user averages, i.e. time-in-point averages that follow shifts in users’ rating practices, and exploit them in both user-user and item-item collaborative filtering implementations. The proposed algorithm has been found to introduce significant gains in rating prediction accuracy, and outperforms other dynamic average computation approaches that are presented in the literature.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 
Research area: 

Enhancing User Rating Database Consistency through Pruning

Dionisis Margaris and Costas Vassilakis
Transactions on Large-Scale Data- and Knowledge-Centered Systems, special issue on Consistency and Inconsistency in Data-centric Applications, Springer
Abstract:

Recommender systems are based on information about users' past behavior to formulate recommendations about their future actions. However, as time goes by the interests and likings of people may change: people listen to different singers or even different types of music, watch different types of movies, read different types of books and so on. Due to this type of changes, an amount of inconsistency is introduced in the database since a portion of it does not reflect the current preferences of the user, which is its intended purpose.
In this paper, we present a pruning technique that removes old aged user behavior data from the ratings database, which are bound to correspond to invalidated preferences of the user. Through pruning (1) inconsistencies are removed and data quality is upgraded, (2) better rating prediction generation times are achieved and (3) the ratings database size is reduced. We also propose an algorithm for determining the amount of pruning that should be performed, allowing the tuning and operation of the pruning algorithm in an unsupervised fashion.
The proposed technique is evaluated and compared against seven aging algorithms, which reduce the importance of aged ratings, and a state-of-the-art pruning algorithm, using datasets with varying characteristics. It is also validated using two distinct rating prediction computation strategies, namely collaborative filtering and matrix factorization. The proposed technique needs no extra information concerning the items' characteristics (e.g. categories that they belong to or attributes' values), can be used in all rating databases that include a timestamp and has been proved to be effective in any size of users-items database and under two rating prediction computation strategies.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 
Research area: 

Experimental results for considering Virtual Near Neighbors in Collaborative Filtering’s Rating Prediction

Dionisis Margaris, Dionysios Vasilopoulos, Costas Vassilakis, Dimitris Spiliotopoulos
SoDa Technical report TR-19001

Abstract:
In this technical report, we present the experimental findings from applying an algorithm that considers virtual near neighbors (VNNs) in the rating prediction formulation process, in order to increase coverage in the context of sparse datasets.
To this end, the algorithm is applied to seven sparse datasets, which are widely used in recommender system research. Additionally, the algorithm is applied to one dense dataset, in order to gain insight on the performance of the proposed algorithm in this class of datasets, as well.
In short, the algorithm introduces the concept of VNNs i.e. virtual users, which are created from the combination of real ones, in order to be used as candidate NNs in the rating prediction computation process.
In these experiments, the optimal values for the parameters that are used in the algorithm are investigated and more specifically, the thresholds that two individual users can constitute a VNN.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

AttachmentSize
PDF icon soda-TR-19001.pdf872.56 KB
Year: 
Research area: 

Exploiting Internet of Things Information to Enhance Venues' Recommendation Accuracy

Dionisis Margaris and Costas Vassilakis
Service Oriented Computing and Applications, Springer, 11(4), pp. 393–409, December 2017

Abstract:
In this paper, we introduce a novel recommendation algorithm, which exploits data sourced from web services provided by the Internet of Things in order to produce more accurate venue recommendations. The proposed algorithm provides added value for the web services offered by the Internet of Things and enhances the state-of-the-art in this algorithm category by taking into account (a) web of things data regarding the contexts of the user and the context of the venues to be recommended (restaurants, movie theatres, etc.), such as the user’s geographical position, road traffic and weather conditions, (b) qualitative aspects of the venues, such as price, atmosphere or service, (c) the semantic similarity of venues and (d) the influencing factors between social network users, derived from user participation in social networks. The combination of these features leads to more accurate and better user-targeted recommendations. We also present a framework which incorporates the above characteristics, and we evaluate the presented algorithm, both in terms of performance and recommendation quality.

Read the article online via Springer Nature SharedIt

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 

Exploiting Rating Abstention Intervals for Addressing Concept Drift in Social Network Recommender Systems

Dionisis Margaris and Costas Vassilakis
Informatics, 5(21): Special Issue “Advances in Recommender Systems”, 2018

Abstract:
One of the major problems that social networks face is the continuous production of successful, user-targeted information in the form of recommendations, which are produced exploiting technology from the field of recommender systems. Recommender systems are based on information about users’ past behavior to formulate recommendations about their future actions. However, as time goes by, social network users may change preferences and likings: they may like different types of clothes, listen to different singers or even different genres of music and so on. This phenomenon has been termed as concept drift. In this paper: (1) we establish that when a social network user abstains from rating submission for a long time, it is a strong indication that concept drift has occurred and (2) we present a technique that exploits the abstention interval concept, to drop from the database ratings that do not reflect the current social network user’s interests, thus improving prediction quality.

Full text

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 
Research area: 

Improving Collaborative Filtering's Rating Prediction Quality by Considering Shifts in Rating Practices

Dionisis Margaris and Costas Vassilakis
Proceedings of the 19th IEEE International Conference on business informatics (CBI17)

Abstract:
Users that populate ratings databases, such as IMDB, follow different marking practices, in the sense that some are stricter, while others are more lenient. This aspect has been captured by the most widely used similarity metrics in collaborative filtering, namely the Pearson Correlation and the Adjusted Cosine Similarity, which adjust each individual rating by the mean of the ratings entered by the specific user, when computing similarities. However, relying on the mean value presumes that the users' marking practices remain constant over time; in practice though, it is possible that a user's marking practices change over time, i.e. a user could start as strict and subsequently become lenient, or vice versa. In this work, we propose an approach to take into account marking practices shifts by (1) introducing the concept of dynamic user rating averages which follow the users' marking practices shifts, (2) presenting two alternative algorithms for computing a user's dynamic averages and (3) performing a comparative evaluation among these two algorithms and the classic static average (unique mean value) that the Pearson Correlation uses.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.
Year: 
Research area: 

Improving Collaborative Filtering's Rating Prediction Quality in Dense Datasets, by Pruning Old Ratings

Dionisis Margaris and Costas Vassilakis
Proceedings of the 22nd IEEE Symposium on Computers and Communications (ISCC17)

Abstract:
In this paper, we introduce a pruning algorithm which removes aged user ratings from the rating database used by collaborative filtering algorithms, in order to (1) improve prediction quality and (2) minimize the rating database size, as well as the rating prediction generation time. The proposed algorithm needs no extra information concerning the items' characteristics (e.g. categories that they belong to or attributes' values) and can be used with all rating databases that include a timestamp. Furthermore, we propose and validate a method for identifying the most prominent combination of a pruning algorithm and a pruning level for datasets, allowing thus to perform the selection of pruning algorithm and pruning level in an unsupervised fashion.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.
Year: 
Research area: 

Improving Collaborative Filtering’s Rating Prediction Accuracy by Considering Users’ Rating Variability

Dionisis Margaris and Costas Vassilakis
Proceedings of The Fourth IEEE International Conference on Big Data Intelligence and Computing (DataCom 2018)

Abstract:
When rating predictions are computed in user-user collaborative filtering, each individual rating is typically adjusted by the mean of the ratings entered by the specific user. This practice takes into account the fact that users follow different rating practices, in the sense that some are stricter when rating items, while others are more lenient. However, users’ rating practices may also differ in rating variability, in the sense that some user may be entering ratings close to her mean, while another user may be entering more extreme ratings, close to the limits of the rating scale. In this work, we (1) propose an algorithm that considers users’ ratings variability in the rating prediction computation process, aiming to improve rating prediction quality and (2) evaluate the proposed algorithm against seven widely used datasets considering three widely used variability measures and two user similarity metrics. The proposed algorithm, using the “mean absolute deviation around the mean” variability measure, has been found to intro-duce considerable gains in rating prediction accuracy, in every dataset and under both user similarity metrics tested.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 
Research area: 

Improving Collaborative Filtering’s Rating Prediction Coverage in Sparse Datasets by Exploiting User Dissimilarity

Dionisis Margaris and Costas Vassilakis
Proceedings of The Fourth IEEE International Conference on Big Data Intelligence and Computing (DataCom 2018)

Abstract:
Collaborative filtering systems analyze the ratings databases to identify users with similar likings and preferences, termed as near neighbors, and then generate rating predictions for a user by examining the ratings of his near neighbors for items that the user has not yet rated; based on rating predictions, recommenda-tions are then formulated. However, these systems are known to exhibit the “gray sheep” problem, i.e. the situation where no near neighbors can be identified for a number of users, and hence no recommendation can be formulated for them. This problem is more intense in sparse datasets, i.e. datasets with relatively small number of ratings, compared to the number of users and items. In this work, we propose a method for alleviating this problem by exploiting user dissimilarity, under the assumption that if some users have exhibited opposing preferences in the past, they are likely to do so in the future. The proposed method has been eval-uated against seven widely used datasets and has been proven to be particularly effective in increasing the percentage of users for which personalized recommendations can be formulated in the context of sparse datasets, while at the same time maintaining or slightly improving rating prediction quality.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 
Research area: 

Making recommendations in Social Networks based on textual reviews: a confidence-based approach (version 2.0)

Dionisis Margaris, Costas Vassilakis, Dimitris Spiliotopoulos

SoDa Technical report TR-19002v2

Note: this version extends and supersedes the first version of the report, which is available here.

Abstract:
In this technical report, we present the experimental findings from applying an algorithm that (1) considers the characteristics of Social Networks (SNs) user reviews which affect the review-to-rating conversion procedure, (2) computes a confidence level for each rating, which reflects the uncertainty level for each conversion process and (3) exploit this metric both in the users’ similarity computation and in the prediction formulation phases in recommender systems.
More specifically, we evaluate the performance of the proposed approach in terms of (i) SN users’ satisfaction and (ii) precision, regarding the recommendations formulated based on the rating predictions generated by the proposed algorithm.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

AttachmentSize
PDF icon TR-19002-v02.pdf451.64 KB
Year: 

Making recommendations in Social Networks based on textual reviews: a confidence-based approach

Dionisis Margaris, Costas Vassilakis, Dimitris Spiliotopoulos

SoDa Technical report TR-19002

Note: a newer version of this report is available here.

Abstract:
In this technical report, we present the experimental findings from applying an algorithm that (1) considers the characteristics of Social Networks (SNs) user reviews which affect the review-to-rating conversion procedure, (2) computes a confidence level for each rating, which reflects the uncertainty level for each conversion process and (3) exploit this metric both in the users’ similarity computation and in the prediction formulation phases in recommender systems.
More specifically, we evaluate the performance of the proposed approach in terms of SN users’ satisfaction regarding the recommendations formulated based on the rating predictions generated by the proposed algorithm.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

AttachmentSize
PDF icon TR-19002.pdf235.34 KB
Year: 

Query personalization using social network information and collaborative filtering techniques

Dionisis Margaris, Costas Vassilakis and Panagiotis Georgiadis
Future Generation of Computer Systems, Special Issue on Recommender Systems for Large-Scale Social Networks, 2017

Abstract:
Query personalization has emerged as a means to handle the issue of information volume growth, aiming to tailor query answer results to match the goals and interests of each user. Query personalization dynamically enhances queries, based on information regarding user preferences or other contextual information; typically enhancements relate to incorporation of conditions that filter out results that are deemed of low value to the user and/or ordering results so that data of high value are presented first. In the domain of personalization, social network information can prove valuable; users’ social networks profiles, including their interests, influence from social friends, etc. can be exploited to personalize queries. In this paper, we present a query personalization algorithm, which employs collaborative filtering techniques and takes into account influence factors between social network users, leading to personalized results that are better-targeted to the user.

Read the article online via ScienceDirect

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Recommendation Information Diffusion in Social Networks Considering User Influence and Semantics

Dionisis Margaris, Costas Vassilakis, and Panagiotis Georgiadis
Social Network Analysis and Mining, Springer, 6(1), pp. 1-22, 2016; DOI: 10.1007/s13278-016-0416-z

Abstract:
One of the major problems in the domain of social networks is the handling and diffusion of the vast, dynamic and disparate information created by its users. In this context, the information contributed by users can be exploited to generate recommendations for other users. Relevant recommender systems take into account static data from users' profiles, such as location, age or gender, complemented with dynamic aspects stemming from the user behavior and/or social network state such as user preferences, items' general acceptance and influence from social friends. In this paper, we enhance recommendation algorithms used in social networks by taking into account qualitative aspects of the recommended items, such as price and reliability, the influencing factors between social network users, the social network user behavior regarding their purchases in different item categories and the semantic categorization of the products to be recommended. The inclusion of these aspects leads to more accurate recommendations and diffusion of better user-targeted information. This allows for better exploitation of the limited recommendation space, and therefore online advertisement efficiency is raised.

Read the article online via Springer Nature SharedIt

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 

Using Time Clusters for Following Users’ Shifts in Rating Practices

Dionisis Margaris and Costas Vassilakis
Complex Systems Informatics and Modeling Quarterly, 75(13), pp. 22–42, December 2017/January 2018; DOI: https://doi.org/10.7250/csimq.2017-13.02
Abstract:

Users that enter ratings for items follow different rating practices, in the sense that, when rating items, some users are more lenient, while others are stricter. This aspect is taken into account by the most widely used similarity metric in user-user collaborative filtering, namely, the Pearson Correlation, which adjusts each individual user rating by the mean value of the ratings entered by the specific user, when computing similarities. However, a user’s rating practices change over time, i.e. a user could start as strict and subsequently become lenient or vice versa. In that sense, the practice of using a single mean value for adjusting users’ ratings is inadequate, since it fails to follow such shifts in users’ rating practices, leading to decreased rating prediction accuracy. In this work, we address this issue by using the concept of dynamic averages introduced earlier and we extend earlier work by (1) introducing the concept of rating time clusters and (2) presenting a novel algorithm for calculating dynamic user averages and exploiting them in user-user collaborative, filtering implementations. The proposed algorithm incorporates the aforementioned concept and is able to follow more successfully shifts in users’ rating practices. It has been evaluated using numerous datasets, and has been found to introduce significant gains in rating prediction accuracy, while outperforming the dynamic average computation approaches that are presented earlier.

Note: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

Year: 
Research area: