We will review some of the major results in random graphs and some of the more challenging open problems. We will cover algorithmic and structural questions. We will touch on newer models, including those related to the WWW.

Random graphs

Factorization approaches provide high accuracy in several important prediction problems, for example, recommender systems. However, applying factorization approaches to a new prediction problem is a nontrivial task and requires a lot of expert knowledge. Typically, a new model is developed, a learning algorithm is derived, and the approach has to be implemented.Factorization machines (FM) are a generic approach since they can mimic most factorization models just by feature engineering. This way, factorization machines combine the generality of feature engineering with the superiority of factorization models in estimating interactions between categorical variables of large domain. libFM is a software implementation for factorization machines that features stochastic gradient descent (SGD) and alternating least-squares (ALS) optimization, as well as Bayesian inference using Markov Chain Monto Carlo (MCMC). This article summarizes the recent research on factorization machines both in terms of modeling and learning, provides extensions for the ALS and MCMC algorithms, and describes the software tool libFM.

Factorization Machines with libFM

Tagging plays an important role in many recent websites. Recommender systems can help to suggest a user the tags he might want to use for tagging a specific item. Factorization models based on the Tucker Decomposition (TD) model have been shown to provide high quality tag recommendations outperforming other approaches like PageRank, FolkRank, collaborative filtering, etc. The problem with TD models is the cubic core tensor resulting in a cubic runtime in the factorization dimension for prediction and learning.In this paper, we present the factorization model PITF (Pairwise Interaction Tensor Factorization) which is a special case of the TD model with linear runtime both for learning and prediction. PITF explicitly models the pairwise interactions between users, items and tags. The model is learned with an adaption of the Bayesian personalized ranking (BPR) criterion which originally has been introduced for item recommendation. Empirically, we show on real world datasets that this model outperforms TD largely in runtime and even can achieve better prediction quality. Besides our lab experiments, PITF has also won the ECML/PKDD Discovery Challenge 2009 for graph-based tag recommendation.

/pdf/pairwise-interaction-tensor-factorization-for-personalized-53romkykbu.pdf

Pairwise interaction tensor factorization for personalized tag recommendation

Since their introduction in the early 1990's, automated recommender systems have revolutionized the marketing and delivery of commerce and content by providing personalized recommendations and predictions over a variety of large and complex product offerings. In this article, we review the key advances in collaborative filtering recommender systems, focusing on the evolution from research concentrated purely on algorithms to research concentrated on the rich set of questions around the user experience with the recommender. We show through examples that the embedding of the algorithm in the user experience dramatically affects the value to the user of the recommender. We argue that evaluating the user experience of a recommender requires a broader set of measures than have been commonly used, and suggest additional measures that have proven effective. Based on our analysis of the state of the field, we identify the most important open research problems, and outline key challenges slowing the advance of the state of the art, and in some cases limiting the relevance of research to real-world applications.

/pdf/recommender-systems-from-algorithms-to-user-experience-zot1ybuvdv.pdf

Recommender systems: from algorithms to user experience

Exploiting temporal context has been proved to be an effective approach to improve recommendation performance, as shown, e.g. in the Netflix Prize competition. Time-aware recommender systems (TARS) are indeed receiving increasing attention. A wide range of approaches dealing with the time dimension in user modeling and recommendation strategies have been proposed. In the literature, however, reported results and conclusions about how to incorporate and exploit time information within the recommendation processes seem to be contradictory in some cases. Aiming to clarify and address existing discrepancies, in this paper we present a comprehensive survey and analysis of the state of the art on TARS. The analysis show that meaningful divergences appear in the evaluation protocols used--metrics and methodologies. We identify a number of key conditions on offline evaluation of TARS, and based on these conditions, we provide a comprehensive classification of evaluation protocols for TARS. Moreover, we propose a methodological description framework aimed to make the evaluation process fair and reproducible. We also present an empirical study on the impact of different evaluation protocols on measuring relative performances of well-known TARS. The results obtained show that different uses of the above evaluation conditions yield to remarkably distinct performance and relative ranking values of the recommendation approaches. They reveal the need of clearly stating the evaluation conditions used to ensure comparability and reproducibility of reported results. From our analysis and experiments, we finally conclude with methodological issues a robust evaluation of TARS should take into consideration. Furthermore we provide a number of general guidelines to select proper conditions for evaluating particular TARS.

/pdf/time-aware-recommender-systems-a-comprehensive-survey-and-40qfzy0zif.pdf

Time-aware recommender systems: a comprehensive survey and analysis of existing evaluation protocols

Collaborative tagging systems are social data repositories, in which users manage resources using descriptive keywords (tags). An important element of collaborative tagging systems is the tag recommender, which proposes a set of tags to each newly posted resource. In this paper we discuss the potential role of three tag sources: resource content as well as resource and user profiles in the tag recommendation system. Our system compiles a set of resource specific tags, which includes tags related to the title and tags previously used to describe the same resource (resource profile). These tags are checked against user profile tags - a rich, but imprecise source of information about user interests. The result is a set of tags related both to the resource and user. Depending on the character of processed posts this set can be an extension of the common tag recommendation sources, namely resource title and resource profile. The system was submitted to ECML PKDD Discovery Challenge 2009 for "content-based" and "graph-based" recommendation tasks, in which it took the first and third place respectively.

/pdf/tag-sources-for-recommendation-in-collaborative-tagging-y6ag9v00jl.pdf

Tag sources for recommendation in collaborative tagging systems

Size and complexity of data repositories collaboratively created by Web users generate a need for new processing approaches. In this paper, we study the problem of detection of fine-grained communities of users in social networks, which can be defined as clustering with a large number of clusters. The practical size of social networks makes the traditional evolutionary based clustering approaches, which represent the entire clustering solution as one individual, hard to apply. We propose an Agglomerative Clustering Genetic Algorithm (ACGA): a population of clusters evolves from the initial state in which each cluster represents one user to a high quality clustering solution. Each step of the evolutionary process is performed locally, engaging only a small part of the social network limited to two clusters and their direct neighborhood. This makes the algorithm practically useful independently of the size of the network. Evaluation on two social network models indicates that ACGA is potentially able to detect communities with accuracy comparable or better than two typical centralized clustering algorithms even though ACGA works under much stricter conditions.

http://web.cs.dal.ca/~eem/cvWeb/pubs/marek-gecco-2009.pdf

Agglomerative genetic algorithm for clustering in social networks

Despite all of the advantages of tags as an easy and flexible information management approach, tagging is a cumbersome task. A set of descriptive tags has to be manually entered by users whenever they post a resource. This process can be simplified by the use of tag recommendation systems. Their objective is to suggest potentially useful tags to the user. We present a hybrid tag recommendation system together with a scalable, highly efficient system architecture. The system is able to utilize user feedback to tune its parameters to specific characteristics of the underlying tagging system and adapt the recommendation models to newly added content. The evaluation of the system on six real-life datasets demonstrated the system’s ability to combine tags from various sources (e.g., resource content or tags previously used by the user) to achieve the best quality of recommended tags. It also confirmed the importance of parameter tuning and content adaptation. A series of additional experiments allowed us to better understand the characteristics of the system and tagging datasets and to determine the potential areas for further system development.

Efficient Tag Recommendation for Real-Life Data

The objective of a tag recommendation system is to propose a set of tags for a resource to ease the tagging process done manually by a user. Tag recommendation is an interesting and well defined research problem. However, while solving it, it is easy to forget about its practical implications. We discuss the practical aspects of tag recommendation and propose a system that successfully addresses the problem of learning in tag recommendation, without sacrificing efficiency. Learning is realized in two aspects: adaptation to newly added posts and parameter tuning. The content of each added post is used to update the resource and user profiles as well as associations between tags. Parameter tuning allows the system to automatically adjust the way tag sources (e.g., content related tags or user profile tags) are combined to match the characteristics of a specific collaborative tagging system. The evaluation on data from three collaborative tagging systems confirmed the importance of both learning methods. Finally, an architecture based on text indexing makes the system efficient enough to serve in real time collaborative tagging systems with number of posts counted in millions, given limited computing resources.

/pdf/learning-in-efficient-tag-recommendation-44jfntgavy.pdf

Learning in efficient tag recommendation

Collaborative tagging systems are popular tools for organization, sharing and retrieval of web resources. Their success is due to their freedom and simplicity of use. To post a resource, the user should only define a set of tags that would position the resource in the system's data structure -- folksonomy. This data structure can serve as a rich source of information about relations between tags and concepts they represent. To make use of information collaboratively added to folksonomies, we need to understand how users make tagging decisions. Three factors that are believed to influence user tagging decisions are: the tags used by other users, the organization of user's personal repository and the knowledge model shared between users. In our work we examine the role of another potential factor -- resource title. Despite all the advantages of tags, tagging is a tedious process. To minimize the effort, users are likely to tag with keywords that are easily available. We show that resource title, as a source of useful tags, is easy to access and comprehend. Given a choice of two tags with the same meaning, users are likely to be influenced by their presence in the title. However, a factor that seems to have stronger impact on users' tagging decisions is maintaining the consistency of the personal profile of tags. The results of our study reveal a new, less idealistic picture of collaborative tagging systems, in which the collaborative aspect seems to be less important than personal gains and convenience.

http://web.cs.dal.ca/~eem/cvWeb/pubs/titleToTag-ht65-lipczak-final-2010.pdf

Marek Lipczak

Papers

Tag sources for recommendation in collaborative tagging systems

Agglomerative genetic algorithm for clustering in social networks

Efficient Tag Recommendation for Real-Life Data

Learning in efficient tag recommendation

The impact of resource title on tags in collaborative tagging systems