- Balázs Hidasi and Domonkos Tikk, General factorization framework for context-aware recommendations, Journal of Data Mining and Knowledge Discovery, Springer, May 2015. [PDF]
The paper proposes the General Factorization Framework (GFF), a single, flexible algorithm that takes the preference model (i.e. the expression which is used to estimate preferences) as an input and computes latent feature matrices for the input dimensions. GFF allows us to easily experiment with various linear models on any context-aware recommendation task, be it explicit or implicit feedback based. GFF opens up a new research path in preference modeling under context. Even though the number of possible models exponentially increases as more context dimensions are used, context-aware factorization related research generally uses either the N-way or the pairwise interaction model. While these models are good ones to start with, they are not necessarily the best fit for the context-aware recommendation problem. For example the traditional models are symmetric, while recommendations have two distinguished dimensions: the users and the items. The lack of experimentation in this field is due to the lack of flexible tools like GFF. We propose and with the help of GFF we examine several novel preference models on a four dimensional context-aware recommendation problem. The results confirm that some of these models are better fit for the problem than the traditional ones used widely in the literature.
- Balázs Hidasi and Domonkos Tikk, Speeding up ALS learning via approximate methods for context-aware recommendations, Journal of Knowledge and Information Systems, Springer, July 2015. [PDF]
The training time and scalability of algorithms is a key aspect in practice, because faster training allows for (1) capturing a more recent state of the system we model; (2) more frequent retraining of the algorithm; (3) finding better trade-offs between training time and accuracy (e.g. by using more features or epochs); (4) more efficient use of computational resources. The ALS learning on implicit feedback context-aware recommendation problems scales linearly with the number of transactions, which makes it viable for practical use. However it scales cubically (quadratically in practice) with the number of features, which makes the use of (the more accurate) high factor models inefficient. The paper presents two approximate strategies for speeding up ALS for the implicit feedback based context-aware recommendation problem. ALS-CD is based on coordinate descent, while ALS-CG builds on the conjugate gradient. With careful implementation both methods can achieve a practically linear scaling with the number of features. This results in up to 10 times speedup in realistic settings. Despite the approximation, the accuracy is very similar to the base ALS. The paper thoroughly compares the proposed methods with each other and with ALS w.r.t. different properties and concludes that ALS-CG is a very good choice for a wide range of problems.
- Balázs Hidasi, Factorization models for context-aware recommendations, Infocommunications Journal, 6(4), pp. 27-34, December 2014. [PDF]
The paper presents an alternative version of iTALS, a context-aware tensor factorization method for the implicit feedback problem. The proposed method, coined iTALSx, uses the pairwise interaction model instead of the N-way model. Otherwise it is very similar to iTALS. However the change in model requires different computational steps in order to do the training efficiently.
The other main contribution of the paper is the comparison of iTALS and iTALSx, i.e. the comparison of the N-way and the pairwise model for the implicit feedback based recommendation problem. We show that a certain model is beneficial for tasks with certain characteristics. For example the N-way model has a higher representational capacity but is more susceptible to noise. Therefore it is better to use it when the data is denser and it is better to use with higher number of factors. On the other hand, the pairwise model is generally more accurate with sparser data and can be still used accurately with low number of factors.
Stream-based, real-time Recommendation:
- Benjamin Kille, Andreas Lommatzsch, Roberto Turrin, András Serény, Martha Larson, Torben Brodt, Jonas Seiler, and Frank Hopfgartner, Stream-based recommendations: Online and offline evaluation as a service, in Mothe, J. et al., editors, Experimental IR Meets Multilinguality, Multimodality, and Interaction (CLEF 2015), pp. 497-517, Springer, 2015. [PDF]
This work presents the NewsREEL 2015 challenge, discussing the main objectives, the data set (refer also to D2.6, “Third Reference Framework Release and Evaluation Report”, Section 4 “Data set used in Reference Framework”) and the metrics introduced to benchmark the recommender algorithms.
The work mainly addresses the real-time objective, a challenging task in the News domain as the set of potentially relevant news items changes continuously, the relevance of news is highly influenced by the context, and the computing times have very tight constraints, both training and recommendation response time. For instance, in the settings of the NewsREEL competition, the response time has to be limited to 200 milliseconds at most. In addition, the paper introduces the Idomaar framework and explains how the framework supports the evaluation and optimization of recommender systems with respect to technical and precision oriented metrics.
- Andreas Lommatzsch and Sahin Albayrak, Real-time recommendations for user-item streams, in Proceedings of the 30th Annual ACM Symposium on Applied Computing, pp. 1039-1046, 2015. [PDF]
Streams manifest as users and items interact in a continuous environment. In contrast to preference matrices, systems cannot apply sophisticated recommendation algorithms in this setting. Streams connect dynamic collections as the sets of users and items continue to change. This induces additional requirements with respect to scalability, robustness, and response times. The work introduces a set of recommendation algorithms including collaborative filtering, content-based filtering, as well as baselines using popularity, recency, or combinations thereof. The evaluation applies those methods to a data set collected as part of the CLEF NewsREEL competition. Results indicated a strong dependence on contextual factors. None of the algorithms performed superior to the rest. Instead, ensembles of recommendation algorithms yielded the best results.
The work relates to the problem of stream-based recommender systems and investigates contextual aspects.
Context and Interaction :
Roberto Turrin, Andrea Condorelli, Roberto Pagano, Massimo Quadrana, and Paolo Cremonesi, Large scale music recommendation, in Proceedings of the Large Scale Recommender System workshop (LSRS 2015) at the 9th ACM Conference on Recommender Systems (RecSys 2015), 2015. [PDF]
This work experiments with three different recommendation solutions for recommending the “next song” given the current user listening session. This work treats the session as a proxy to the actual user context, and, as such, pushes forward the state-of-the-art in session-driven recommendation. The work exploits the “30M music and playlist dataset”, a music dataset extracted from last.fm and formatted according to the Idomaar format, as described in D2.6 “Third Reference Framework Release and Evaluation Report”, Section 4 “Data sets used in Reference Framework”. Two baseline algorithms—namely SAGH (same artist greatest hits) and CAGH (collocated artist greatist hits)—are compared with IPR (Implicit Playlist recommender) which analyses the user consumption patterns of the community of users to generate ”popular” listening paths and recommend tracks on the basis of the placement of a user within these paths.
The work uses several information sources, in particular—with references to the RICHeS paradigm—the context as driven by the recent user interaction, which is considered to drive the recommendation at real-time.
- Benjamin Kille, Fabian Abel, Balázs Hidasi, and Sahin Albayrak Using Interaction Signals for Job Recommendations (to appear).
In this work, we focused on the problem of reciprocal recommendations and job recommendations in particular. When recommending jobs to a user, the recommender system actually needs to consider both (1) the requirements and interests that the user has regarding her future job position and (2) the requirements that the recruiter and her company has with respect to the candidate. Job recommender systems therefore depend on accurate feedback to improve recommendation quality (see e.g. XB01 in Section 3, Deliverable 2.5). Implicit feedback may arise in terms of clicks, bookmarks and replies (e.g. bookmarking a job advertisement). As part of this showcase, we present results from a large-scale analysis on XING. We analyse different types of information sources and study correlations between explicit ratings and implicit signals to detect situations where members liked their suggestions. With reference to the RICHeS paradigm, we thus study the incorporation of interaction with the user and relevance feedback. Results show that there are correlations between explicit and implicit user signals. In particular, we see that replies and bookmarks reflect explicit preferences much better than clicks.