Sie sind auf Seite 1von 2

Adaptive Music News Recommendations based on Large Semantic Datasets

Till Plumbaum, Andreas Lommatzsch


Technische Universitt Berlin Berlin, Germany

Stefan Rudnitzki
Neofonie GmbH Berlin, Germany

{Firstname.Lastname}@dai-labor.de ABSTRACT
We present our recent work on recommending music news articles to users based on their music preferences and large scale semantic datasets. In todays online world, people are overwhelmed with the amount of available information. Therefore, we exploit the fact that semantically linked and structured information becomes more and more available driven by a strong research community. Our solution combines these semantic, encyclopedic knowledge sources and a large news article dataset. In a rst step we compute semantically related entities of interest, like similar artists or genres, based on a user model using graph-based algorithms. In a second step, we utilize these entities to compute best matching news article recommendations.

Stefan.Rudnitzki@neofonie.de 1.1 The Recommendation Architecture

The computation ow, as shown in Figure 1, consists of several steps: Based on the knowledge graph (entities connected by labeled edges) dierent recommendation models are created. This is done due to eciency reasons and to support several recommender algorithms. A model denes which types of entities and edges are considered and how they are weighted. Moreover, each model contains a weight normalization strategy and an algorithm dening how to connect edges. We implement both simple models and latent models, which reduce the graph complexity by aggregating similar nodes into latent classes [3, 4]. The recommender models considered for a request are chosen based on the search context. The results from the chosen recommenders are aggregated into one result set (meta-recommender [6]).

1.

THE MUSIC NEWS RECOMMENDER


semantic dataset

Our proposed music news recommender system supports users in nding interesting and up-to-date news articles about their favorite music artists and events. The knowledge base of the recommender system is a music dataset obtained from Freebase1 [1] consisting of 400,000 artists, 4,600,000 tracks and albums, and 2,000 genres, connected by 1.5 million edges. This information is linked to a news corpus, provided by Neofonie GmbH, currently containing 7,200,000 news articles with 40,000 articles added on average each day. Both datasets are combined by extracting entities from the news articles using named entity recognition methods. These entities are then linked to the Freebase dataset building a large knowledge graph, modeled as an ontology. Recommendations are computed in a user-centric way by combining a user model with the knowledge graph. In the demo, we simulate the user model by entering interests in a search eld. Explicit ratings will be added in the future. The recommender applies several semantic recommendation algorithms and creates an aggregated result set considering the search context and the provided user input.
1

recommender models

recommendations

result list

Figure 1: Computation ow diagram of our recommender The recommendations are calculated based on precomputed models. 1. We use a path-based recommender analyzing the length and the type of links between the user interests and entities relevant to the user. For this purpose we learn adequate normalization and edge weighting strategies. 2. We calculate a latent model and apply this for deriving recommendations.

http://www.freebase.com

WOMRAD 2010 Workshop on Music Recommendation and Discovery, colocated with ACM RecSys 2010 (Barcelona, SPAIN) Copyright c . This is an open-access article distributed under the terms of the Creative Commons Attribution License 3.0 Unported, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

The computed recommendations of the dierent recommendation models are aggregated into a single sorted list, ensuring the nal result list covers dierent types of entities (related artists, genres, etc.) [5]. The recommended entities are then used to select appropriate news articles nding the best matching articles. The articles selection strategy is

based on the recentness of the news and the correlation of computed entities and their occurrence in the news. As a result, a list containing the computed recommendations and news articles is presented to the user.

1.2

The Recommendation User Interface

The recommendation user interface (UI) currently consists of two pages. The rst page lets the user choose the best matching entity if the interest the user entered before is ambiguous. Therefore, a list of matching entities is presented to the user, from which he can select the right entity (depicted in Figure 2). The list shows all known entities and of what type they are, genre, actor, etc. Figure 3: Dierent types of recommendations are displayed. On the left, a list of similar artists and related news is presented. On the right, in the box, recommendations for tags and music albums is shown. to improve the recommendation accuracy. Also multilingual aspects must be considered, as proposed by the authors in [2].

Additional Authors.
Figure 2: First, based on the interest input, a list of matching entities from the knowledge graph is presented to the user where he must select the correct entity. Afterwards, the recommendations are presented to the user. The second page, see Figure 3, shows an example UI where the user has entered Madonna as an interest. As shown in Figure 3, dierent types of recommendations are presented to the user. A list of corresponding artists and related news is shown on the left side. A collection of matching keywords and albums is presented on the right side. All these entities are linked and can be selected to compute more recommendations. Ernesto William De Luca (Technische Universit at Berlin, email: ernesto.deluca@dai-labor.de), Holger D uwiger (Neofonie GmbH), email: rothand@neofonie.de and Sahin Albayrak (Technische Universit at Berlin), email: sahin.albayra k@dai-labor.de).

3.

REFERENCES

2.

CONCLUSIONS

The use of semantic information oers many new possibilities to recommend news to users. The use of linked information allows us to extend the systems knowledge about a user. For example, if the user only entered Madonna as an interest, we can add genre information (e. g. pop) and related artists. Therefore, we need signicantly less information about the music taste of the user to be able to recommend appropriate news. Next steps in our work are an enhanced extraction of semantic information from news articles using lexicon and model based entity disambiguation methods in combination with dependency parsing and a pattern based (Jape pattern2 ) fact detection approach. Furthermore, the user model has to become more sophisticated, supporting explicit user feedback and implicit feedback deduced from usage data and session information. This feedback should be incorporated in the recommendation process
2

[1] K. Bollacker, C. Evans, P. Paritosh, T. Sturge, and J. Taylor. Freebase: a collaboratively created graph database for structuring human knowledge. In SIGMOD 08: Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 12471250, New York, NY, USA, 2008. ACM. [2] E. W. De Luca, T. Plumbaum, J. Kunegis, and S. Albayrak. Multilingual ontology-based user prole enrichment. In Proc. Workshop on the Multilingual Semantic Web, 2010. [3] J. Kunegis and A. Lommatzsch. Learning spectral graph transformations for link prediction. In ICML 09: Proceedings of the 26th Annual International Conference on Machine Learning, pages 18, New York, NY, USA, 2009. ACM. [4] J. Kunegis, A. Lommatzsch, and C. Bauckhage. Alternative similarity functions for graph kernels. In Proc. Int. Conf. on Pattern Recognition, 2008. [5] J. H. Lee. Analyses of multiple evidence combination. In SIGIR 97: Proceedings of the 20th annual international ACM SIGIR conference on Research and development in information retrieval, pages 267276, New York, NY, USA, 1997. ACM Press. [6] R. Polikar. Ensemble based systems in decision making. Circuits and Systems Magazine, IEEE, 6(3):21 45, 9 2006. 1531-636X.

http://gate.ac.uk/sale/tao/

Das könnte Ihnen auch gefallen