Multi-Tasking with Joint Semantic Spaces for Large-Scale Music Annotation and Retrieval

Weston, Jason; Bengio, Samy; Hamel, Philippe

doi:10.1080/09298215.2011.603834

Cited by 28 publications

(23 citation statements)

References 26 publications

(29 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Another interesting use of factorizations is presented in (Weston et al, 2010(Weston et al, , 2011. The target is information retrieval, and so the aim was to optimize the ranking of labels attached to queries (images or music).…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Mapping preferences into Euclidean space

Luaces

Díez

Joachims

et al. 2015

Expert Systems with Applications

View full text Add to dashboard Cite

Understanding and modeling human preferences is one of the key problems in applications ranging from marketing to automated recommendation. In this paper, we focus on learning and analyzing the preferences of consumers regarding food products. In particular, we explore machine learning methods that embed consumers and products in an Euclidean space such that their relationship to each other models consumer preferences. In addition to predicting preferences that were not explicitly stated, the Euclidean embedding enables visualization and clustering to understand the overall structure of a population of consumers and their preferences regarding the set of products.Notice that consumers' clusters are market segments, and products clusters can be seen as groups of similar items with respect to consumer tastes. We explore two types of Euclidean embedding of preferences, one based on inner products and one based on distances. Using a real world dataset about consumers of beef meat, we find that both embeddings produce more accurate models than a tensorial approach that uses a SVM to learn preferences. The * Corresponding author: Tel: +34 985 182 588Email addresses: oluaces@uniovi.es (Oscar Luaces), jdiez@uniovi.es (Jorge Díez), tj@cs.cornell.edu (Thorsten Joachims), abahamonde@uniovi.es (Antonio Bahamonde) Preprint submitted to Expert Systems with ApplicationsJuly 15, 2015 reason is that the number of parameters to learned in embeddings can be considerably lower than in the tensorial approach. Furthermore, we demonstrate that the visualization of the learned embeddings provides interesting insights into the structure of the consumer and product space, and that it provides a method for qualitatively explaining consumer preferences. Additionally, it is important to emphasize that the approach presented here is flexible enough to allow its use with different levels of knowledge about consumers or products; therefore the application field is very wide to grasp an accurate understanding of consumers' preferences.

show abstract

Section: Related Workmentioning

confidence: 99%

“…Additionally, to ensure numerical stability, following Weston et al (2010Weston et al ( , 2011, we use a parameter R (a radius) such that the size of θ is always smaller or equal than R.…”

Section: Maximum Margin Approachmentioning

confidence: 99%

Mapping preferences into Euclidean space

Luaces

Díez

Joachims

et al. 2015

Expert Systems with Applications

View full text Add to dashboard Cite

show abstract

“…Another successful factorization approach was presented in [21,22], in this case the application field is information retrieval. The aim was to learn a multilabel classifier.…”

Section: Related Workmentioning

confidence: 99%

“…Notice that this approach has been successfully used for a variety of application fields including, for instance, recommender systems [7], information retrieval [21,22] and construction of music playlists [14,2]. On the other hand, the factorization algorithm presented here can be seen as a visualization method to arrange in a common metric space the components of inputs in a sense that interactions are proportional to distances.…”

Section: Introductionmentioning

confidence: 99%

Analysis of nutrition data by means of a matrix factorization method

et al. 2015

View full text Add to dashboard Cite

We present a factorization framework to analyze the data of a regression learning task with two peculiarities. First, inputs can be split into two parts that represent semantically significant entities. Second, the performance of regressors is very low. The basic idea of the approach presented here is to try to learn the ordering relations of the target variable instead of its exact value. Each part of the input is mapped into a common Euclidean space in such a way that the distance in the common space is the representation of the interaction of both parts of the input. The factorization approach obtains reliable models from which it is possible to compute a ranking of the features according to their responsibility in the variation of the target variable. Additionally, the Euclidean representation of data provides a visualization where metric properties have a clear semantics. We illustrate the approach with a case study: the analysis of a dataset about the variations of Body Mass Index for Age of children after a Food Aid Program deployed in poor rural communities in Southern México. In this case the two parts of inputs are the vectorial representation of children and their diets. In addition to discover latent information, the mapping of inputs allows us to visualize children and diets in a common metric space.

show abstract

“…Apart from the aforementioned sources, semantics can also be obtained from multiple sources. For example, the introduction of relatedness-aware refinements with different sources of information over the concept detection task (Fink & Perona, 2003;Miotto & Lanckriet, 2012;Xiang et al, 2010) and the fusion of different high-level types of relatedness (Globerson et al, 2007;Weston, Bengio & Hamel, 2011). Nevertheless, multiple sources of relatedness may not be complementary.…”

Section: Introductionmentioning

confidence: 99%

Learning contextualized semantics from co-occurring terms via a Siamese architecture

Sandouk

Chen

2016

Neural Networks

View full text Add to dashboard Cite

One of the biggest challenges in Multimedia information retrieval and understanding is to bridge the semantic gap by properly modeling concept semantics in context. The presence of out of vocabulary (OOV) concepts exacerbates this difficulty. To address the semantic gap issues, we formulate a problem on learning contextualized semantics from descriptive terms and propose a novel Siamese architecture to model the contextualized semantics from descriptive terms. By means of pattern aggregation and probabilistic topic models, our Siamese architecture captures contextualized semantics from the co-occurring descriptive terms via unsupervised learning, which leads to a concept embedding space of the terms in context. Furthermore, the co-occurring OOV concepts can be easily represented in the learnt concept embedding space. The main properties of the concept embedding space are demonstrated via visualization. Using various settings in semantic priming, we have carried out a thorough evaluation by comparing our approach to a number of state-of-the-art methods on six annotation corpora in different domains, i.e., MagTag5K, CAL500 and Million Song Dataset in the music domain as well as Corel5K, LabelMe and SUNDatabase in the image domain. Experimental results on semantic priming suggest that our approach outperforms those state-of-the-art methods considerably in various aspects.

show abstract

Multi-Tasking with Joint Semantic Spaces for Large-Scale Music Annotation and Retrieval

Cited by 28 publications

References 26 publications

Mapping preferences into Euclidean space

Mapping preferences into Euclidean space

Analysis of nutrition data by means of a matrix factorization method

Learning contextualized semantics from co-occurring terms via a Siamese architecture

Contact Info

Product

Resources

About