Multi-Modal Adversarial Autoencoders for Recommendations of - PowerPoint PPT Presentation

TraininG towards a society of data-saVvy inforMation prOfessionals to enable open leadership INnovation Multi-Modal Adversarial Autoencoders for Recommendations of Citations and Subject Labels Florian Mai Iacopo Vagliano Ansgar Scherp Lukas Galke Kiel University, Germany UMAP ’18: 26th Conference on User Modeling, Adaptation and Personalization, July 8–11, 2018, Singapore www.moving-project.eu UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [1/26]

“Avoid using GANs, if you care for your mental health” - Alfredo Canziani UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [2/26]

Motivation ◮ Adversarial regularization improves autoencoders on images (Makhzani et al. 2015) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [3/26]

Motivation ◮ Adversarial regularization improves autoencoders on images (Makhzani et al. 2015) ◮ Autoencoders enable flexible treatment of multi-modal input (Barbieri et al. 2017) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [3/26]

Motivation ◮ Adversarial regularization improves autoencoders on images (Makhzani et al. 2015) ◮ Autoencoders enable flexible treatment of multi-modal input (Barbieri et al. 2017) Research Questions 1. Does adversarial regularization improve autoencoders for recommendation? 2. To what extent do preferable input modalities depend on task? 3. What is the effect of sparsity? UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [3/26]

Two Different Tasks Recommendations for citations (left) and subject labels (right) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [4/26]

Two Different Tasks Recommendations for citations (left) and subject labels (right) ◮ Two recommendation tasks on scientific documents ◮ Items are either citations or subject labels ◮ Assumption: test documents unknown UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [4/26]

Example ◮ A researcher is writing a new paper UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [5/26]

Example ◮ A researcher is writing a new paper ◮ the draft cites already 10 other publications ( item set ) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [5/26]

Example ◮ A researcher is writing a new paper ◮ the draft cites already 10 other publications ( item set ) ◮ “Am I missing any relevant related work?” UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [5/26]

Example ◮ A researcher is writing a new paper ◮ the draft cites already 10 other publications ( item set ) ◮ “Am I missing any relevant related work?” → Recommend citation candidates UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [5/26]

Example ◮ A researcher is writing a new paper ◮ the draft cites already 10 other publications ( item set ) ◮ “Am I missing any relevant related work?” → Recommend citation candidates Important: the draft is unseen by the system (New User) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [5/26]

Exploit Multiple Modalities Use only citations of draft? UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [6/26]

Exploit Multiple Modalities Use only citations of draft? Hmm, there is more... UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [6/26]

Exploit Multiple Modalities Use only citations of draft? Hmm, there is more... Use more data of the draft? UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [6/26]

Exploit Multiple Modalities Use only citations of draft? Hmm, there is more... Use more data of the draft? Yes. UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [6/26]

Exploit Multiple Modalities Use only citations of draft? Hmm, there is more... Use more data of the draft? Yes. → Start with title , but it could be more ( condition ) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [6/26]

Related Work ◮ Document-level citation recommendation: collaborative filtering (McNee et al. 2002), SVD (Caragea et al. 2013) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [7/26]

Related Work ◮ Document-level citation recommendation: collaborative filtering (McNee et al. 2002), SVD (Caragea et al. 2013) ◮ Subject Labeling: MLP for Multi-label classification (Galke et al. 2017), but professionals use predictions only as recommendations UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [7/26]

Approach Multi-Modal Adversarial Autoencoder ◮ Train autoencoder on item sets UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [8/26]

Approach Multi-Modal Adversarial Autoencoder ◮ Train autoencoder on item sets ◮ Supply condition to the decoder (multi-modal) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [8/26]

Approach Multi-Modal Adversarial Autoencoder ◮ Train autoencoder on item sets ◮ Supply condition to the decoder (multi-modal) ◮ Jointly train encoder to produce code indistinguishable from a sample of indepentend Gaussians (adversarial) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [8/26]

Rationale ◮ Recommendation tasks are highly sparse ◮ Good representations (Bengio, Courville, and Vincent 2012) might be helpful, e.g. smoothness a ≈ b → f ( a ) ≈ f ( b ) ◮ Enforce smoothness on the code (adversarial regularization) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [9/26]

Rationale ◮ Recommendation tasks are highly sparse ◮ Good representations (Bengio, Courville, and Vincent 2012) might be helpful, e.g. smoothness a ≈ b → f ( a ) ≈ f ( b ) ◮ Enforce smoothness on the code (adversarial regularization) ◮ Leads to more generalizable reconstruction? → RQ 1 UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [9/26]

Adversarial Autoencoders Model Overview UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [10/26]

Multi-Layer Perceptron Parts used for the Multi-Layer Perceptron (MLP) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [11/26]

Undercomplete Autoencoders Parts used for the Autoencoder (AE) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [12/26]

Adversarial Autoencoders Parts used for the Adversarial Autoencoder (AAE) UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [13/26]

Experimental Setup Close to real-world evaluation: ◮ Train and test split on the time axis → disjoint “only published resources are citable” ◮ Number of considered items is crucial (Beel et al. 2016) → pruning thresholds as controlled variable ◮ Title as additional input (as condition) vs. only item sets ◮ Datasets: PubMed for citations, Econ62k for subject labels ◮ Evaluate mean reciprocal rank (MRR) of one dropped item among the predictions. ◮ Re-run three times → 408 experiments. UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [14/26]

Time Split: PubMed UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [15/26]

Time Split: Economics UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [16/26]

Task Definition Task: Given a partial set of items x \ { i ∗ } , find the missing item i ∗ . x row of binary ratings over documents × items. c condition: documents’ title y predicted probabilities for items: p ( y | x , c ) Goal: Missing item on high rank i ∗ = arg max y UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [17/26]

Method Summary Only item sets IC Item Co-occurence (McNee et al. 2002) Only titles MLP y = MLP dec ( c ) Multi-Modal SVD Singular value decomposition (Caragea et al. 2013) AE y = MLP dec (MLP enc ( x )[ , c ]) AAE y = MLP dec (MLP enc ( x )[ , c ]) . Encoder MLP enc jointly optimized to fool discriminator MLP disc . UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [18/26]

Citation Recommendation PubMed: MRR of methods by pruning threshold on minimum item count UMAP ’18 L. Galke, F. Mai, I. Vagliano, A. Scherp Adversarial Autoencoders for Recommendations [19/26]