A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems

    WWW, pp. 278-288, 2015.

    Cited by: 333|Bibtex|Views34|Links
    real worldMulti-View Deep Neural NetworkMulti-view DNNDeep LearningRecommendation SystemMore(19+)
    Wei bo:
    The Canonical Correlation Analysis model performed no better than random guess in Apps data which shows that using the non-linear mapping in Deep Structured Semantic Models plus the ranking based objective are important to the system


    Recent online services rely heavily on automatic personalization to recommend relevant content to a large number of users. This requires systems to scale promptly to accommodate the stream of new users visiting the online services for the first time. In this work, we propose a content-based recommendation system to address both the recomm...More



    • Recommendation systems and content personalization play increasingly important role in modern online web services.
    • The rest of this paper is organized as following, first we review major approaches in recommendation systems including papers that focus on the cold start problem in Section 2; in Section 3, we describe the data sets we work with and detail the type of features we use to model the user and the items in each domain, respectively.
    • In our recommendation systems setup, we set the pivot view Xu to user features and create auxiliary view for each different type of items we aim to recommend.
    • Due to the heterogeneity nature of recommendation systems, it is quite likely that the user view and item view pose different input features.
    • We have discovered that the letter tri-gram representation works ideally in the case that the input raw text is short, but becomes inappropriate to model user level features which often contain a large collection of queries and URL domains.
    • The proposed deep learning approach often needs to handle a huge amount of training examples in high dimensional feature spaces for the user view.
    • For both Apps and News data sets, we first run three sets of experiments to train singleview DNN models, each of which corresponds to a dimension reduction method in Section 6 (SV-TopK,SV-Kmeans and SV-LSH).
    • The third set of experiments train a joint model between Apps, News and Movie/TV features with TopK user features (MV-TopK w/Xbox).
    • As a head-to-head comparison between our single-view models (Type II) and traditional recommendation methods (Type I), our best model (SV-TopK) outperformed the best baselines CTR [32], which leveraged item features for recommendation, by 11% for all users (0.497 vs 0.448 MRR score), and 36.7 % for new users (0.436 vs 0.319 MRR score), relatively.
    • To find the answer to this hypothesis, we further added the Xbox data into the framework and trained a MV-DNN model with three user-item view pairs.
    • To further show the strength of our approach in modeling cross-domain users, we perform a set of experiments on the public data2 by the authors in [28].
    • We presented a general recommendation framework that uses deep learning to match rich user features to items features.
    • We want to make our DNN learning more scalable so that eventually the entire set of user features can be used for training without dimension reduction.
    • Experiments on several large-scale real-world data sets indicated that the proposed approach worked much better than other systems by large margin
    • Results indicate that our approach is significantly better than the state-of-the-art algorithms (up to 49% enhancement on existing users and 115% enhancement on new users)
    • Are eager to know: can we safely conclude that more views can indeed improve the performance of the system? To find the answer to this hypothesis, we further added the Xbox data into the framework and trained a MV-DNN model with three user-item view pairs
    • On the other hand, by comparing to the state-of-the-art algorithm, our best MV-DNN (with Xbox view) with top-K features performed 25.2% better than the CTR model for all users (from 0.277 to 0.347) , and 115% better for new users (from 0.142 to 0.306), for P@1
    • Similar results can also be observed for the News data in Table 4, where MV-DNN scored 49% better for all users and 101% better for new users than the CTR model, relatively
    • One reason is that we continue to see improved performance for all views even though the improvement becomes smaller and smaller over time
    Your rating :