Joint Deep Modeling of Users and Items Using Reviews for Recommendation

• DeepCoNN uses parallel networks on a user's reviews and an item's reviews, coupling their learned representations in a shared top layer to predict ratings.The interaction layer is motivated by matrix factorization, allowing user and item latent factors to interact.Source: §1 · p. 2
• The authors identify DeepCoNN as the first neural model to jointly represent users and items from reviews, with rating prediction guiding their coupled latent features.They also describe it as scalable and suitable for online updates as new data arrives.Source: §1 · p. 2
• DeepCoNN outperforms all compared baselines in prediction accuracy on Yelp, 21 Amazon categories, and Beer.The introduction also says it outperforms state-of-the-art techniques [35] on all evaluated datasets.Source: §1 · p. 2
• DeepCoNN improves prediction more for users and items with fewer ratings, alleviating the sparsity problem by leveraging review text.Source: §1 · p. 2
• The authors report that pre-trained word embeddings capture review semantics and sentiment that can improve rating accuracy; topic-modeling competitors use traditional bag-of-words.Source: §1 · p. 2

2. METHODOLOGY

• DeepCoNN learns hidden user and item latent factors from review text to estimate users’ ratings.The model is based on two parallel CNN networks coupled at the top by a shared layer.Source: §2 · p. 2
• The two networks are trained jointly to predict ratings while minimizing prediction error.One network models user behaviors from reviews written by users, and the other models item properties from reviews written for items.Source: §2 · p. 2
• A shared layer at the top couples the parallel user and item networks.Source: §2 · p. 2

2.1 Definition and Notation

• The training set T consists of N tuples, each representing a review written by a user for an item.Source: §2.1 · p. 2
• Each training tuple (u, i, rui, wui) contains the user u, item i, rating rui, and review text wui.Source: §2.1 · p. 2
• Table 1 summarizes the mathematical notations used throughout the paper.Source: §2.1 · p. 2, Table 1 · p. 2

2.2 Architecture

• The objective measures rating-prediction error using latent factors produced by Netu and Neti.Source: §2.2 · p. 3
• Netu and Neti differ only in their inputs, so the paper details Netu’s process and applies the same process to Neti.Source: §2.2 · p. 3

2.3 Word Representation

• DeepCoNN merges all reviews written by a user into one document and represents it as a word-embedding matrix in the look-up layer to extract semantic information.The embedding function maps dictionary words to distributed vectors.Source: §2.3 · p. 3
• Each word position maps to its c-dimensional vector, and the vectors are concatenated into the matrix for the user document.The document consists of n words in total.Source: §2.3 · p. 3
• The word-vector matrix preserves word order, unlike bag-of-words representations.Source: §2.3 · p. 3

2.4 CNN Layers

• Following [7], convolution filters Kj ∈ℜc×t apply to word-vector windows of size t to produce features.Source: §2.4 · p. 3
• The model uses ReLU activation, and deep convolutional networks with ReLUs train several times faster than tanh equivalents.Source: §2.4 · p. 3
• Max pooling takes each feature map’s maximum, capturing its strongest feature and converting variable-length text results into a fixed-size vector [7].Source: §2.4 · p. 3
• Multiple filters yield varied features, whose pooled results pass through a weighted fully connected layer to produce user and item CNN outputs.Source: §2.4 · pp. 3–4

2.5 The Shared Layer

• The shared layer couples Netu and Neti by concatenating their outputs, xu and yi, into one vector because their original feature spaces may not be comparable.This maps the outputs into the same feature space.Source: §2.5 · p. 4
• A Factorization Machine estimates ratings from the concatenated vector by modeling all nested variable interactions.The paper gives the resulting training cost in Eq. 7.Source: §2.5 · p. 4
• The FM’s pairwise second-order interactions are modeled by inner products between latent vectors for variables in the concatenated representation.The interaction term is defined as an inner product summed over latent dimensions.Source: §2.5 · p. 4

2.6 Network Training

• Training minimizes Eq. 7, with the derivative of J with respect to z given by Eq. 8.Source: §2.6 · p. 4
• Gradients for parameters in other layers are computed using the differentiation chain rule.Source: §2.6 · p. 4
• The model optimizes parameters with RMSprop over shuffled minibatches, using adaptive step sizes scaled by a running average of gradient norms.λ denotes the learning rate, and ǫ is added for numerical stability.Source: §2.6 · p. 4
• Dropout is applied to the fully connected layers of both networks to prevent overfitting.Source: §2.6 · p. 4

2.7.1 Word Order Preservation

• DeepCoNN preserves word order with word embeddings and convolution operations that use text’s internal structure, unlike bag-of-words topic models.Topic models infer latent topics while ignoring word order.Source: §2.7.1 · p. 4

2.7.2 Online Learning

• DeepCoNN scales with training-data size and can be trained or updated with new data, with user and item latent factors updated independently of historical data.The paper contrasts these advantages with topic-modeling approaches, which do not benefit from them to the same extent.Source: §2.7.2 · p. 4

3.1 Datasets and Evaluation Metric

• The experiments use Yelp, Amazon [19], and Beer [18] reviews; Yelp and Amazon average fewer than six rating-review pairs per customer, and MSE is the evaluation metric.Each dataset contains more than half a million reviews, and reviews average fewer than 150 words.Source: §3.1 · p. 4

3.2 Baselines

• Baselines span rating-only MF and PMF, review-informed topic models LDA, CTR, HFT, and deep recommender CDL [35].All baselines except MF and PMF incorporate review information.Source: §3.2 · p. 5

3.3 Experimental Settings

• Each dataset is split into 80% training, 10% validation, and the remainder test data; validation performance selects hyperparameters for all models.Source: §3.3 · p. 5

3.4 Performance Evaluation

• PMF outperforms MF on Yelp, Beer, and most Amazon categories, while HFT-10 and HFT-50 outperform other topic-modeling baselines on all three datasets.The paper attributes the gains of CTR and HFT over LDA to jointly modeling ratings and reviews.Source: §3.4 · p. 6
• DeepCoNN beats the best baseline on Yelp, Beer, and Amazon, with an average improvement of 8.3%.CDL outperforms topic-modeling recommenders, but DeepCoNN surpasses the best baseline across all three datasets.Source: §3.4 · p. 6, Table 3 · p. 7

3.5 Model Analysis

• DeepCoNN performs best when both user and item reviews are modeled; TFIDF and random word representations are weaker, and factorization-machine coupling outperforms a dot product.The coupling advantage may stem from modeling both first- and second-order interactions between user and item latent factors.Source: §3.5 · p. 6, Table 4 · p. 6

3.6 The Impact of the Number of Reviews

• MSE reductions over MF are positive for all tested user and item groups, and are greater for groups with fewer ratings.The analysis covers Yelp, Beer, and Amazon Music Instruments; the results indicate that review text helps alleviate sparsity and cold-start problems.Source: §3.6 · p. 7, Figure 3 · p. 7

Related work

• Earlier review-based rating studies predicted ratings for an existing review, whereas this work predicts from a user’s review history to recommend products.Source: §4 · p. 7
• Review-based methods explored aspects, topic modeling, user/item review modeling, sentiment, and rating-aligned item topics.These approaches variously used ratings or reviews, and some did not incorporate ratings when modeling review text.Source: §4 · p. 7
• Prior review methods relied on lexical similarity and bag-of-words, overlooking semantic similarity and word order; topic-modeling approaches also face scalability and new-user/item challenges.Source: §4 · pp. 7–8
• Neural recommenders using rating matrices [26, 37, 15] or music and movie content [35] ignored review text.The content-based models learned item factors from content, while the rating-matrix models used techniques such as denoising auto-encoders and RBMs.Source: §4 · p. 8
• The multi-view model in maps user and item factors to a common space, but uses content rather than reviews and cosine similarity rather than rating-linked learning.The passage notes that this does not guarantee the learned factors help recommendation.Source: §4 · p. 8
• Unlike DeepCoNN, the deep review-based model in uses reviews to regularize matrix factorization rather than jointly modeling users and items from reviews.Its RNN models review likelihood from item factors, and its matrix-factorization component makes handling new users and items nontrivial.Source: §4 · p. 8

• DeepCoNN uses two review-driven neural networks to model user behavior and item properties, coupled by a shared layer that maps representations into a common feature space for latent-factor interaction.Source: §5 · p. 8
• DeepCoNN improved over state-of-the-art baselines by 8.5% on Yelp and 7.6% on Beer, outperformed all Amazon baselines with 8.7% average improvement, and attained 8.3% overall.Source: §5 · p. 8
• Experiments showed that each network can separately learn user or item factors from reviews while cooperation improves rating prediction; word embeddings also helped capture review semantics versus random or TF-IDF representations.Source: §5 · p. 8
• For users and items with few reviews or ratings, DeepCoNN reduced MSE more than MF, with its greatest reduction when only one review was available, supporting sparsity alleviation.Source: §5 · p. 8