Collaborative Knowledge Base Embedding for Recommender Systems

Fuzheng ZhangNicholas Jing YuanDefu LianXing XieWei-Ying Ma

article2016KDD1,834 citations

Proposes the Collaborative Knowledge Base Embedding framework to tackle interaction sparsity by jointly training collaborative filtering models with structural, textual, and visual representations extracted from knowledge bases via deep learning and TransR.

Listen

Online recommendation systems heavily rely on user interaction history, but they often struggle when data is sparse or when recommending newly introduced items. While auxiliary information from knowledge bases can alleviate these issues, existing solutions typically focus on single network structures and depend on labor-intensive manual feature engineering, ignoring rich unstructured signals such as text descriptions and visual imagery.

The article develops and evaluates Collaborative Knowledge Base Embedding, a unified recommendation framework that automatically extracts semantic features from structured relational data, text summaries, and visual images, jointly learning them alongside user interaction patterns.

The authors designed three specialized embedding modules: a Bayesian translation network model to capture heterogeneous entity relationships, stacked denoising auto-encoders for text summaries, and stacked convolutional auto-encoders to extract visual representations from item images like movie posters and book covers. These representations are unified into an item latent vector and trained end-to-end with collaborative filtering using implicit feedback ranking. The framework was evaluated across two large real-world datasets: the benchmark MovieLens-1M dataset (5,883 users, 3,230 movies) and IntentBooks, a web-scale dataset derived from Bing search logs (92,564 users, 18,475 books, mapped to Microsoft's Satori knowledge base).

Key findings show that the unified framework significantly and consistently outperforms competitive baseline recommendation methods across ranking accuracy and retrieval metrics. Structural relationship data provided the largest single performance boost among auxiliary inputs, followed by text summaries, while visual features provided a smaller yet statistically meaningful improvement. Automated deep embedding techniques demonstrated clear superiority over manual feature engineering and conventional topic modeling. Furthermore, joint end-to-end training outperformed two-stage pipelines where embeddings were learned separately from user feedback.

These results demonstrate that enterprises can enhance recommendation quality and user discovery without burdensome manual feature engineering by integrating diverse knowledge base data directly into model training. The approach reduces reliance on dense user interaction histories and provides a flexible blueprint for leveraging structured and unstructured data across broader information retrieval and search domains.

Organizations operating content-rich platforms should evaluate integrating multimodal knowledge bases into their core recommendation pipelines. Initial deployments can prioritize relational and textual features for the greatest immediate return, incorporating visual models as infrastructure allows. Further research should explore expanding this architecture to broader domain contexts and assessing computational scalability in real-time production environments.

Confidence in these findings is strong given the rigorous multi-run evaluations across two distinct domains. However, practical applicability depends on the availability of a well-maintained knowledge base and reliable entity-matching pipelines, which showed an observed mapping error rate of approximately 8% in the study.

Microsoft University of Electronic Science and Technology of China collaborative filtering encoders heterogeneous networks collaborative knowledge base embedding recommender systems collaborative filtering

Zhang et al (2016).pdf

Paper: Collaborative Deep Learning for Recommender Systems, Hao Wang et al. (2014). Read this first to understand how collaborative filtering can be jointly trained with deep item-content representations, a core design principle the source extends to knowledge-graph and visual signals.
Paper: Collaborative topic modeling for recommending scientific articles, Chong Wang et al. (2011). Its collaborative topic regression model establishes the earlier method of combining item text with user feedback that the source replaces with learned text embeddings.
Paper: Translating Embeddings for Modeling Multi-relational Data, Antoine Bordes et al. (2013). TransE introduces the translation-based relational embeddings underlying the source’s knowledge-graph module, making its structural representation easier to follow.
Paper: Embedding Entities and Relations for Learning and Inference in Knowledge Bases, Bishan Yang et al. (2014). This work compares foundational knowledge-base embedding models, including TransE, and prepares readers to understand the source’s relational representation choices.

Paper: RippleNet: Propagating user preferences on the knowledge graph for recommender systems, Hongwei Wang et al. (2018). RippleNet carries knowledge-graph recommendation beyond item-side embeddings by propagating each user’s interests across graph relations to predict preferences.
Paper: KGAT: Knowledge Graph Attention Network for Recommendation, Xiang Wang et al. (2019). KGAT advances knowledge-graph recommendation with attention-based multi-hop propagation over user behaviors and item attributes, extending the source’s use of graph structure.
Paper: Graph Convolutional Neural Networks for Web-Scale Recommender Systems, Rex Ying et al. (2018). PinSage extends multimodal recommendation embeddings to web-scale graph convolution, combining visual and textual item features with large interaction neighborhoods.

Cover for Collaborative Knowledge Base Embedding for Recommender Systems

Abstract

Among different recommendation techniques, collaborative filtering usually suffer from limited performance due to the sparsity of user-item interactions. To address the issues, auxiliary information is usually used to boost the performance. Due to the rapid collection of information on the web, the knowledge base provides heterogeneous information including both structured and unstructured data with different semantics, which can be consumed by various applications. In this paper, we investigate how to leverage the heterogeneous information in a knowledge base to improve the quality of recommender systems. First, by exploiting the knowledge base, we design three components to extract items' semantic representations from structural content, textual content and visual content, respectively. To be specific, we adopt a heterogeneous network embedding method, termed as TransR, to extract items' structural representations by considering the heterogeneity of both nodes and relationships. We apply stacked denoising auto-encoders and stacked convolutional auto-encoders, which are two types of deep learning based embedding techniques, to extract items' textual representations and visual representations, respectively. Finally, we propose our final integrated framework, which is termed as Collaborative Knowledge Base Embedding (CKE), to jointly learn the latent representations in collaborative filtering as well as items' semantic representations from the knowledge base. To evaluate the performance of each embedding component as well as the whole system, we conduct extensive experiments with two real-world datasets from different scenarios. The results reveal that our approaches outperform several widely adopted state-of-the-art recommendation methods.

1. INTRODUCTION
2. PRELIMINARY
2.1 User Implicit Feedback
2.2 Knowledge Base
2.3 Problem Formulation
3. OVERVIEW
4. KNOWLEDGE BASE EMBEDDING
4.1 Structural Embedding
4.2 Textual Embedding
4.3 Visual Embedding
5. COLLABORATIVE JOINT LEARNING
6. EXPERIMENTS
6.1 Datasets Description
6.2 Evaluation Schema
6.3 Study of Structural Knowledge Usage
6.4 Study of Textual Knowledge Usage
6.5 Study of Visual Knowledge Usage
6.6 Study of the Whole Framework
7. RELATED WORK
7.1 Knowledge Base for Recommendation
7.2 Structural Knowledge Embedding
7.3 Deep Learning for Recommendation
8. CONCLUSION
9. REFERENCES

Knowls

Knowl 1 — Collaborative Knowledge Base Embedding Item Representation and Joint Ranking Model
model/method
Collaborative Knowledge Base Embedding (CKE) integrates implicit collaborative filtering with structural, textual, and visual representations extracted from a knowledge base. Given an implicit feedback matrix $\mathbf{R} \in \mathbb{R}^{m \times n}$ where $R_{ij} = 1$ indicates an observed interaction between user $i$ and item $j$ and $R_{ij} = 0$ denotes unobserved interaction, CKE models user preferences via latent user vectors $\mathbf{u}_i \in \mathbb{R}^{\text{dim}}$ and composite item latent vectors $\mathbf{e}_j \in \mathbb{R}^{\text{dim}}$ .

The composite item representation $\mathbf{e}_j$ integrates an offset vector and three semantic embeddings:

$\mathbf{e}_j = \boldsymbol{\eta}_j + \mathbf{v}_j + \mathbf{X}_{\frac{L_t}{2}, j*} + \mathbf{Z}_{\frac{L_v}{2}, j*}$

where $\boldsymbol{\eta}_j \sim \mathcal{N}(0, \lambda_I^{-1} \mathbf{I})$ is an item-specific offset latent vector, $\mathbf{v}_j$ is the entity structural embedding from Bayesian TransR, $\mathbf{X}_{\frac{L_t}{2}, j*}$ is the central hidden layer textual embedding from an $L_t$ -layer Bayesian Stacked Denoising Auto-encoder (SDAE), and $\mathbf{Z}_{\frac{L_v}{2}, j*}$ is the central hidden layer visual embedding from an $L_v$ -layer Bayesian Stacked Convolutional Auto-encoder (SCAE).

For a user $i$ and a pair of items $(j, j')$ where $R_{ij} = 1$ and $R_{ij'} = 0$ , the pairwise preference probability is parameterized as:

$p(j > j'; i \mid \theta) = \sigma\left(\mathbf{u}_i^T \mathbf{e}_j - \mathbf{u}_i^T \mathbf{e}_{j'}\right)$

where $\sigma(x) = \frac{1}{1 + e^{-x}}$ is the logistic sigmoid function, and $\theta$ denotes the model parameters.
Knowl 2 — Complete CKE Joint MAP Objective Function
equation
The parameters $\mathbf{u}, \mathbf{e}, \mathbf{r}, \mathbf{M}, \mathbf{W}, \mathbf{b}, \mathbf{Q}, \mathbf{c}$ of the Collaborative Knowledge Base Embedding (CKE) framework are learned by maximizing the joint log-posterior objective function $\mathcal{L}$ :

$\begin{aligned} \mathcal{L} &= \sum_{(i, j, j') \in \mathcal{D}} \ln \sigma\left(\mathbf{u}_i^T \mathbf{e}_j - \mathbf{u}_i^T \mathbf{e}_{j'}\right) \\ &\quad - \frac{\lambda_X}{2} \sum_l \|\sigma(\mathbf{X}_{l-1} \mathbf{W}_l + \mathbf{b}_l) - \mathbf{X}_l\|_2^2 \\ &\quad + \sum_{(v_h, r, v_t, v_{t'}) \in \mathcal{S}} \ln \sigma\left(\|\mathbf{v}_h \mathbf{M}_r + \mathbf{r} - \mathbf{v}_{t'} \mathbf{M}_r\|_2^2 - \|\mathbf{v}_h \mathbf{M}_r + \mathbf{r} - \mathbf{v}_t \mathbf{M}_r\|_2^2\right) \\ &\quad - \frac{\lambda_Z}{2} \sum_{l \notin \{\frac{L_v}{2}, \frac{L_v}{2}+1\}} \|\sigma(\mathbf{Z}_{l-1} * \mathbf{Q}_l + \mathbf{c}_l) - \mathbf{Z}_l\|_2^2 \\ &\quad - \frac{\lambda_Z}{2} \sum_{l \in \{\frac{L_v}{2}, \frac{L_v}{2}+1\}} \|\sigma(\mathbf{Z}_{l-1} \mathbf{Q}_l + \mathbf{c}_l) - \mathbf{Z}_l\|_2^2 \\ &\quad - \frac{\lambda_U}{2} \sum_i \|\mathbf{u}_i\|_2^2 - \frac{\lambda_v}{2} \sum_v \|\mathbf{v}\|_2^2 - \frac{\lambda_r}{2} \sum_r \|\mathbf{r}\|_2^2 - \frac{\lambda_M}{2} \sum_r \|\mathbf{M}_r\|_F^2 \\ &\quad - \frac{1}{2} \sum_l \left(\lambda_W \|\mathbf{W}_l\|_F^2 + \lambda_b \|\mathbf{b}_l\|_2^2\right) - \frac{1}{2} \sum_l \left(\lambda_Q \|\mathbf{Q}_l\|_F^2 + \lambda_c \|\mathbf{c}_l\|_2^2\right) \\ &\quad - \frac{\lambda_I}{2} \sum_j \|\mathbf{e}_j - \mathbf{v}_j - \mathbf{X}_{\frac{L_t}{2}, j*} - \mathbf{Z}_{\frac{L_v}{2}, j*\|_2^2 \end{aligned}$

where $\mathcal{D} = \{(i, j, j') \mid R_{ij} = 1 \land R_{ij'} = 0\}$ denotes pairwise interaction triples, $\mathcal{S}$ is the set of quadruples $(v_h, r, v_t, v_{t'})$ consisting of a valid relation triple $(v_h, r, v_t)$ and a corrupted negative triple $(v_h, r, v_{t'})$ , $\mathbf{X}_l$ and $\mathbf{Z}_l$ are layer activations for textual SDAE and visual SCAE, $*$ denotes convolution, and $\lambda_U, \lambda_I, \lambda_v, \lambda_r, \lambda_M, \lambda_X, \lambda_Z, \lambda_W, \lambda_b, \lambda_Q, \lambda_c$ are Gaussian precision hyperparameters.
Knowl 3 — Bayesian TransR for Structural Knowledge Base Embedding
model/method
Structural knowledge is represented as a heterogeneous network $G = (\mathcal{V}, \mathcal{E})$ where $\mathcal{V}$ is the entity set and $\mathcal{E}$ is the set of typed relation edges. Bayesian TransR models entities and relations in distinct semantic spaces bridged by relation-specific projection matrices. For a triple $(v_h, r, v_t)$ , entities are embedded as vectors $\mathbf{v}_h, \mathbf{v}_t \in \mathbb{R}^k$ , the relation as $\mathbf{r} \in \mathbb{R}^d$ , and the relation projection matrix as $\mathbf{M}_r \in \mathbb{R}^{k \times d}$ .

The projected entity representations in relation space are defined as:

$\mathbf{v}_h^r = \mathbf{v}_h \mathbf{M}_r, \quad \mathbf{v}_t^r = \mathbf{v}_t \mathbf{M}_r$

The dissimilarity score function for the triple is:

$f_r(v_h, v_t) = \|\mathbf{v}_h^r + \mathbf{r} - \mathbf{v}_t^r\|_2^2$

The Bayesian generative process specifies zero-mean spherical Gaussian priors:

$\mathbf{v} \sim \mathcal{N}(0, \lambda_v^{-1} \mathbf{I}), \quad \mathbf{r} \sim \mathcal{N}(0, \lambda_r^{-1} \mathbf{I}), \quad \mathbf{M}_r \sim \mathcal{N}(0, \lambda_M^{-1} \mathbf{I})$

For each quadruple $(v_h, r, v_t, v_{t'}) \in \mathcal{S}$ , where $(v_h, r, v_t)$ is a valid knowledge base triple and $(v_h, r, v_{t'})$ is a corrupted negative triple (formed by replacing $v_t$ with an entity of the same type), the pairwise ranking probability is:

$p((v_h, r, v_t, v_{t'})) = \sigma\left(f_r(v_h, v_{t'}) - f_r(v_h, v_t)\right)$

For an item entity $j$ , its embedding vector $\mathbf{v}_j$ learned via this process serves as its structural representation in the recommendation framework.
Knowl 4 — Bayesian Stacked Denoising Auto-Encoder for Textual Embedding
model/method
Textual knowledge for items (e.g., plots or book descriptions) is embedded using an $L_t$ -layer Bayesian Stacked Denoising Auto-encoder (Bayesian SDAE). The clean textual input matrix is denoted by $\mathbf{X}_{L_t}$ , where the $j$ -th row $\mathbf{X}_{L_t, j*}$ is the bag-of-words representation for item entity $j$ . A corrupted input matrix $\mathbf{X}_0$ is constructed by randomly masking entries of $\mathbf{X}_{L_t}$ to zero with noise masking level $\epsilon$ .

The encoder comprises the first $\frac{L_t}{2}$ layers mapping $\mathbf{X}_0$ to a latent representation $\mathbf{X}_{\frac{L_t}{2}}$ , and the decoder comprises the remaining $\frac{L_t}{2}$ layers reconstructing the clean input $\mathbf{X}_{L_t}$ .

The generative process for each layer $l \in \{1, \dots, L_t\}$ is:
1. Draw weight matrix $\mathbf{W}_l \sim \mathcal{N}(0, \lambda_W^{-1} \mathbf{I})$ .
2. Draw bias vector $\mathbf{b}_l \sim \mathcal{N}(0, \lambda_b^{-1} \mathbf{I})$ .
3. Draw layer output activations $\mathbf{X}_l \sim \mathcal{N}\left(\sigma(\mathbf{X}_{l-1} \mathbf{W}_l + \mathbf{b}_l), \lambda_X^{-1} \mathbf{I}\right)$ , where $\sigma$ is the element-wise sigmoid activation function.
The vector $\mathbf{X}_{\frac{L_t}{2}, j*}$ at the bottleneck layer serves as the textual semantic embedding for item $j$ .
Knowl 5 — Bayesian Stacked Convolutional Auto-Encoder for Visual Embedding
model/method
Visual knowledge of items (e.g., movie poster images or book covers) is embedded using an $L_v$ -layer Bayesian Stacked Convolutional Auto-encoder (Bayesian SCAE). The clean input $\mathbf{Z}_{L_v}$ is a 4-dimensional tensor whose $j$ -th slice $\mathbf{Z}_{L_v, j*}$ is a $3 \times 64 \times 64$ RGB pixel tensor for item $j$ . The corrupted input $\mathbf{Z}_0$ is generated by adding zero-mean Gaussian noise with standard deviation $\sigma$ to masked entries of $\mathbf{Z}_{L_v}$ .

The architecture uses convolutional layers for $l \notin \{\frac{L_v}{2}, \frac{L_v}{2}+1\}$ and fully connected layers for the central bottleneck transition $l \in \{\frac{L_v}{2}, \frac{L_v}{2}+1\}$ .

The generative process for each layer $l \in \{1, \dots, L_v\}$ is:
1. Draw weight parameter tensor/matrix $\mathbf{Q}_l \sim \mathcal{N}(0, \lambda_Q^{-1} \mathbf{I})$ .
2. Draw bias parameter $\mathbf{c}_l \sim \mathcal{N}(0, \lambda_c^{-1} \mathbf{I})$ .
3. Draw layer output activations:
  - If layer $l$ is fully connected: $\mathbf{Z}_l \sim \mathcal{N}\left(\sigma(\mathbf{Z}_{l-1} \mathbf{Q}_l + \mathbf{c}_l), \lambda_Z^{-1} \mathbf{I}\right)$ .
  - If layer $l$ is convolutional/deconvolutional: $\mathbf{Z}_l \sim \mathcal{N}\left(\sigma(\mathbf{Z}_{l-1} * \mathbf{Q}_l + \mathbf{c}_l), \lambda_Z^{-1} \mathbf{I}\right)$ , where $*$ is the convolution operator.
The row vector $\mathbf{Z}_{\frac{L_v}{2}, j*}$ from the central representation matrix $\mathbf{Z}_{\frac{L_v}{2}}$ serves as the visual embedding of item entity $j$ .
Knowl 6 — Stochastic Gradient Descent Training Procedure for CKE
algorithm
CKE parameters are jointly optimized by maximizing the log-likelihood objective function using Stochastic Gradient Descent (SGD).
Input: User implicit feedback matrix $R \in \mathbb{R}^{m \times n}$ , structural knowledge graph $G=(\mathcal{V}, \mathcal{E})$ , textual data $X_{L_t}$ , visual data $Z_{L_v}$ , hyperparameters $\lambda_U, \lambda_I, \lambda_v, \lambda_r, \lambda_M, \lambda_X, \lambda_Z, \lambda_W, \lambda_b, \lambda_Q, \lambda_c$ , learning rate $\gamma$
Output: User latent vectors $\{\mathbf{u}_i\}$ , item latent vectors $\{\mathbf{e}_j\}$
Initialize parameters $\mathbf{u}, \mathbf{e}, \mathbf{v}, \mathbf{r}, \mathbf{M}, \mathbf{W}, \mathbf{b}, \mathbf{Q}, \mathbf{c}$ randomly from their Gaussian priors
Construct corrupted textual matrix $X_0$ and corrupted visual tensor $Z_0$
repeat
    Sample a positive user-item interaction $(i, j)$ where $R_{ij} = 1$
    Sample an unobserved item $j'$ where $R_{ij'} = 0$
    Identify subset $\mathcal{S}_{j, j'} \subset \mathcal{S}$ containing quadruples involving item entities $j$ or $j'$

    Compute gradient of joint objective $\mathcal{L}$ with respect to $\mathbf{u}_i, \mathbf{e}_j, \mathbf{e}_{j'}$
    Compute gradient of $\mathcal{L}$ with respect to structural parameters $\mathbf{v}, \mathbf{r}, \mathbf{M}_r$ using quadruples in $\mathcal{S}_{j, j'}$
    Compute gradient of $\mathcal{L}$ with respect to textual SDAE parameters $\mathbf{W}_l, \mathbf{b}_l$ and activations $X_l$
    Compute gradient of $\mathcal{L}$ with respect to visual SCAE parameters $\mathbf{Q}_l, \mathbf{c}_l$ and activations $Z_l$

    Update all parameters using SGD: $\theta \leftarrow \theta + \gamma \nabla_\theta \mathcal{L}$
until convergence
return $\{\mathbf{u}_i\}$ and $\{\mathbf{e}_j\}$
Recommendations for a user $i$ are generated by sorting candidate items $j$ in descending order of the predicted score:

$\hat{y}_{ij} = \mathbf{u}_i^T \mathbf{e}_j$

Knowl 7 — Experimental Datasets and Knowledge Base Statistics

data/table

The experiments evaluate recommendation performance on two datasets linked to Microsoft's Satori knowledge base:

MovieLens-1M: Implicit feedback extracted by taking rating 5 as positive interactions and filtering users with fewer than 3 positive ratings. Movies are matched to Satori knowledge entities with a verified match precision of 92% (134 movies could not be mapped). Structural subgraphs contain 1-step relations including genre, director, writer, actors, language, country, production date, rating, nominated awards, and received awards.
IntentBooks: User book interests derived from Bing search query logs (Sep 2014 to Jun 2015) using similarity computation and classification (precision: 91.5%), filtering users with fewer than 5 book interactions. Structural subgraphs include 1-step entities: genre, author, publish date, belonged series, language, and rating.

Statistic	MovieLens-1M	IntentBooks
#user	5,883	92,564
#item	3,230	18,475
#interactions	226,101	897,871
#sk nodes	84,011	26,337
#sk edges	169,368	57,408
#sk edge types	10	6
#tk items	2,752	17,331
#vk items	2,958	16,719

In the table, #sk nodes, #sk edges, and #sk edge types indicate structural knowledge node, edge, and relation counts; #tk items and #vk items indicate the number of items possessing textual summaries and visual images, respectively.

Knowl 8 — Hyperparameter Settings for CKE on Benchmark Datasets

data/table

Optimal hyperparameters determined via validation sets (70% train / 30% test split randomly partitioned 5 times) for collaborative filtering (cf), structural knowledge embedding (sk), textual knowledge embedding (tk), and visual knowledge embedding (vk) in CKE:

Component	MovieLens-1M	IntentBooks
cf	$\text{dim}=150, \lambda_U=\lambda_I=0.0025$	$\text{dim}=100, \lambda_U=\lambda_I=0.005$
sk	$\lambda_v=\lambda_r=0.001, \lambda_M=0.01$	$\lambda_v=\lambda_r=0.001, \lambda_M=0.1$
tk	$\lambda_W=\lambda_b=0.01, \lambda_X=0.0001,$	$\lambda_W=\lambda_b=0.01, \lambda_X=0.001,$
	$\epsilon=0.2, L_t=4, N_l=300$	$\epsilon=0.1, L_t=6, N_l=200$
vk	$\lambda_Q=\lambda_c=0.01, \lambda_Z=0.0001, \sigma=3,$	$\lambda_Q=\lambda_c=0.01, \lambda_Z=0.001, \sigma=2,$
	$L_v=6, N_f=20, S_f=(5,5)$	$L_v=8, N_f=20, S_f=(5,5)$

Here, $\text{dim}$ is the latent vector dimension, $\epsilon$ is the input masking ratio, $\sigma$ is the visual Gaussian noise standard deviation, $L_t$ and $L_v$ are total network layer depths, $N_l$ is the hidden unit count in intermediate textual SDAE layers, $N_f$ is the number of convolutional filter maps, and $S_f$ is the convolutional kernel filter size.

Knowl 9 — Ablation and Empirical Performance of Individual Knowledge Base Modalities
empirical result
Evaluating individual knowledge embedding components under MAP@K and Recall@K across MovieLens-1M and IntentBooks demonstrates the following:
1. Structural Knowledge (CKE(S)): CKE(S) outperforms all structural baselines (BPRMF+TransE, PER, LIBFM(S), PRP, and BPRMF). Standard BPRMF performs worst, showing the substantial value of structural data. CKE(S) outperforms BPRMF+TransE, confirming that TransR's distinct relation projection spaces capture entity and relation heterogeneity better than TransE's single-space translation. Network embedding approaches (CKE(S) and BPRMF+TransE) outperform manual meta-path feature engineering (PER) and feature factorization (LIBFM(S)).
2. Textual Knowledge (CKE(T)): CKE(T) outperforms LIBFM(T), CMF(T), and pure collaborative filtering (BPRMF). CKE(T) also consistently surpasses Collaborative Topic Regression (CTR), demonstrating that deep representation learning via SDAE extracts semantic representations superior to topic modeling.
3. Visual Knowledge (CKE(V)): CKE(V) outperforms LIBFM(V), CMF(V), and BPRMF+SDAE(V). The clear performance margin of CKE(V) over BPRMF+SDAE(V) demonstrates that convolutional layers (SCAE) preserve spatial locality and image neighborhood relations substantially better than fully connected denoising auto-encoders.
4. Relative Modality Contribution: Across both datasets, structural knowledge delivers the largest individual performance improvement, followed by textual knowledge, while visual knowledge delivers a smaller but statistically significant improvement.
Knowl 10 — Performance Superiority of Full CKE Integration and Collaborative Joint Learning
empirical result
Empirical comparison of the complete model, CKE(STV), against partial multimodal variants and competing baselines yields the following key results:
1. Multimodal Synergy: CKE(STV) achieves the highest MAP@K and Recall@K scores across both MovieLens-1M and IntentBooks, consistently outperforming two-modality ablations: CKE(ST), CKE(SV), and CKE(TV). This confirms that structural, textual, and visual knowledge sources provide complementary semantic information.
2. Joint Learning vs. Decoupled Learning: CKE(STV) significantly outperforms BPRMF+STV (which trains structural TransR, textual SDAE, and visual SCAE independently in advance and feeds fixed embeddings into collaborative ranking). Jointly optimizing knowledge embeddings and implicit feedback ranking aligns feature extraction directly with the recommendation objective.
3. Deep Knowledge Embedding vs. Feature Engineering: CKE(STV) substantially outperforms LIBFM(STV) (which combines raw structural, textual, and visual features in a factorization machine), confirming the superiority of automated deep network embedding over raw feature combinations.

Coverage note — None was omitted; all contributed models, auto-encoder architectures, objective functions, optimization algorithms, dataset statistics, hyperparameters, and experimental ablation results are covered.

References

1.ahmet uyar, f. m. a. Evaluating search features of google knowledge graph and bing satori. Online Information Review (2015).
2.Bobadilla, J., Ortega, F., Hernando, A., and Gutiérrez, A. Recommender systems survey. Knowledge-Based Systems 46 (2013), 109–132.
3.Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., and Yakhnenko, O. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems (2013), 2787–2795.
4.Cambria, E., Schuller, B., Liu, B., Wang, H., and Havasi, C. Knowledge-based approaches to concept-level sentiment analysis. IEEE Intelligent Systems, 2 (2013), 12–14.
5.Cheekula, S. K., Kapanipathi, P., Doran, D., Jain, P., and Sheth, A. Entity recommendations using hierarchical knowledge bases.
6.Chui, M., Manyika, J., and Kuiken, S. What executives should know about open data. McKinsey Quarterly, January 2014 (2014).
7.Ciresan, D. C., Meier, U., Masci, J., Maria Gambardella, L., and Schmidhuber, J. Flexible, high performance convolutional neural networks for image classification. In IJCAI Proceedings-International Joint Conference on Artificial Intelligence, vol. 22 (2011), 1237.
8.Elkahky, A. M., Song, Y., and He, X. A multi-view deep learning approach for cross domain user modeling in recommendation systems. In Proceedings of the 24th International Conference on World Wide Web, International World Wide Web Conferences Steering Committee (2015), 278–288.
9.Freitas, A., Oliveira, J. G., OâA˘ ZRiain, S., Curry, E., and Da Silva, J. ´ C. P. Querying linked data using semantic relatedness: a vocabulary independent approach. In Natural Language Processing and Information Systems. Springer, 2011, 40–51.
10.Gantner, Z., Rendle, S., Freudenthaler, C., and Schmidt-Thieme, L. Mymedialite: A free recommender system library. In Proceedings of the fifth ACM conference on Recommender systems, ACM (2011), 305–308.
11.Grad-Gyenge, L., Filzmoser, P., and Werthner, H. Recommendations on a knowledge graph.
12.Huang, P.-S., He, X., Gao, J., Deng, L., Acero, A., and Heck, L. Learning deep structured semantic models for web search using clickthrough data. In Proceedings of the 22nd ACM international conference on Conference on information & knowledge management, ACM (2013), 2333–2338.
13.Ji, H., and Grishman, R. Knowledge base population: Successful approaches and challenges. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1, Association for Computational Linguistics (2011), 1148–1158.
14.Lian, D., Zhao, C., Xie, X., Sun, G., Chen, E., and Rui, Y. Geomf: joint geographical modeling and matrix factorization for point-of-interest recommendation. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM (2014), 831–840.
15.Lin, Y., Liu, Z., Sun, M., Liu, Y., and Zhu, X. Learning entity and relation embeddings for knowledge graph completion. In Proceedings of AAAI (2015).
16.Masci, J., Meier, U., Cire¸san, D., and Schmidhuber, J. Stacked convolutional auto-encoders for hierarchical feature extraction. In Artificial Neural Networks and Machine Learning–ICANN 2011. Springer, 2011, 52–59.
17.Nguyen, P., Tomeo, P., Di Noia, T., and Di Sciascio, E. An evaluation of simrank and personalized pagerank to build a recommender system for the web of data. In Proceedings of the 24th International Conference on World Wide Web Companion, International World Wide Web Conferences Steering Committee (2015), 1477–1482.
18.Passant, A. dbrecâA˘Tmusic recommendations using dbpedia. In ˇ The Semantic Web–ISWC 2010. Springer, 2010, 209–224.
19.Purushotham, S., Liu, Y., and Kuo, C.-C. J. Collaborative topic regression with social matrix factorization for recommendation systems. arXiv preprint arXiv:1206.4684 (2012).
20.Ren, J. S., and Xu, L. On vectorization of deep convolutional neural networks for vision tasks. In Twenty-Ninth AAAI Conference on Artificial Intelligence (2015).
21.Rendle, S. Factorization machines with libfm. ACM Transactions on Intelligent Systems and Technology (TIST) 3, 3 (2012), 57.
22.Rendle, S., Freudenthaler, C., Gantner, Z., and Schmidt-Thieme, L. Bpr: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, AUAI Press (2009), 452–461.
23.Schafer, J. B., Frankowski, D., Herlocker, J., and Sen, S. Collaborative filtering recommender systems. In The adaptive web. Springer, 2007, 291–324.
24.Scheffler, T., Schirru, R., and Lehmann, P. Matching points of interest from different social networking sites. In KI 2012: Advances in Artificial Intelligence. Springer, 2012, 245–248.
25.Tiddi, I., d’Aquin, M., and Motta, E. Using linked data traversal to label academic communities. In Proceedings of the 24th International Conference on World Wide Web Companion, International World Wide Web Conferences Steering Committee (2015), 1029–1034.
26.Van den Oord, A., Dieleman, S., and Schrauwen, B. Deep content-based music recommendation. In Advances in Neural Information Processing Systems (2013), 2643–2651.
27.Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., and Manzagol, P.-A. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. The Journal of Machine Learning Research 11 (2010), 3371–3408.
28.Wang, C., and Blei, D. M. Collaborative topic modeling for recommending scientific articles. In Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM (2011), 448–456.
29.Wang, H., Wang, N., and Yeung, D.-Y. Collaborative deep learning for recommender systems. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’15, ACM (New York, NY, USA, 2015), 1235–1244.
30.Yu, X., Ren, X., Sun, Y., Gu, Q., Sturt, B., Khandelwal, U., Norick, B., and Han, J. Personalized entity recommendation: A heterogeneous information network approach. In Proceedings of the 7th ACM international conference on Web search and data mining, ACM (2014), 283–292.
31.Zhou, K., and Zha, H. Learning binary codes for collaborative filtering. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM (2012), 498–506.

Citation

MLA

Zhang, F., et al. “Collaborative Knowledge Base Embedding for Recommender Systems”. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 353–62, https://doi.org/10.1145/2939672.2939673.

APA

Zhang, F., Yuan, N. J., Lian, D., Xie, X., & Ma, W.-Y. (2016). Collaborative Knowledge Base Embedding for Recommender Systems. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 353–362. https://doi.org/10.1145/2939672.2939673

Chicago

Zhang, F., N. J. Yuan, D. Lian, X. Xie, and W.-Y. Ma. 2016. “Collaborative Knowledge Base Embedding for Recommender Systems”. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 353–62. https://doi.org/10.1145/2939672.2939673.

Harvard

Zhang, F. et al. (2016) “Collaborative Knowledge Base Embedding for Recommender Systems”, Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, pp. 353–362. Available at: https://doi.org/10.1145/2939672.2939673.

Vancouver

1. Zhang F, Yuan NJ, Lian D, Xie X, Ma W-Y (2016) Collaborative Knowledge Base Embedding for Recommender Systems. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, pp 353–362

BibTeX

@inproceedings{Zhang_2016, series={KDD ’16}, title={Collaborative Knowledge Base Embedding for Recommender Systems}, url={http://dx.doi.org/10.1145/2939672.2939673}, DOI={10.1145/2939672.2939673}, booktitle={Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining}, publisher={ACM}, author={Zhang, Fuzheng and Yuan, Nicholas Jing and Lian, Defu and Xie, Xing and Ma, Wei-Ying}, year={2016}, month=Aug, pages={353–362}, collection={KDD ’16} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Collaborative Knowledge Base Embedding for Recommender Systems

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Collaborative Knowledge Base Embedding Item Representation and Joint Ranking Model

Knowl 2 — Complete CKE Joint MAP Objective Function

Knowl 3 — Bayesian TransR for Structural Knowledge Base Embedding

Knowl 4 — Bayesian Stacked Denoising Auto-Encoder for Textual Embedding

Knowl 5 — Bayesian Stacked Convolutional Auto-Encoder for Visual Embedding

Knowl 6 — Stochastic Gradient Descent Training Procedure for CKE