A Deep Hybrid Model for Recommendation Systems

Abstract: Recommendation has been a long-standing problem in many areas ranging from e-commerce to social websites. Most current studies focus only on traditional approaches such as content-based or collaborative filtering while there are relatively fewer studies in hybrid recommender systems. Due to the latest advances of deep learning achieved in different fields including computer vision and natural language processing, deep learning has also gained much attention in Recommendation Systems. There are several studies that utilize ID embeddings of users and items to implement collaborative filtering with deep neural networks. However, such studies do not take advantage of other categorical or continuous features of inputs. In this paper, we propose a new deep neural network architecture which consists of not only ID embeddings but also auxiliary information such as features of job postings and candidates for job recommendation system which is a reciprocal recommendation system. Experimental results on the dataset from a job-site show that the proposed method improves recommendation results over deep learning models utilizing ID embeddings.

• The paper proposes a deep hybrid model that fuses matrix factorization with deep neural networks to capture both linear and nonlinear interactions.
• The methodology integrates a shared interaction layer combining ID embeddings and rich auxiliary features, leading to notable improvements in HR and NDCG.
• Experimental results on a large job dataset demonstrate that the model significantly outperforms baseline methods in recommendation accuracy.

A Deep Hybrid Model for Recommendation Systems

Overview of the Proposed Model

The paper "A Deep Hybrid Model for Recommendation Systems" introduces a novel architecture for hybrid recommendation systems, specifically targeting job recommendation scenarios where both parties' preferences need to be satisfied. The proposed model, termed DeepHybrid, combines Matrix Factorization (MF) with Deep Neural Networks (DNN) to leverage both user-item interactions and auxiliary data, enhancing the system's ability to make precise recommendations.

Matrix Factorization and Deep Neural Networks

Matrix Factorization is employed to map users and items onto a latent factor space, primarily capturing the linear interactions derived from historical user-item interactions. Conversely, the Deep Neural Network component processes both continuous and categorical features through rich nonlinear transformations, capturing complex patterns beyond those represented by MF.

Matrix Factorization (MF)

• Utilizes user and item ID embeddings.
• Computes latent vectors $p_u$ and $q_i$, representing user and item properties.
• Represents interactions via product $$\hat{y}_{ui} = \alpha_{out}(h^T(p_{u} \odot q_{i}))$$, where $\alpha_{out}$ is a nonlinear activation function ensuring richer representation.

Deep Neural Networks (DNN)

• Processes additional user and item properties, categorizing them into continuous and categorical features.
• Normalizes continuous features and embeds categorical data, providing a detailed feature representation.
• Provides greater flexibility through a multilayer perceptron (MLP) architecture that supports complex interaction modeling.

DeepHybrid Model Architecture

The DeepHybrid model fuses outputs from MF and DNN by concatenating their representations in a shared interaction layer, designed to capture both linear and nonlinear interactions. This hybrid approach is particularly advantageous for recommendations requiring nuanced handling of auxiliary data such as job titles and applicant details.

Interaction Layer

• Combines MF outputs (focused on ID interactions) with DNN outputs (encompassing extensive feature data).
• Facilitates improved learning and ranking performance by addressing both ID and content-based recommendations.

Experimental Evaluation

The paper reports experimental results using a dataset from one of Turkey's largest job sites, Kariyernet. The DeepHybrid model demonstrated substantial performance gains over benchmark models, such as eALS and NeuMF, across implicit feedback scenarios.

Key Metrics

• High improvement in Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG) over baseline models.
• DeepHybrid significantly enhances ranking capabilities, thus providing more relevant recommendations.

Analysis and Future Prospects

The model's success highlights the significance of combining content-based and collaborative filtering approaches in recommendation systems. Future directions include optimizing the contribution of auxiliary information, particularly through text-based features and more robust candidate generation methods.

In conclusion, the DeepHybrid model represents a meaningful integration of hybrid recommendation techniques, offering significant improvements for specialized applications such as job recruitment systems. The integration strategy provides a foundation that can be extended and adapted to other sectors requiring reciprocal recommendation mechanisms.