Abstract:Recommender systems (RS) play a pivotal role in establishing user trust by suggesting relevant items that meet their needs and enhance reliability. The recent trend involves integrating deep learning models into RS, enabling the creation of intricate, precise, and application-specific models. However, to ensure the dependability of these deep RS models, it is crucial to implement explainable systems that illuminate the internal mechanisms of the models and enhance user satisfaction. Additionally, in specific domains such as medicine and law, users require a comprehension of the model's operational processes. This is where explainable deep neural networks (xDNN) come into play, aiding in the analysis and refinement of complex models, ultimately resulting in increased user satisfaction. Furthermore, there are several challenges in current explainable models, such as non-general models, the need for pretraining models, constraints on input variations, and compromising accuracy. To tackle the challenge of comprehending which neurons in specific layers contribute to particular outputs, this paper introduces a novel approach known as xDNN based on shadow models (xDNN-SM) to maintain interpretability without sacrificing accuracy. xDNN-SM incorporates inputs, hidden layers, and outputs to reveal the relationship between inputs and their corresponding outputs through the utilization of shadow models and shadow inputs. A shadow model is a basic component of the network, and its inputs are determined by its location within the network and its connections. Shadow inputs are computed based on real inputs to enhance the explainability capabilities of the proposed model. By leveraging shadow models and inputs, the primary model becomes explainable, allowing for traceability along various paths. The xDNN-SM approach categorizes the latent feature vectors of each shadow model separately, determining the cluster to which test inputs belong. Each shadow model possesses distinct input features, making the role of each model quantifiable. Consequently, the connections between inputs and outputs become traceable. Furthermore, the proposed model identifies similar samples for each test instance through both local and global interpretation approaches. The effectiveness of xDNN-SM is demonstrated on benchmark datasets, specifically KDDcup 2012 and Yelp, highlighting the explanatory capabilities of the proposed model.

The explainable structure of deep neural network for recommendation systems

Which Neural Network Makes More Explainable Decisions? an Approach Towards Measuring Explainability

A Neural Network Based Explainable Recommender System

Dynamic Explainable Recommendation Based on Neural Attentive Models.

A Reinforcement Learning Framework for Explainable Recommendation

Exploiting Structural and Temporal Influence for Dynamic Social-Aware Recommendation

Explainable Recommendation Through Attentive Multi-View Learning.

SDNN: Symmetric deep neural networks with lateral connections for recommender systems

A Deep Neural Network Model for Rating Prediction Based on Multi-layer Prediction and Multi-granularity Latent Feature Vectors.

Hierarchical Attention based Neural Network for Explainable Recommendation.

Explaining Deep Neural Networks by Leveraging Intrinsic Methods

Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications

Learning to Rank Rationales for Explainable Recommendation

Visually Explainable Recommendation

A Reusable Model-agnostic Framework for Faithfully Explainable Recommendation and System Scrutability

Knowledge Enhanced Graph Neural Networks for Explainable Recommendation

On the Relationship Between Counterfactual Explainer and Recommender

Distilling Structured Knowledge into Embeddings for Explainable and Accurate Recommendation

Embedding deep networks into visual explanations

A Systematic Review of Deep Knowledge Graph-Based Recommender Systems, with Focus on Explainable Embeddings