Abstract:The challenge of creating domain-centric embeddings arises from the abundance of unstructured data and the scarcity of domain-specific structured data. Conventional embedding techniques often rely on either modality, limiting their applicability and efficacy. This paper introduces a novel modeling approach that leverages structured data to filter noise from unstructured data, resulting in embeddings with high precision and recall for domain-specific attribute prediction. The proposed model operates within a Hybrid Collaborative Filtering (HCF) framework, where generic entity representations are fine-tuned through relevant item prediction tasks. Our experiments, focusing on the cloud computing domain, demonstrate that HCF-based embeddings outperform AutoEncoder-based embeddings (using purely unstructured data), achieving a 28% lift in precision and an 11% lift in recall for domain-specific attribute prediction.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to generate high - quality domain - centric embeddings from abundant unstructured data and scarce domain - specific structured data in order to improve the precision and recall rate of domain - specific attribute prediction. Specifically, existing embedding techniques usually rely on single - modality data, which limits their application scope and effectiveness. This paper proposes a novel modeling method, which filters the noise in unstructured data by using structured data, thereby generating domain - specific embeddings with high precision and high recall rate. ### Problem Background In many application scenarios, such as customer targeting or behavior prediction, there is often a large amount of unstructured data, but this data may not be domain - specific. In contrast, domain - specific structured information is relatively scarce and highly sparse, and it is difficult to extract meaningful signals from it. This raises an important question: how can B2B enterprises customize machine - learning models so that they can learn from abundant unstructured data and customize them to specific business domains? ### Core Problem of the Paper To solve the above problems, the paper proposes a Hybrid Deep Collaborative Filtering (HCF) model. This model aims to extract knowledge from common unstructured data sources and business - specific structured data, and create embeddings representing item - domain - level interactions. Specifically, the HCF model achieves this goal through the following steps: 1. **First Stage: Unsupervised Pre - training** Use pre - trained language models such as BERT to convert unstructured text data into vector representations. The BERT model can capture the context information in the text and generate 768 - dimensional embedding vectors. 2. **Second Stage: Supervised Fine - tuning** Use domain - specific structured data to fine - tune the company embeddings generated in the first stage. Through matrix decomposition and multi - layer fully - connected networks, further optimize the quality of the embeddings to ensure that only useful information related to the structured data is retained. ### Experimental Results The experimental results show that the HCF model performs excellently in the technical product recommendation task in the cloud service domain. Compared with the embeddings generated only by the AutoEncoder, the HCF model has a 28% improvement in precision and an 11% improvement in recall rate. In addition, evaluated by metrics such as AUC, the HCF model outperforms other traditional methods in multiple benchmark tests. ### Summary The core contribution of this paper is to propose a hybrid model that combines unstructured and structured data, which can effectively deal with the challenges of data sparsity and computational complexity while maintaining information richness, thereby generating high - quality domain - specific embeddings. This method is not only applicable to the cloud service domain, but can also be extended to other tasks that require cross - modality data fusion.

Domain Specific Data Distillation and Multi-modal Embedding Generation

Multi-view Heterogeneous Fusion and Embedding for Categorical Attributes on Mixed Data.

User Distribution Mapping Modelling with Collaborative Filtering for Cross Domain Recommendation

Cross-domain Recommender Systems via Multimodal Domain Adaptation

Transfer Meets Hybrid: A Synthetic Approach for Cross-Domain Collaborative Filtering with Text

Cross-Domain Recommendation: An Embedding and Mapping Approach

Multi-Domain Collaborative Filtering

Deep Transfer Collaborative Filtering with Geometric Structure Preservation for Cross-Domain Recommendation

Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach

Neural Cross-Domain Collaborative Filtering with Shared Entities

Cross-Domain Feature Learning in Multimedia

Cross-Domain Federated Data Modeling on Non-IID Data

Domain Generalization with Small Data

Cross-Domain Deep Collaborative Filtering for Recommendation.

Inter-Class and Inter-Domain Semantic Augmentation for Domain Generalization

Deep Collaborative Filtering Incorporating Auxiliary Multi-Media Information.

Learning Accurate and Bidirectional Transformation Via Dynamic Embedding Transportation for Cross-Domain Recommendation

Review-Based Cross-Domain Recommendation via Hyperbolic Embedding and Hierarchy-Aware Domain Disentanglement

Modeling Embedding Dimension Correlations Via Convolutional Neural Collaborative Filtering

Cross-modal Semantic Autoencoder with Embedding Consensus

A Centralized-Distributed Transfer Model for Cross-Domain Recommendation Based on Multi-Source Heterogeneous Transfer Learning