Introduction
Traditional job recommendation systems often struggle with the limitations of keyword-based matching. These "sparse" representations fail to account for vocabulary mismatches (e.g., "J2EE Developer" vs. "Software Engineer"), rely on overly broad categories, and suffer from significant metadata gaps. To address these issues, researchers at XING and the University of Hamburg developed a "More-Like-This" (MLT) component that utilizes dense vector representations to capture the semantic similarity between job postings.
Proposed Methodology
The core of the system involves a transition from traditional TF-IDF models to dense document embeddings. The researchers evaluated several architectures, including Word2Vec and Doc2VecC. The latter is a robust version of Doc2Vec that uses a corruption-based training objective to improve document-level representations.
Title-Weighted Descriptions
A key innovation of this research is the introduction of a title-weighting scheme. Recognizing that job titles carry the highest semantic signal, the authors calculated document vectors as a weighted average of word vectors. Words were weighted by their TF-IDF scores, with an additional multiplier (constant $c = 5.0$) applied to words appearing in the job title. This approach effectively anchors the document vector to the core intent of the posting, preventing it from being diluted by boilerplate text in the full description.
Experimental Evaluation
To evaluate the system, the authors created a "silver standard" dataset derived from 112,000 distinct job postings and user interactions (bookmarks and reply intentions) on the XING platform. This allowed them to measure how well the models could identify jobs that users found similar in practice.
Offline Performance
Offline tests demonstrated a dramatic improvement over the baseline. While the standard TF-IDF model achieved a Precision@10 of only 8.69%, the introduction of dense representations saw performance climb significantly.
As shown in the table above, the basic Doc2VecC model reached 64.23%. When the custom title-weighting scheme was applied, the performance further increased to a peak of 73.05% Precision@10.
Results and Discussion
Online Impact and Qualitative Shift
In a live A/B test involving over one million users, the dense representation system delivered an 8.0% relative increase in click-through rate (CTR) compared to the existing Elasticsearch-based ranking. Beyond quantitative metrics, the system exhibited a notable qualitative shift. It moved away from recommending generic roles toward more specific, research-oriented positions that better aligned with user expertise.
Implicit Geography and Efficiency
An unexpected benefit of the dense model was its ability to learn implicit geography. Recommendations were found to be 30% to 60% closer to the user's physical location than those of previous systems. This occurred because the model learned the semantic relationships between city names in titles without explicit geographic metadata. Furthermore, the system remains computationally efficient, adding less than 10% to average response times, making it highly suitable for production environments.
Conclusion
By combining dense document embeddings with a novel title-weighting heuristic, the researchers successfully built a recommender system that understands the semantic nuances of job postings. The significant gains in both offline precision and online CTR demonstrate the power of moving beyond keyword matching in the recruitment domain.
Summary updated