A challenge-based survey of e-recruitment recommendation systems

1.1 Differences with previous recent surveys

• Earlier surveys [35] generally organized work by recommendation method; this survey instead emphasizes e-recruitment’s distinguishing difficulties and technical solutions proposed for them.The authors present this challenge-based framing as complementary to existing method-focused surveys.Source: §1 · p. 1, §1.1 · p. 3

1 INTRODUCTION

• The survey identifies eight frequently addressed challenges, reviews papers targeting them, and discusses future research directions, including challenges less investigated recently.It also presents a structured overview of the collected 123 papers, recording recommendation type, method type, and addressed challenges.Source: §1 · p. 2

1.2 Literature search methodology

• The literature search collected 515 papers, filtered them to 99, and added 24 papers selected from industry leaders and experts, yielding 123 papers.Filtering included publication date, citation-count criteria for earlier papers, and whether papers recommended actual jobs or job seekers.Source: §1.2 · p. 3

1.3 Structure of the survey

• The survey covers distinguishing system properties, identified challenges, papers not covered in challenge sections, limitations, open challenges, and future directions.Its structure assigns separate sections to the identified challenges and concludes with limitations and future directions.Source: §1.3 · p. 3

• Job seekers and positions compete, and a recommendation’s value may depend on a seeker’s relative chance of success, so competition should inform recommendations.A person can work only a few jobs, and a posting typically hires only a few people; consequently, recommendation lists may need to remain relatively small.Source: §2 · p. 4
• Recruitment recommendations are two-sided: a seeker’s application signals interest, but success depends on the recruiter making an offer and ultimately on employment.Both job seekers and employers are stakeholders in the recommendation process.Source: §2 · p. 4
• Matching must account for both suitability—fit between a seeker’s qualifications and job requirements—and preferences, which can differ for seekers and recruiters.A seeker may qualify for several jobs but prefer one company, while a recruiter may choose among equally suitable candidates based on preferences.Source: §2 · p. 4
• Recommendation inputs are multi-faceted, spanning seeker and job attributes such as skills, experience, interests, location, and organizational culture.The available data may be heterogeneous and require attention in system design.Source: §2 · p. 4
• Because recruitment affects careers and companies’ competitive edge, fairness and trustworthiness are especially important considerations.The paper notes that e-recruitment is defined as a high-risk domain in the EU AI Act proposal.Source: §2 · p. 4
• Seekers’ interaction histories are short, and jobs and seekers become inactive after employment or a hire, unlike persistent users and items in traditional recommenders.New jobs also appear and disappear frequently.Source: §2 · pp. 4–5

3 SURVEY STRUCTURED ACCORDING TO CHALLENGES FACED IN THE DEVELOPMENT OF E-RECRUITMENT RECOMMENDATION SYSTEMS

• The survey focuses on common challenges, first describing them and then reviewing proposed approaches, with visual overviews of problems and literature solutions.It also discusses papers not covered by the challenge sections.Source: §3 · p. 5

3.1 A preview of the challenges

• Data quality is challenged by noisy textual content, variable-quality facets, and implicit information that must be extracted from unstructured data.The survey highlights cleaning and preprocessing as necessary for noisy text.Source: §3.1 · p. 5
• Skill extraction matters because matching relies on seeker skills and required job skills; multilingual platforms also need to support text in multiple languages.The survey focuses on skills as a particularly important feature for matching.Source: §3.1 · p. 5
• Data sparsity and cold start are acute when seekers leave after finding work and jobs disappear or recur as new items with few interactions.The survey notes that using data beyond interactions can often alleviate cold start.Source: §3.1 · pp. 5–6
• Systems must handle heterogeneous profiles and job data, multiple interaction types, and data from external sources.Interactions can include views, clicks, applications, chats, favorites, likes, and comments.Source: §3.1 · pp. 5–6
• Recommendations need to balance seekers’ qualifications and job requirements with personal preferences, which may change over time.The survey also identifies interpretability and explainability as important in this high-stakes domain.Source: §3.1 · p. 6
• Systems may need to satisfy seekers, recruiters, and providers while balancing recommendation counts, success prospects, and false positives; bias mitigation is also essential.The survey identifies gender bias and recency bias among concerns.Source: §3.1 · pp. 6–7
• Growing data volumes create scalability demands in training and inference, including speed and storage or memory consumption.Source: §3.1 · p. 7

• Noisy resumes and job descriptions make cleaning and preprocessing important, but the survey limits its review to work explicitly focused on NLP techniques for these issues.Common preprocessing includes tokenization, stop-word removal, stemming, and lemmatization.Source: §3.2 · p. 7
• A semantic gap arises when different terms express the same concept or the same term has different meanings across resumes and job descriptions.A common reviewed approach maps skills or concepts to shared ontology nodes, using language models, NER, or NED.Source: §3.2 · pp. 7–8
• Because job seekers’ profiles and job descriptions are often free text with no structure, some e-recruitment systems extract skills using n-grams, NER, part-of-speech tagging, dictionaries or ontologies, and skill headwords.Source: §3.2 · p. 8
• Cross-lingual matching is a challenge when resumes and job descriptions use different languages; reviewed work used multilingual language models, including a Siamese architecture fine-tuned with recruiter-interaction history [80].Source: §3.2 · p. 8
• Short seeker histories and distinct job-position entities contribute to sparsity; reviewed remedies include splitting titles from company names, clustering similar positions, and adding content-similarity links to interaction graphs,,.The survey discusses only work that explicitly studies sparsity, although content-based approaches may alleviate it to some extent (e.g. [13]).Source: §3.2 · p. 8

• Seeker and job data can span different fields and data types, while behavioral signals such as clicks, applications, chats, and invites are also available.Using such behavioral data is challenging, and the survey notes that external sources can enrich the data.Source: §3.3 · p. 8
• Reviewed methods combine similarity scores across corresponding resume and job-posting fields or all fields,, while others learn field- and data-source-specific embeddings whose interactions support matching [141].Source: §3.3 · p. 9
• External information such as social-network friends’ features and personal websites has been used to build more comprehensive profiles and improve recommendations,.Personality and traits are also identified as potentially useful beyond skills.Source: §3.3 · p. 9

3.4 Cold start

• Job openings can be treated as new items even when their titles match, so e-recruitment can face both job-seeker and job cold starts.The section describes cold start as especially acute when users or items have no or few interactions.Source: §3.4 · p. 10
• Other approaches use prior interactions and content-derived features, while content can alleviate cold start only to some extent.The review focuses on studies that explicitly address cold start; some methods derive job-seeker features from previously interacted-with jobs.Source: §3.4 · pp. 10–11
• Interaction-based approaches find similar job seekers or jobs and transfer their known matching scores to the new entity.Examples include recommending to new graduates from similar graduates’ job offers and using similar job seekers’ interaction histories.Source: §3.4 · p. 10
• Feature-based methods predict matches or interested job categories from job-seeker and job content or attributes.Some content-based approaches address cold start for both job seekers and jobs.Source: §3.4 · p. 10

3.5 User preferences as well as suitability

• Suitability matches job requirements to job-seeker skills, while preferences may come from profile factors or behavioral interactions.The review focuses on studies that explicitly model preferences, although preferences matter more broadly in recommendation.Source: §3.5 · p. 11
• Because preferences may change over time, reviewed methods model temporal interactions with recurrent networks, time-weighted scoring, matrix factorization, or sequence models.A person-job fit model also maps heterogeneous dynamic preferences from different interaction types into a unified preference space.Source: §3.5 · pp. 11–12
• Preference signals include clicks, applications, invites, profile interests or location, and explicit dashboard settings.Examples include a vacancy-exploration dashboard and fuzzy matching based on job seekers’ profile preferences.Source: §3.5 · p. 11
• Some systems learn preference models from explicit feedback or combine implicit feedback with group-level preference matrices.One matching model forms job-seeker and employer preference vectors for Gale-Shapley stable matching; another uses matrices generated from historical data.Source: §3.5 · p. 11

3.6 Interpretability and explainability

• Explanations matter because recommendations affect career choices and can help people make more insightful decisions.The section uses interpretability and explainability interchangeably.Source: §3.6 · p. 12
• Deep-model explanations can visualize attention over resume and job-description text, while other routes use interpretable models or explicit data relations.Examples include human-readable decision-tree features, skill-affinity dashboards, knowledge-graph templates, and skill-ontology similarity.Source: §3.6 · p. 13

3.7 Specific objectives

• Specific objectives reflect e-recruitment’s multi-stakeholder nature, including reciprocal benefits for job seekers and employers.Some studies also account for a market setting where each seeker is hired for one or a few positions and vice versa.Source: §3.7 · pp. 13–14
• A stable matching method addresses reciprocal recommendations, while application redistribution targets postings with too many or too few applications.The redistribution approach boosts or penalizes recommendation scores using a forecast of application counts.Source: §3.7 · p. 14
• Another specific objective is preventing job seekers from receiving spam; one hybrid recommendation method harshly penalizes false positives.Source: §3.7 · p. 14

3.8 Bias and fairness

• An online user study found that job seekers preferred biased recommendations, indicating that human bias should be addressed from new perspectives or with new technology.The survey notes limits of purely algorithmic debiasing and points to legal and ethical considerations.Source: §3.8 · p. 14
• Fairness concerns span job seekers and job postings, including racial or gender discrimination, popularity bias, and selection bias.The review covers fairness approaches for both sides of the recommendation process.Source: §3.8 · p. 14
• For job-seeker fairness, reviewed methods include deterministic reranking and a neural fair collaborative filtering model that removes gender components from embeddings.The debiased embeddings are used in fine-tuning so sensitive attributes do not affect recommendation outputs.Source: §3.8 · p. 15
• For job-posting fairness, one method treats recency bias as selection bias and uses inverse propensity weighting in an unbiased loss.Source: §3.8 · p. 15

3.9 Large scale

• CareerBuilder used a job-similarity graph rather than user-based or user-item graphs to address training-time execution and storage constraints.The graph’s edges represent job similarities based on behavioral and content-based signals.Source: §3.9 · p. 15
• Two-stage systems first retrieve candidates cheaply, then rerank them with a more expensive model, as in CareerBuilder’s architecture [141].Its first stage selects hundreds of candidates from millions using FAISS nearest-neighbor search in an embedding space.Source: §3.9 · p. 16
• LinkedIn’s CasMoS learns the candidate-selection stage of a two-stage recommendation framework using the WAND query operator.A separate LinkedIn study designed a parallel block-wise coordinate-descent algorithm to learn GLMix for predicting user response.Source: §3.9 · pp. 15–16
• Other large-scale approaches include Spark for content-based recommendation and clustering jobs or job seekers.The survey frames these as approaches to scalability in training and inference.Source: §3.9 · pp. 15–16

3.10 Papers not included in previous sections

• The survey omits some papers on specific issues, including missing features and group-specific recommendation, because few address them or they were judged less practically significant.It also notes LinkedIn practice discussed in two talks [54].Source: §3.10 · p. 16

4.1 Summary

• Cold start merits particular attention because different job entities can share a title or content, while e-recruitment offers many facets that could help alleviate the problem.The survey also notes that job seekers may be employed in only one or a few positions, creating distinct recommendation objectives.Source: §4.1 · p. 17
• E-recruitment recommendations must match skills and job requirements while accounting for user preferences, diverse data and interactions, and the task’s multiple stakeholders.The survey identifies eight challenges, including data quality and heterogeneity, cold start, explainability, fairness, objectives, and scale.Source: §4.1 · pp. 16–17
• Explainability and fairness matter because recommendations can influence careers and employers, while real-world systems must serve large populations.The survey also identifies large scale as a design challenge for systems serving thousands or millions of job seekers and positions.Source: §4.1 · p. 17

4.2 Limitations of the survey

• The authors selected the main challenges from their perspective, acknowledge that others may exist, and excluded papers from other reciprocal-recommendation domains to limit scope.Textual feature extraction at different granularities is one possible additional challenge, though not specific to e-recruitment.Source: §4.2 · p. 17

4.3 Open challenges and future research directions

• One-worker-one-job constraints and fairness remain important directions, with fairness complicated by reciprocal recommendations and multiple stakeholders.The authors also connect recommended jobs to career choices and possible long-term labor-market effects.Source: §4.3 · pp. 17–19
• Future work could investigate career-path recommendation, domain adaptation across sectors or settings, and better multilingual matching of resumes and job postings.Career-path systems recommend intermediate jobs toward a desired career goal; domain adaptation has seen little e-recruitment study beyond a few papers such as.Source: §4.3 · p. 18
• Conversational recommendation and the needs of more specific job-seeker groups are promising underexplored directions.The paper cites eliciting preferences, explanations, and explicit feedback as potential value of conversational systems, and proposes a taxonomy of groups and their needs.Source: §4.3 · p. 18
• Few public datasets are available, and the authors argue that more public data could strengthen benchmarks and encourage new ideas.They identify two public datasets and note that the Xing RecSys challenge datasets are not publicly available.Source: §4.3 · p. 19

A SUPPLEMENTARY MATERIALS

• Table 1 provides an overview of collected e-recruitment recommendation systems, sorted by publication year.It records recommended entities and method categories, including content-based, collaborative-filtering, knowledge-based, and hybrid or other methods.Source: A SUPPLEMENTARY MATERIALS · p. 26
• The table does not label a paper reciprocal merely because its design could be reciprocal; the paper must claim reciprocity and experiment with both recommendation tasks.The stated rule concerns recommendations of jobs and job seekers.Source: A SUPPLEMENTARY MATERIALS · p. 26
• Method type is left unreported for papers focused on preprocessing, postprocessing, or reranking when they do not specify it in detail.Source: A SUPPLEMENTARY MATERIALS · p. 26