Source-linked AI summary
Personalized and Aspiration-Oriented Career Path Recommendation
Kuleshwar Sahu, Girish Keshav Palshikar, Rajiv Srivastava
TL;DR
The paper addresses personalized career-path recommendation from an employee’s existing career and aspirations. It combines career-path and aspiration similarity using DKD and URLA methods, with URLA learning event representations and aligning sequences. URLA achieves better DCG than DKD while the two methods’ rankings remain coherent.
Problem
Career profiling and advisory are challenging because employees face expanding career options and guidance must match both their past profiles and aspirations.
Method
The method combines career-path similarity for personalization with aspiration similarity for fulfillment, using DKD feature matching and URLA embedding-based temporal alignment.
Results
URLA produces higher mean DCG than DKD, while rankings show positive concordance; at rank 5, mean Kendall’s tau is 0.36.
Takeaways & Limitations
URLA provides consistent recommendation quality without requiring domain knowledge to model career-path similarity and retains temporal information through alignment.
Takeaways & Limitations
DKD relies on exact matching and therefore cannot model partial or non-discrete similarity between different event values or attributes.
Abstract
from arXiv · showhide
Fulfilling career aspirations is important for growth of employee and organization. We propose a data driven methodology to recommend personalized career path for a given aspirant's career path and aspirations. The pro-posed method uses the career path similarity (CPS) between aspirant's career and candidate career path, and 'aspirational similarity' (AS) between aspiration and candidate career paths to find suitable career path. CPS ensures personalized recommendation while AS ensures aspiration fulfillment. We defined two methods to compute the CPS between career paths which are (a) domain knowledge driven (DKD) and, (b) unsupervised representation learning and alignment (URLA) based, along with different AS measures. The DKD based similarity is defined in the terms of features extracted and summarized over career paths. In the URLA, we use the sequence of event names present in the career paths of the employees to learn the embedding for each event name. In URLA we use learned embedding vector of the career path event names and as-sociated event attributes (skill cluster and domain) to find the best alignment between two career paths. We hypothesized that relative position of event names in the sequence represents semantics of event name and that can be learned. We use LSTM neural network to learn the embedding vector of each career event name. We also define the matching method to compute the AS be-tween aspiration and career path in both proposed methods. We combine CPS and AS to rank available 'candidate career paths' of employees to find the suitable one. We get better DCG value in URLA as compare to DKD. We also showed that ranking are coherent using both the methods. URLA method is better since it does not require domain knowledge to model the similarity and includes temporal aspect by optimal Levenshtein alignment using weighted cosine distance.
1 Introduction
The paper frames personalized career guidance as a challenging sequence-similarity problem and proposes two unsupervised, aspiration-aware methods using employee career data.
- Rapidly expanding technologies, tasks, roles, and service offerings increase employees’ available career options and organizations’ training needs.
- Career profiling and advisory can use eHRM data to identify similar, senior employees whose career paths match an employee’s aspirations.
- The proposed methods recommend personalized career paths from employee profiles containing project assignments, skills, roles, domains, and learnings.
- Career paths are modeled as chronological sequences of events including project allocations, role changes, training, and certification.
- DKD extracts cumulative career-path features using domain knowledge, whereas URLA learns event representations and computes sequence-to-sequence similarity with weighted edit distance.
- URLA addresses DKD’s exact-matching limitation by learning embeddings for events and attributes to support more granular similarity.
2 Related work
Prior work studies cumulative and sequential profile similarity for recruitment, career prediction, and professional matching; this paper adds aspiration-aware, unsupervised recommendation with a domain-independent method.
- Earlier recruitment research treated next-job prediction as a supervised problem using candidate, organizational, and job-transition data.
- Other profile-matching approaches model career trajectories as job nodes with attributes, decaying the weights of recent events and penalizing dissimilar gaps.
- A cited shortcoming is that some existing approaches completely discard events that are not similar within a sequence.
- Related methods learn temporally smooth attribute weights from social-media data to predict an employee’s next role.
- This paper incorporates aspirations, uses unsupervised representation learning, and proposes a domain-independent method with comparable recommendation quality.
3 Problem Formulation
The problem formulation represents employee careers as time-ordered event tuples and defines candidate career paths over project, role, training, and certification events.
- Each career path c_i is represented as a list of time-ordered tuples containing event type, event name, start date, and end date.
- Event types are Project Allocation, Role Set, Training, and Certification.
- Event names carry skill-cluster and domain attributes except for Role Set events.
4 Solution approaches
The paper ranks personalized career recommendations by combining career-path similarity with aspiration similarity. It compares domain-knowledge-driven matching with embedding-based temporal alignment across career-event sequences.
- Candidate paths are ranked by averaging career-path similarity, which supports personalization, and aspiration similarity, which measures aspiration fulfillment.The highest-ranked candidate is reported as the recommended career path.
- Domain knowledge driven career advisory: DKD uses a non-symmetric, duration-weighted match, so candidate events with durations closer to the aspirant receive higher similarity.For example, a 600-day Developer role is preferred over a 300-day Developer role for a 500-day aspirant role.
- Domain knowledge driven career advisory: DKD computes career-path similarity from learning, role, project, and domain similarity components derived from career-path features.Its aspiration similarity separately matches aspired role, domain, and skill clusters to candidate paths.
- Domain knowledge driven career advisory: DKD cannot model semantic similarity between different event values and does not account for temporal alignment across career paths.The paper motivates embedding-space matching to address partial matches and sequence alignment.
- Unsupervised representation learning and alignment: URLA learns event-name embeddings from career-event sequences and represents each career path as temporally ordered embedded sub-paths for four event types.The event types are Project Allocation, Role Set, Training, and Certification.
- Unsupervised representation learning and alignment: URLA computes similarity by aligning event-type-specific sub-paths with normalized Levenshtein similarity using duration-weighted cosine distances.The method combines similarities across the separate sub-career paths.
5 Case study data set
The case study uses a large IT-industry career-event repository and derives current paths, future paths, and aspirations from employees’ observed trajectories. Candidate paths are then selected for aspirants whose role, domain, and skill-cluster aspirations are available.
- The dataset contains 24,184 employee career paths and 250,470 events spanning Project Allocation, Role Set, Training, and Certification.The career paths are drawn from a large IT industry.
- Each career path is split into a current 30% segment and a future 70% segment by event duration.The future segment supplies the aspiration information used in the case study.
- The study identifies 2,461 current–future path pairs containing aspired role, domain, and skill-cluster information.Aspired domain is the dominant future domain, while aspired role and skill clusters are extracted from the future path.
- For each aspirant, candidate paths are searched across available employee career paths using the extracted aspiration components and experience requirements.Table 4 summarizes the number of candidate paths per aspirant employee.
6 Experimental results
The experiments compare DKD and URLA on career-path recommendation, finding that DKD better recalls the original path while URLA performs better for unseen alternatives and achieves higher DCG.
- DKD reports the highest precision@k because exact matching makes the original career path likely to be retrieved.For alternatives beyond the original path, DKD has mean error 0.117, compared with 0.017 for URLA.
- URLA finds similar unseen career paths more effectively, retaining similarity after the original path is identified.
- URLA achieves higher mean DCG values than DKD, indicating better or comparable recommendation quality under the reported relevance-based evaluation.Relevance is normalized by the similarity of the top-ranked career path before averaging DCG across aspirants.
- At rank 5, the mean Kendall’s tau coefficient between DKD and URLA rankings is 0.36, indicating positive pairwise concordance.
- Across top ranks 1 to 100, mean Kendall’s tau remains positive at more than 0.33, indicating possible coherence between the two rankings.
7 Conclusion and future work
The proposed algorithms provide personalized career-path recommendations with rich, temporally ordered events. URLA improves exploration without requiring extensive domain knowledge, while future work targets whole-path embeddings and broader evaluation.
- Both algorithms recommend personalized career paths containing rich event sequences organized along a timeline.
- DKD uses domain-knowledge-based exact matching and is better at recalling seen career paths according to Precision@1.
- URLA reduces dependence on granular domain knowledge and achieves better DCG by learning higher-dimensional event representations.
- Future work includes learning a single embedding vector for the complete sequence of temporally ordered career events.
- Planned extensions evaluate joint ranking and embedding, ensembles across IT, BPO, and banking, and recruitment applications.