Using autoencoders for session-based job recommendations

Abstract

In this work, we address the problem of providing job recommendations in an online session setting, in which we do not have full user histories. We propose a recommendation approach, which uses different autoencoder architectures to encode sessions from the job domain. The inferred latent session representations are then used in a k-nearest neighbor manner to recommend jobs within a session. We evaluate our approach on three datasets, (1) a proprietary dataset we gathered from the Austrian student job portal Studo Jobs, (2) a dataset released by XING after the RecSys 2017 Challenge and (3) anonymized job applications released by CareerBuilder in 2012. Our results show that autoencoders provide relevant job recommendations as well as maintain a high coverage and, at the same time, can outperform state-of-the-art session-based recommendation techniques in terms of system-based and sessionbased novelty.

Job Recommender Systems: A Review

“…Contributions using textbook CF methods include Lee et al [72], Reusens et al [102], Ahmed et al [4], Liu et al [79], where the latter two applied it to the RecSys 2016 dataset. Lacic et al [68] compare several auto-encoders to encode user interactions, based on which similar users are determined.…”

“…However, it is frequently used to evaluate content-based recommender systems (see Table 1). Second, to our knowledge, only Lacic et al [68] compare their proposed recommender system over multiple datasets: the Careerbuilder dataset, the 2017 RecSys dataset, and a private dataset originating from the student job portal Studo Jobs [112]. What is striking, is that the performance of the different models differs considerably across datasets, with no unanimous agreement across data sets and across error measures which model should be preferred.…”

Improving Next-Application Prediction with Deep Personalized-Attention Neural Network

“…In [11], they use a hierarchical LSTM to predict the next potential employer of a person and how long he/she will stay in the new position. However, predicting the next job application based on application records received relatively little consideration except [12], which recommends the next job posting in a K-nearest neighbor manner. It is one of our baselines.…”

“…The textual job content includes a Job Title, a Job Description and some Job Requirements. From this initial dataset, we created two datasets: (i) CB12 s like in [12], in which sessions are created via a time-based split of 30 minutes inactivity threshold, and we discarded sessions with less than two applications for nextjob prediction purpose. (ii) CB12 l uses all application records during 13 weeks to model the career profile of each job seeker.…”

“…• Association Rule (AR): is a simplified version of association rule [31] with a maximum rule size of two; • Content Similarity (CS): recommends similar jobs based on the cosine similarity between representations of each applied job and the N most recently applied jobs in the global buffer; • Item-kNN (IkNN): recommends jobs that are similar to the last applied job during the current session as in [17]; • Session-kNN (SkNN) [32]: compares the entire current session with the past sessions in the training dataset, rather than considering only the last job; • Vector Multiplication SkNN (V-SkNN) [33]: is a variant of SkNN that emphasizes jobs more recently interacted within the current session, when computing the similarities with past sessions (a linear decay function is used); • VAE Comb: is the best model proposed in [12]; • GRU4Rec: is the most recent version of GRU4Rec [19];…”