{"id":"04234c0b-df49-4893-b6c8-42bfe61c5c71","arxiv_id":"2502.01660","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A cross-component attention transformer predicts individual employee turnover across firms by combining employee, company, competitor, and departed-colleague signals, reporting large top-K precision gains over baselines.","lead":"This paper trains a transformer model with five interlinked components to predict which individual employees will leave their company within the next six or twelve months, using public career profiles plus company financials. If the reported gains are real, recruiters and HR teams could focus outreach on employees who are actually likely to switch jobs, cutting search costs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Competitor and title graphs are built from all career trajectories without temporal restriction, so test-period transitions may leak into features and the reported Precision@30 lift is unverified.","rationale":"I read the paper as making a practical predictive claim: CATCICE achieves 0.800 Precision@30 versus 0.667 for HCPNN on 6-month turnover prediction, with the gain attributed to competitor influence and contagious effect. For that claim to hold, the features must be computable at prediction time. Sections 3.3.2 and 3.3.3 construct both graphs from 'all employee career trajectories', and the data were collected in March 2017, so the graphs contain the 2016 transitions used as test labels. This is the weakest assumption because the reported gains are concentrated in exactly the two components fed by those graphs, as shown by the Section 5.4 ablation. The paper's explicit statement in Section 5.1.1 that non-overlap of prediction periods means no leakage is insufficient: leakage can occur at feature-construction time rather than at label-split time. I am not claiming the authors intentionally leaked labels; the omission may be an oversight, but it is decisive for the central claim. The proposed test, rebuilding graphs using only transitions completed before each prediction window and re-running the evaluation, would settle it. If performance persists, the concern is resolved; if it collapses, the headline lift is an artifact. I therefore agree with the reader's weakest_assumption and see no reason to change the REJECT verdict, though the test is cheap enough that a CONDITIONAL acceptance with code and temporally restricted graph results could also be considered if the authors address the issue.","tokens_in":24000,"tokens_out":5062,"duration_ms":46795,"concrete_test":"Re-run the Test6 and Test12 evaluations with a temporal restriction: construct the competitor and title graphs using only job transitions whose end date is before the start of each prediction window (before 2016-01-01 for Test6, and for each monthly feature, only transitions completed before that month), then retrain CATCICE and all baselines. If Precision@30 drops materially from 0.800 toward the HCPNN level of 0.667 (or AUC from 0.704 toward 0.696), the reported advantage is attributable to lookahead in graph construction. As a supporting check, compute how many top-10 competitor pairs for companies with many 2016 leavers are connected by edges that exist only because of 2016 transitions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CATCICE's 19.9% Precision@30 lift (0.800 vs 0.667) and AUC improvement (0.704 vs 0.696) come from modeling competitor influence and contagious effect. This requires those features to be computed from information available at prediction time. The paper does not establish this. Sections 3.3.2 and 3.3.3 state that the competitor graph and title graph are built from 'all employee career trajectories', with no temporal restriction; profiles were scraped in March 2017 and include 2016 transitions. For Test6 (observation through 2015/12, prediction 2016/01-06), the graphs therefore contain the very test-period transitions that define the labels. A test employee who leaves firm A for firm B in June 2016 adds edge (A,B) to the competitor graph; HOPE embeddings then make B look similar to A, so B's financial features enter the competitor-influence vector x_CI for all A employees, including the leaver. The title graph similarly uses future title transitions to choose similar peers for the contagious-effect feature. The statement in Section 5.1.1 that non-overlap of prediction periods implies no data leakage addresses only the temporal split of labels, not graph construction. The ablation in Section 5.4 is consistent with the leak: removing CI or CE drops Precision@30 from 0.800 to 0.600/0.633, so the headline gain is concentrated in the two components built from all-data graphs. Without a temporally restricted graph construction, the claimed superiority is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes CATCICE, a cross-component attention transformer for predicting individual employee turnover across firms using public career profiles and company financial statements. The model operationalizes employee, company, employee-company, competitor-influence, and contagious-effect features, with competitor and title graphs embedded by HOPE and aggregated via similarity-weighted averages. The paper reports evaluations on six-month and twelve-month test windows against eleven baselines, reporting AUC, Precision@k, NDCG@k, an ablation study, a recruiter cost-savings simulation, and SHAP-based interpretation. The authors claim superior performance and business value, with headline numbers including Precision@30 of 0.800 versus 0.667 for the best baseline and AUC of 0.704 versus 0.696.","tokens_in":24372,"tokens_out":5761,"duration_ms":57490,"significance":"The studied problem is practically important and underexplored, and the design-science framing with job embeddedness theory is a strength. The proposed five-component transformer with cross-component attention is architecturally novel relative to standard transformers, and the paper includes a temporal split, multiple metrics, an ablation study, and interpretability analysis. I credit the authors for a clearly described feature operationalization and for stating the temporal split explicitly. However, the empirical evidence for the central claim is compromised by a likely future-information leak in the graph construction, and the reported gains are not accompanied by confidence intervals, significance tests, or code/data release. The main comparative results therefore are not trustworthy as presented.","major_comments":[{"comment":"The claim in Section 5.1.1 that non-overlap of prediction periods implies 'no data leakage issue' addresses only the label split, not the construction of the competitor and title graphs. Sections 3.3.2 and 3.3.3 state that both graphs are built from 'all employee career trajectories' with no temporal restriction, and the profiles were collected in March 2017. The Test6 prediction period is 2016/01-06 and Test12 extends through 2016/12 (Table 4), so test-period transitions are included in the graph edges; HOPE embeddings and the features x_CI and x_CE in Eqs. (2) and (4) therefore encode information from the future relative to the prediction time. This is exactly the kind of leak that can inflate the reported comparisons.","section":"Sections 3.3.2, 3.3.3 vs. Section 5.1.1"},{"comment":"The ablation results are consistent with the leak being the driver of the reported advantage. Removing the two components built from all-data graphs, competitor influence (CATCICE-CI) and contagious effect (CATCICE-CE), drops Precision@30 from 0.800 to 0.600 and 0.633, respectively, while removing the other components leaves higher Precision@30 values (0.300 for E is an exception, but E removal also sharply reduces AUC). The headline improvement over HCPNN (0.800 vs. 0.667) is thus concentrated in exactly the features that can see the test labels, so the results do not support the claim that modeling competitor influence and contagion improves prediction. A re-evaluation with graphs constructed only from data available before each prediction period is required.","section":"Section 5.4, Table 11"},{"comment":"All results are single point estimates on a single temporal split, with no confidence intervals, standard errors, or significance tests. The AUC improvement over the best baseline is small (0.704 vs. 0.696 for GBT on Test6; 0.716 vs. 0.708 on Test12). Without uncertainty quantification and with a single split, these differences are not sufficient to support the stated claim of 'superior performance over several state-of-the-art benchmark methods,' even leaving aside the leak in graph construction.","section":"Section 5.2, Tables 8-9"},{"comment":"The cost-savings simulation assumes that employees who will leave their current employers within six months will respond to InMail messages. This makes the number of responses a deterministic rescaling of Precision@30, so the cost comparison adds no independent evidence of business value; it simply restates the ranking metric under an assumption that already presumes the prediction is correct. The simulation should instead use observed response behavior or a sensitivity analysis with explicit, label-independent response rates.","section":"Section 5.3, Table 10"}],"minor_comments":[{"comment":"There is a typo in the second paragraph: 'muti-head self-attention' should be 'multi-head self-attention.'","section":"Section 6"},{"comment":"Several cross-references are unresolved, leaving placeholders such as 'Error! Reference source not found.' and raw REF fields in the text; these must be fixed before publication.","section":"Sections 3.3.1-3.3.4"},{"comment":"The paper reports only the number of layers and attention heads as the final architecture, with other training details relegated to Appendix A-3, which is not included in the submitted text; the final version should provide the full hyperparameter search and training details.","section":"Section 5.2"},{"comment":"The competitor graph contains 88,108 directed edges among 753 companies, which is very dense; the paper should discuss whether the top-10 similar competitors selected after HOPE are meaningfully distinct or whether the density makes the similarity measure mostly uniform.","section":"Table 5"},{"comment":"No code or data is provided; given the proprietary nature of the data, at least detailed feature-construction pseudocode and model training code should be released so that the temporal-leak concern can be checked by readers.","section":"Reproducibility"}],"recommendation":"reject","confidential_remarks":"The stress-test concern is well-founded and central. I agree with the reject recommendation: the main empirical claim is invalidated by the graph-construction leak, and the absence of significance tests and code leaves no way to verify the reported gains. If the authors can redo the evaluation with temporally restricted graphs and provide proper uncertainty quantification, a future resubmission would be worth considering, but the current manuscript does not support its headline conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the problem is real and not just repackaged: predicting individual employees' cross-firm turnover with competitor influence and contagion is not in the cited literature, and the design-science/job-embeddedness framing is a reasonable way to organize the feature set. Second, the evaluation has a load-bearing flaw: the competitor and title graphs are built from all career trajectories, with no temporal restriction, so for the 2016 test windows the graphs can include the very transitions being predicted. The paper's statement that non-overlapping prediction periods preclude leakage addresses the label split but not the graph features. Given that the ablation attributes the largest precision gains to the CI/CE components, the central claim cannot be trusted as written.\n\nWhat's good: the task formulation is new, the five-component architecture is a legitimate extension of transformer/cross-attention, and the feature operationalization from public profiles and financial statements is thorough. The cost-saving simulation and SHAP interpretations are standard but done cleanly. The literature review covers the relevant streams and positions the work correctly.\n\nSoft spots besides the leak: no code/data, no significance tests or confidence intervals, a single temporal split, and the AUC gain over the best baseline is small (0.704 vs 0.696). The Precision@30 lift (0.800 vs 0.667) is large but exactly where the leaked components matter most. The reviewer cannot verify whether the graph embeddings were fit before or after the prediction cutoff; the text says 'all employee career trajectories' without qualification, and that is the authors' burden.\n\nI would not desk-reject this. The method is plausible, the applied problem is important, and the leak is fixable by rebuilding graphs on training-period transitions only and rerunning with multiple seeds and released code. But as written, the empirical claims are unverified. Send it to review with that as the central condition.","headline":"The paper opens a genuinely new applied task and builds a sensible model, but graph construction leaks test-period transitions into features, so the headline precision gains are unverified as written.","tokens_in":24830,"tokens_out":2031,"would_cite":false,"duration_ms":22088,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A cross-component attention transformer that fuses competitor influence and contagious effects predicts individual employee turnover across firms with a 19.9% lift in Precision@30 over the best baseline.","keywords":["employee turnover prediction","cross-component attention","transformer","competitor influence","contagious effect","job embeddedness theory","HOPE graph embedding","talent analytics"],"falsifier":"Re-build the competitor and title graphs using only career moves that occurred strictly before each prediction window (for example, before January 2016 for the Test6 set), retrain and re-evaluate CATCICE under the same split, and compare Precision@30 and AUC against the baselines. If the gap over HCPNN shrinks to near zero, the reported 19.9% lift is leakage rather than model quality; if the gap persists, the cross-component attention is genuinely learning from the influence signals.","tokens_in":23818,"feed_emoji":"💼","tokens_out":4948,"duration_ms":44820,"temperature":0.7,"pith_summary":"This paper tries to predict, for individual employees across many companies, who will quit within the next six to twelve months, using only public career profiles and financial statements. The proposed model, CATCICE, combines five streams of information (personal attributes, company health, the employee-company match, the pull of rival employers, and the push of recently departed colleagues) in a transformer architecture whose attention mechanism lets each stream query the others at each time step. On a dataset of 87,070 U.S. tech workers, the model reports an AUC of 0.704 for six-month predictions versus 0.696 for the best baseline, and a 19.9% lift in Precision@30 (0.800 vs 0.667). The authors also estimate that recruiters using the model would pay $7.08 per responding candidate instead of $8.50 under the next-best method. The paper's deeper claim is that job embeddedness theory can be operationalized into measurable features, and that competitor influence and contagious effects are learnable, additive signals for turnover.","feed_headline":"Turnover model beats best baseline by 20% on top-30 precision","feed_subtitle":"CATCICE fuses competitor pull and colleague contagion to predict who switches jobs within six months.","key_machinery":"The load-bearing object is the five-component cross-component attention transformer. Each component (employee factors, company factors, employee-company factors, competitor influence, and contagious effect) is a stack of layers, where each layer first applies masked multi-head self-attention over that component's own time series, then runs cross-component attention: the component's query attends to the keys and values of all five components, the outputs are concatenated and compressed, and skip connections and feed-forward layers follow. The competitor-influence and contagious-effect inputs are themselves constructed by HOPE graph embeddings of a company-flow graph and a title-flow graph, respectively, with cosine-similarity-weighted aggregation of neighbors' features. The mechanism's purpose is to let the model learn inter-component time dependencies, for example how a rival's hiring spree in one month combines with the departure of similar-titled peers in a previous month to raise an employee's turnover risk in the next month.","core_discovery":"The central claim is that individual-level turnover across firms can be predicted more accurately than existing methods allow by explicitly modeling two social forces that prior work ignores: the magnetic pull of competing companies and the contagious effect of colleagues who have already left. CATCICE operationalizes these forces by building a directed company-flow graph and a directed job-title-flow graph from the full employment histories, embedding those graphs with HOPE, and using cosine similarities to identify each company's top competitors and each employee's closest departed peers. These influence features, together with employee, company, and employee-company factors, feed a five-component transformer in which masked multi-head self-attention captures intra-component time dependencies and a novel cross-component attention lets each component attend to the other four. The authors report consistent wins over eleven baselines spanning survival analysis, classical classifiers, RNNs, and graph embedding, with the largest gains in top-k metrics, and ablation shows every component contributes, with employee factors most important.","pith_inferences":["A direct test of the leakage hypothesis would rebuild the competitor and title graphs with a strict temporal cutoff before each prediction window; if the gains over baselines fade, the reported 19.9% lift is an artifact of lookahead rather than model quality.","The cross-component attention design is generic and could transfer to other multi-source sequential prediction problems, such as customer churn across competing services or student dropout with peer and institutional signals.","The cost-saving estimate assumes a simple response model (all leavers respond, all stayers do not) and fixed InMail prices; real recruiter workflows with different response rates would shift the dollar figures, though the ranking improvement would likely persist.","If the graphs leak future information, the SHAP-based interpretability analysis may also be misleading, since features would encode future moves rather than true drivers of turnover."],"forward_implications":["If the model's accuracy holds, recruiters can concentrate their outreach on the top fraction of predicted leavers, cutting the average cost per productive response from $8.50 to $7.08 relative to the next-best method.","HR departments can act before departure: the same features that predict a quit, such as many past position changes or recent departures of similar-titled peers, suggest which retention levers to pull.","The ablation results imply that competitor influence and contagious effect each add predictive value on top of standard employee and company features, supporting the theory that turnover is partly a social and competitive phenomenon.","The five-component attention architecture outperforms RNN-based models such as GRU and HCPNN, indicating that inter-component temporal dependencies are better captured by cross-component attention than by a single hidden-state sequence."],"supporting_citations":[{"why":"Supplies the kernel theory of job embeddedness (links, fit, sacrifice) that generates the meta-requirements and the five factor groups.","marker":"(Mitchell et al. 2001)"},{"why":"Provides the HOPE graph-embedding method used to produce company and title embeddings for competitor influence and contagious effect.","marker":"(Ou et al. 2016)"},{"why":"Supplies the transformer and masked multi-head self-attention architecture that CATCICE extends with its cross-component attention.","marker":"(Vaswani et al. 2017)"},{"why":"Provides the turnover-contagion evidence that departed colleagues' job search behavior influences quitting, justifying the contagious-effect component.","marker":"(Felps et al. 2009)"},{"why":"Establishes interfirm competition for human capital through employee-skill networks, grounding the competitor-influence operationalization.","marker":"(Liu et al. 2020)"},{"why":"Defines the HCPNN baseline, the strongest competitor on top-k metrics that CATCICE must beat.","marker":"(Meng et al. 2019)"},{"why":"Supplies Gradient SHAP, the attribution method used for interpreting which factors drive each turnover prediction.","marker":"(Lundberg and Lee 2017)"},{"why":"Prior work on contagious effect for turnover prediction that this paper extends and compares against.","marker":"(Teng et al. 2019)"}],"fun_headline_variants":["CATCICE fuses competitor pull and contagion to predict turnover","Attention transformer predicts job switches across firms","Modeling competitor influence and peer contagion improves turnover prediction","Turnover prediction: cross-component attention beats 11 benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The graphs that define who counts as a competitor and who counts as a departed peer are built from all employee career trajectories in the dataset, with no temporal cutoff; if those trajectories include moves that happened during or after the prediction windows, then the influence features for an employee at a given month are computed from future job moves, meaning the test labels have already shaped the features.","fun_headline_variants_meta":{"raw":{"variants":["CATCICE fuses competitor pull and contagion to predict turnover","Attention transformer predicts job switches across firms","Modeling competitor influence and peer contagion improves turnover prediction","Turnover prediction: cross-component attention beats 11 benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1481,"prompt_tokens":911,"completion_tokens":570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":527,"tokens_out":570,"duration_ms":5742,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:50:39.995344+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-build the competitor and title graphs using only career moves that occurred strictly before each prediction window (for example, before January 2016 for the Test6 set), retrain and re-evaluate CATCICE under the same split, and compare Precision@30 and AUC against the baselines. If the gap over HCPNN shrinks to near zero, the reported 19.9% lift is leakage rather than model quality; if the gap persists, the cross-component attention is genuinely learning from the influence signals.","supporting_citations":[],"review_version":1}