B2B Customer Conversion Prediction: A Document Representation, Graph Theory, and CatBoost Driven Methodology
Tianqi Wang, Sheikh Shams Azam, Wan Eih Huang, Anton Wiranata, Christopher G. Brinton, Jan P. Allebach
cs.LG
Quick verdict
Claims a B2B conversion pipeline reaching 91% accuracy with CatBoost on proprietary HP data; solid engineering, but weak evaluation and no external validation.
Summary
This KDD 2023 workshop paper presents an end-to-end framework for predicting whether B2B customers convert into the sales funnel, using proprietary marketing data from HP Inc. The pipeline has three stages. First, 126.1k individual contacts are aggregated into 93,568 customer accounts via connected components over multi-key criteria combining company name, country, city, and account ID. Second, noisy company names are standardized by a clustering architecture: bigram one-hot encodings, PCA retaining 95% variance, k-means (k=30,000), TF-IDF over cluster indices, dot-product similarity (threshold 0.7), and iterative largest-maximal-clique grouping. Third, 156 firmographic and campaign-engagement features are reduced to 40 (correlation filtering, LightGBM importance, SFFS) and fed to CatBoost, compared against random forest, logistic regression, and an RNN. On a class-balanced subset (2,194 converted versus 2,194 sampled unconverted accounts, all with at least two campaign activities; 75/25 split), CatBoost with the 40 selected features reaches 91.0% test accuracy (90.2% with all 156 features), beating random forest (82.5%), logistic regression (74.7%), and the RNN (77.4%). SHAP analysis motivates campaign recommendations (e.g., inter-campaign intervals of three quarters or less), which the authors concede are correlational. Main limitations: proprietary data and no code prevent replication; a single split with accuracy as the only metric; no quantitative validation of the entity-resolution step; and undisclosed baseline tuning.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 0 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Contacts can be aggregated to B2B customer accounts from incomplete CRM data using connected components over company name, country, city, and account ID keys.
- The proposed bigram-PCA-k-means-TF-IDF-maximal-clique architecture clusters company names containing spelling mistakes and variants without a dictionary or context.
- CatBoost on 40 selected features predicts sales-funnel conversion at 91.0% test accuracy, outperforming random forest, logistic regression, and an RNN on the class-balanced evaluation set.
- SHAP analysis identifies campaign cadence (intervals of three quarters or less) and multi-person attendance (more than one person per campaign) as actionable correlates of conversion, though the authors concede causality is unverified.
Strengths
End-to-end treatment of genuinely messy industrial CRM data
Concrete multi-key aggregation (§5) with two explicit criteria and connected components turns 126.1k fragmented contact records (Table 1) into 93,568 customer accounts, handling the same-name-different-subsidiary case via country/city constraints.
Well-specified, tunable text-clustering architecture
Thresholds are disclosed throughout (fuzzy ratio 70, dot product 0.7, PCA 95% variance) and k is explored over {10000,...,50000} with a worked sensitivity table (Table 2), which is more design transparency than typical industrial papers offer.
Principled, marketing-grounded feature engineering and selection
Interpretability with candid limits
Multi-model comparison across feature sets
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This KDD 2023 workshop paper (2nd Workshop on End-to-End Customer Journey Optimization, research supported by HP Inc.) tackles three linked problems in one-time-selling B2B marketing: (1) aggregating 126.1k individual contacts into customer accounts from incomplete CRM data (Table 1), (2) clustering noisy company-name strings, and (3) predicting conversion into the sales funnel. The headline claim — 91% prediction accuracy with CatBoost — is supported by the paper's own single-split experiment (§7.2), but its evidential weight is limited by the evaluation design discussed below. A fourth, weaker claim is that SHAP-based model inspection can ground personalized campaign recommendations (§7.3), which the authors themselves frame cautiously.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does the B2B conversion prediction paper achieve?
The paper reports 91.0% test accuracy for CatBoost using the 40 selected features (90.2% with all 156 features), evaluated on a class-balanced dataset of 2,194 converted versus 2,194 sampled unconverted accounts with a single 75/25 train/test split. No precision, recall, F1, or AUC values are reported, and there is no cross-validation or significance testing.
Is the code or data for this paper available?
No. The data are proprietary HP Inc. marketing data (stated in the Acknowledgments) and the paper mentions no public code repository. The pipeline steps and hyperparameters are described in enough detail to reimplement, but the reported results cannot be independently replicated.
How does the method match messy company names with spelling errors?
It lowercases the name and removes legal-form stopwords, searches the web for the company profile and adopts the scraped name when the fuzzy ratio exceeds 70, encodes words as summed one-hot bigram vectors (dimension 1,296), applies PCA (95% variance retained) and k-means with k=30,000, represents each name as a TF-IDF vector of cluster indices, links names whose normalized dot product exceeds 0.7, and groups the resulting graph by iteratively extracting the largest maximal clique.
What actionable marketing insights does the paper give?
SHAP analysis indicates that an average interval of three quarters or less between successive campaigns and more than one person attending each campaign are associated with higher conversion predictions. The authors themselves note these are correlational findings, not guaranteed levers, and propose causal inference as future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Tables 2, 3, 4, 5 and 7 were extracted as flattened run-on text; Table 2's four example-cluster columns interleave, so the per-k cluster memberships cited are a best-effort alignment of clearly delimited cell contents (all quoted strings appear verbatim in the extraction).
- Figures 1-5 are images; only their captions were extractable, so visual details of the SHAP plots beyond the features discussed in the body text could not be checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.