Source-linked AI summary
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
TL;DR
The paper addresses incomplete contact-level B2B marketing data and the challenge of predicting purchases during long buying cycles. It combines multi-key contact aggregation, irregular company-name clustering, feature generation, and CatBoost conversion prediction, achieving over 90% accuracy while treating campaign recommendations as feasible but not guaranteed to produce significant differences.
Problem
Long, infrequent B2B buying cycles make it important to aggregate contact data, identify likely purchasers, and support campaign targeting.
Method
The framework aggregates contacts with multiple keys, clusters irregular company names, generates firmographic and campaign-engagement features, and applies CatBoost for conversion prediction.
Results
Over 90% prediction accuracy is achieved with CatBoost for customer conversion into the sales funnel.
Takeaways & Limitations
Prediction results and SHAP analysis support discussion of personalized campaign actions intended to foster conversion.
Takeaways & Limitations
The proposed actionable recommendations are not guaranteed to make significant differences.
Abstract
from arXiv · showhide
In the one-time selling B2B context, the buying cycle may last months or even years. During the long process, targeting customers that have a high potential to make purchases and recommending personalized campaigns accordingly are important for effective marketing. For this goal, we study the following problems, B2B customer data aggregation, customer feature generation, and prediction of whether a B2B customer would show interest in making a purchase (i.e., prediction of conversion into sales funnel). We propose an algorithm to aggregate individual contacts to the B2B customer level based on multiple keys. For non-standardized keys such as company names, we propose a novel architecture to cluster them in a domain encompassing irregularities such as spelling mistakes and spelling variants. We then define and generate a set of features and apply the CatBoost model for customer conversion prediction. Our framework achieves 91\% prediction accuracy. Based on the prediction results and analysis of the model, we then discuss personalized campaign recommendations to foster conversion.
1 INTRODUCTION
B2B marketing analysis must aggregate incomplete contact-level records, generate suitable customer features, and predict conversion during long, infrequent buying cycles. The paper addresses these challenges with contact aggregation, feature generation, CatBoost prediction, and campaign recommendations.
- B2B marketing data can support questions about purchase likelihood and campaign sequences, including variation across regions and industries.
- Contact-level records create aggregation difficulties because complete customer behavior requires combining activities from all employees, despite missing engagement data.
- LRFM models and customer-behavior time series are unsuitable for one-time or infrequent B2B selling, motivating demographic, firmographic, and campaign-engagement features.
- Personalized B2B campaign recommendations are rarely studied, so the paper discusses actions intended to increase conversion ratios from its prediction results.
- The framework aggregates contacts using multiple keys, handles irregular company names, predicts conversion, and discusses actions based on prediction results.
2 RELATED WORK
Prior work covers word clustering, feature selection, conversion prediction, and recommender systems, but important gaps remain for irregular company-name data and B2B conversion settings. Existing conversion studies mainly use more accessible eCommerce or B2C behavioral data.
- Company-name cleaning lacks dictionary and contextual support because made-up words and absent context limit standard spelling-correction approaches.
- Feature-selection research emphasizes identifying relevant attributes across financial, operational, marketing, and textual business data.
- B2B conversion prediction has received little attention compared with eCommerce and B2C studies using more complete online behavior or click-stream data.
- B2B campaign recommendation research includes a social and temporal model, within broader recommendation approaches such as content-based, collaborative, and hybrid methods.
3 DATASET AND PRE-PROCESSING
The dataset combines incomplete individual-contact and customer-company records, requiring additional keys and external firmographic sources before customer-level analysis. Pre-processing therefore focuses on completing company links and standardizing geographic and company information.
- The marketing data are incomplete: some contacts lack Account IDs, while Industry, Revenue, and Employee Number appear only in some records.
- The first objective is to aggregate contacts into companies so campaign participation and temporal sequences can be assembled, including for records without Account IDs.
- Country and City are cleaned through translation, lowercasing, special-character removal, stop-word removal, and manual checking before serving as keys.
- Web scraping supplements missing firmographic features, followed by string processing to reconcile formats between scraped and existing data.
4 DOCUMENT REPRESENTATION AND CLUSTERING ARCHITECTURE
The document-representation architecture standardizes irregular company-name vocabulary through subword clustering, then groups strings in a cluster-based vector space. Similarity thresholding and connected-component structure address spelling variation while limiting overly broad transitive groups.
- 4.1 Architecture Overview: Company names are difficult to standardize with vocabulary checks or direct word embeddings because they contain many made-up words.
- 4.1 Architecture Overview: The architecture has two phases: word-level clustering and string grouping in the resulting cluster vector space.
- 4.2 Word-level Clustering: Input company names are pre-processed by lowercasing, removing customized stop words, and using processed names to search public company profiles.
- 4.2 Word-level Clustering: Words are encoded as bigram vectors, reduced with PCA retaining 95% variance, and clustered with k-means so cluster indices become standardized vocabulary representations.For inputs limited to 26 letters and 10 numerals, each bigram encoding has dimension 1,296.
- 4.3 String Grouping in Cluster Vector Space: Cluster indices form a bag of words, which is transformed with TF-IDF and normalization into vector representations for input strings.
- 4.3 String Grouping in Cluster Vector Space: A dot-product threshold of 0.7 identifies similar strings, while maximal cliques of connected components prevent long similarity chains from merging very different names.
- 4.2 Word-level Clustering: Increasing the number of word clusters reduces cluster diversity, creating a trade-off between preserving spelling variation and admitting noisy clustering.
5 CONTACT GROUPING ALGORITHM
The contact grouping algorithm combines multiple keys and connected components to aggregate individual contacts into customer accounts. Applied separately across company names and countries, it produced 93,568 accounts.
- 5 CONTACT GROUPING ALGORITHM: Two criteria link contacts to one customer account: matching Company Name, Country, and City, or matching Company Name, Account ID, and Country.Company Name alone is insufficient because subsidiaries may make independent purchasing decisions.
- 5.1 Connected Components Based Contact Grouping Algorithm: The method first groups contacts by Company Name, Country, and City, then uses connected components to merge groups sharing Account IDs.Each graph node represents a contact group, and each connected component corresponds to one customer account.
- 5.1 Connected Components Based Contact Grouping Algorithm: For each Company Name and Country, the implementation maps cities to contacts and Account IDs, constructs a shared-Account-ID graph, and groups each connected component.Records within each resulting company are then ordered temporally for campaign-sequence construction.
- 5.1 Connected Components Based Contact Grouping Algorithm: The algorithm organizes related contacts into temporal campaign sequences and records multi-person participation in campaigns identified by the same Campaign ID.Campaign records are sorted by time and Campaign ID after grouping.
- 5.2 Experimental Results: 93,568 customer accounts resulted from applying the two contact-grouping criteria separately.The resulting accounts can include contacts from various cities, leaving City blank in such cases.
6 FEATURE EXTRACTION AND FEATURE SELECTION
The framework extracts firmographic and campaign-activity features at the customer-account level, truncating activity sequences at conversion or purchase time. It then reduces and selects features, yielding a final set of 40 features for modeling.
- Feature extraction: The method extracts firmographic and campaign-activity features after aggregating contact data into customer accounts.Because contact grouping is approximate and the data are incomplete, an account can occasionally contain contacts with differing firmographic information.
- Feature extraction: Campaign-activity sequences aggregate activities from contacts belonging to each customer account but are truncated at conversion and purchase times.This removes activities occurring after conversion, such as website contact for after-sales services.
- Feature selection: 156 generated features comprised 149 numerical and 7 categorical features, with selection applied only to the numerical variables.Feature selection is intended to reduce computational cost and improve model performance.
- Feature selection: 40 features remain for modeling: 33 selected numerical features plus 7 categorical features.The categorical features are Region, Sub-region, Industry, Products, last campaign, second to last campaign, and third to last campaign.
- Feature selection: 23 constant, 13 highly correlated, and 28 zero-importance features were removed before SFFS selected 33 of the remaining 85 numerical features.The high-correlation threshold was greater than 0.99.
7 CONVERSION PREDICTION BASED ON CATBOOST
The study evaluates CatBoost and other models for predicting B2B customer conversion using account-level features, then interprets CatBoost predictions with SHAP values to support campaign recommendations.
- 7.2 Experiment and Results: The experiment compares CatBoost, random forest, logistic regression, and an RNN using accounts with at least two campaign activities.The dataset contains 93,568 customer accounts, including 15,724 with at least two campaign activities.
- 7.2 Experiment and Results: 91.0% conversion prediction accuracy was achieved by CatBoost trained on 40 selected features.The selected set combines 33 numerical features with 7 categorical features.
- 7.3 Action Recommendation Based on CatBoost Model and SHAP Values: SHAP values explain how features affect CatBoost conversion predictions, with positive values associated with higher predicted conversion and negative values with lower prediction.TreeExplainer is fitted to the trained CatBoost model, and the top 20 features are summarized.
- 7.3 Action Recommendation Based on CatBoost Model and SHAP Values: Average campaign intervals of no more than 3 quarters and participation by more than one person per campaign tend to have positive SHAP impacts on conversion.These patterns are presented for campaign timing and average attendance regardless of campaign type.
- 7.3 Action Recommendation Based on CatBoost Model and SHAP Values: For an unconverted account, SHAP-based analysis suggests encouraging campaign participation within a shorter interval than five quarters, while the authors caution that recommendations are not guaranteed to produce significant differences.The paper identifies causal inference as a future step for determining causality between variables.
8 CONCLUSION
The paper presents a framework for cleaning marketing data, aggregating contacts at the account level, predicting conversion with CatBoost, and discussing actions based on model interpretations.
- 8 CONCLUSION: The framework combines contact aggregation, irregular company-name clustering, feature development, and CatBoost conversion prediction for B2B marketing data.It achieves prediction accuracy of over 90% and discusses action recommendations based on prediction results and model interpretation.
- 8 CONCLUSION: The proposed aggregation algorithm groups individual contacts at the B2B customer level using multiple keys, including non-standardized company names with spelling irregularities.