Deposit subscribe Prediction using Data Mining Techniques based Real Marketing Dataset

Recently, economic depression, which scoured all over the world, affects business organizations and banking sectors. Such economic pose causes a severe attrition for banks and customer retention becomes impossible. Accordingly, marketing managers are in need to increase marketing campaigns, whereas organizations evade both expenses and business expansion. In order to solve such riddle, data mining techniques is used as an uttermost factor in data analysis, data summarizations, hidden pattern discovery, and data interpretation. In this paper, rough set theory and decision tree mining techniques have been implemented, using a real marketing data obtained from Portuguese marketing campaign related to bank deposit subscription [Moro et al., 2011]. The paper aims to improve the efficiency of the marketing campaigns and helping the decision makers by reducing the number of features, that describes the dataset and spotting on the most significant ones, and predict the deposit customer retention criteria based on potential predictive rules.

💡 Research Summary

The paper addresses the challenge of customer attrition in the banking sector during periods of economic downturn by applying data‑mining techniques to a real‑world marketing dataset. The authors use the well‑known Portuguese bank direct‑marketing data (originally described by Moro et al., 2011), which contains information on 17 marketing campaigns and 79 354 contacts. Two CSV files are provided: the full set (45 211 records) and a 10 % random sample (4 521 records). Because raw CSV files are inconvenient for mining, the authors first transform the data into a relational database (RDB) consisting of three tables—Client_info, Bank_info, and Campaign_info. Categorical attributes such as “Job” are stored in a separate lookup table and linked via foreign keys; some continuous attributes are discretized to simplify subsequent analysis.

The study focuses on two classification techniques: Rough Set Theory (RST) and Decision Tree (DT) learning.

Rough Set Theory (RST) reduction
Using Pawlak’s indiscernibility relation, the authors construct a discernibility matrix for the 16 conditional attributes (excluding the decision attribute Deposit_s). By checking the independence condition for each attribute, they identify a minimal reduct that preserves the lower and upper approximations of the original information system. The resulting core set consists of three attributes: Age, Balance, and Duration. These three features alone are sufficient to approximate the decision space with minimal loss, demonstrating that most of the original 16 attributes are redundant for the target classification. The authors then generate decision rules from the core set and report several rules whose error rate is below 25 %. Example rules include combinations of Age, Balance, and Duration that achieve accuracies ranging from 82 % to 99.9 % when validated on the full dataset.

Decision Tree (C4.5) induction
For the same dataset, the authors compute the information‑gain ratio for each attribute. Duration obtains the highest gain ratio (≈0.108) and becomes the root node of the tree. Using the C4.5 algorithm, they grow a tree with 104 leaves and 146 nodes. The first split separates records with Duration ≤ 211 seconds from those with Duration > 211 seconds. Subsequent splits involve attributes such as Poutcome, pdays, marital status, age, contact type, and month. From the tree, a set of explicit IF‑THEN rules is extracted. The most straightforward rule—“If Duration ≤ 211 seconds then Deposit_s = No”—achieves 97.13 % accuracy. Other rules combine Duration ranges with Poutcome, pdays, marital status, and contact channel, yielding accuracies between 61 % and 99.9 %. The authors note that the tree is easy to interpret, can be pruned, and can be directly used for predicting new cases.

Analysis and discussion
The paper compares the two approaches. RST provides a compact representation (three core attributes) and highlights the most informative variables, which is valuable for reducing data collection costs and simplifying models. DT, on the other hand, exploits the full attribute set to capture complex interactions, offering finer-grained segmentation useful for targeted marketing actions. The authors argue that a hybrid workflow—first applying RST to select a minimal, high‑impact feature subset, then training a decision tree on that reduced space—could combine interpretability with predictive power.

Several limitations are acknowledged. The discretization and foreign‑key mapping may introduce information loss; the use of a 10 % sample for initial experiments raises concerns about sampling bias; the paper lacks a detailed description of pruning parameters, which hampers reproducibility; and performance evaluation relies solely on overall accuracy, omitting precision, recall, F1‑score, and cost‑sensitive metrics that are crucial in a banking context where false positives and false negatives have asymmetric financial implications.

Conclusion
The study demonstrates that rough‑set‑based feature reduction and C4.5 decision‑tree classification can effectively predict whether a bank client will subscribe to a term deposit, using a publicly available marketing dataset. The identification of Age, Balance, and Duration as the core predictive attributes aligns with domain intuition (older clients with higher balances and longer call durations are more likely to convert). The extracted rules achieve high accuracy and can guide marketing managers in allocating resources toward the most promising prospects, thereby reducing campaign costs and improving revenue. Future work is suggested to incorporate cross‑validation, cost‑sensitive analysis, and comparisons with modern ensemble methods such as Random Forests or Gradient Boosting Machines to further validate and enhance the robustness of the proposed approach.