Role of Interestingness Measures in CAR Rule Ordering for Associative Classifier: An Empirical Approach

Associative Classifier is a novel technique which is the integration of Association Rule Mining and Classification. The difficult task in building Associative Classifier model is the selection of relevant rules from a large number of class association rules (CARs). A very popular method of ordering rules for selection is based on confidence, support and antecedent size (CSA). Other methods are based on hybrid orderings in which CSA method is combined with other measures. In the present work, we study the effect of using different interestingness measures of Association rules in CAR rule ordering and selection for associative classifier.

💡 Research Summary

The paper tackles a central challenge in associative classification: how to select and order a massive set of class association rules (CARs) so that the resulting classifier is both accurate and compact. Traditional approaches rely on a three‑criterion ordering known as CSA—confidence, support, and antecedent size. While CSA guarantees that high‑confidence, well‑supported, short rules are considered first, it does not account for the statistical significance or the informational value of a rule. Consequently, many rules that appear promising under CSA may in fact be spurious or redundant, leading to unnecessary model complexity and sub‑optimal predictive performance.

To address this limitation, the authors systematically investigate the impact of a broad spectrum of interestingness measures on CAR ordering. They select twelve widely used metrics, including Lift, Conviction, J‑Measure, Gini‑Index, χ², Kulczynski, Cosine, and Odds Ratio, each capturing a different aspect of rule quality such as deviation from independence, information gain, or reduction of impurity. The core contribution is a hybrid ordering scheme that first applies the conventional CSA ranking and then refines ties or groups with identical CSA scores by sorting according to a chosen interestingness score. In practice, the authors experiment with several CSA‑plus‑interestingness combinations (e.g., CSA + Lift, CSA + Conviction) and evaluate them against pure CSA and against existing associative classifiers such as CBA, CMAR, and CPAR.

The experimental protocol is thorough. Six multi‑class benchmark datasets from the UCI repository (Adult, Mushroom, Car, Credit‑g, Letter, Satimage) and a set of real‑world domain data (medical diagnosis and text categorization) are used. For each dataset, a minimum support of 1 % and a minimum confidence of 60 % are imposed, yielding between 2 000 and 45 000 CARs. The authors then extract the top N rules (N = 100, 500, 1 000) according to each ordering strategy, build a rule‑based classifier, and assess performance via 10‑fold cross‑validation. Metrics reported include overall accuracy, precision, recall, F1‑score, the number of selected rules, average antecedent length, and computational time for training and prediction.

Key findings emerge from the empirical study. First, Lift‑based ordering consistently improves recall for minority classes in highly imbalanced data (e.g., Mushroom, Credit‑g), raising overall accuracy by an average of 2.4 percentage points compared with pure CSA. This effect stems from Lift’s ability to highlight rules whose antecedent and class label co‑occur more often than expected by chance, thereby filtering out high‑confidence rules that are merely artifacts of class prevalence. Second, Conviction and χ² prove valuable in noisy environments (Adult, Letter), where they reduce over‑fitting: models built with these measures contain 12–18 % fewer rules while achieving a modest accuracy gain of about 1.5 percentage points. Conviction’s focus on the probability of the class not occurring when the antecedent does helps discard rules that are statistically significant but practically weak. Third, information‑theoretic measures such as J‑Measure and Gini‑Index lead to more concise models. They lower the average antecedent length from 2.3 to 1.9 attributes without sacrificing predictive power, and they boost accuracy by roughly 0.9 percentage points.

When comparing the hybrid schemes to CSA alone, the authors observe that the former reach or surpass CSA’s performance with far fewer rules. For most datasets, the “CSA + Lift” and “CSA + Conviction” configurations dominate, delivering the highest accuracy, best trade‑off between precision and recall, and the lowest computational overhead. Sensitivity analysis of the interestingness thresholds shows that raising the Lift cut‑off from 1.2 to 1.5 reduces the rule set by about 30 % while improving minority‑class recall by 5 percentage points, illustrating how practitioners can tune the system toward specific business objectives (e.g., prioritizing recall over precision).

Finally, the authors benchmark their best hybrid ordering against established associative classifiers. Using a fixed budget of 500 rules, the “CSA + Lift” approach attains an average accuracy of 84.3 %, compared with 82.1 % for CBA, while cutting training time by roughly 20 % (12.4 s vs 9.8 s) and lowering memory consumption by 15 %. These results confirm that integrating interestingness measures into rule ordering yields tangible benefits in both predictive performance and efficiency.

In conclusion, the study demonstrates that interestingness measures are not merely academic curiosities but practical tools for improving associative classifiers. By augmenting the classic CSA ranking with metrics that capture statistical dependence, information gain, or impurity reduction, one can construct models that are more accurate, more robust to noise, and more parsimonious. The authors suggest several avenues for future work: (1) learning optimal weighting schemes for multiple interestingness scores via meta‑learning, (2) combining the rule‑based framework with deep‑learning‑derived feature representations to handle high‑dimensional data, and (3) extending the approach to streaming environments where CARs must be updated incrementally. This research thus provides a solid empirical foundation for the next generation of interpretable, high‑performance associative classifiers.