Maximum Sensitive Utility–Minimum maxPer Item (MU-MIP)

• MU-MIP is a utility-based privacy-preserving data mining formulation that modifies transactions to hide sensitive periodic high-utility itemsets by violating utility or periodicity constraints.
• It employs a two-stage selection process by first choosing the transaction with maximum sensitive utility and then targeting the item with the minimum maxPer value for modification.
• The approach is further interpreted in varied-threshold utility mining and random utility best-item learning, demonstrating versatility in maintaining data utility and privacy.

Maximum sensitive Utility–Minimum maxPer item (MU-MIP) is a utility-based privacy-preserving data mining formulation for sanitizing transactional databases so that user-designated sensitive periodic high-utility itemsets cease to be discoverable under fixed utility and periodicity constraints. In the literature represented here, the term is defined explicitly as an algorithmic principle in periodic privacy-preserving utility mining, where sanitization selects a transaction with maximum sensitive utility contribution and then modifies the item with minimum item-level maximum periodicity gap, or maxPer, within the sensitive itemset (Zhou et al., 19 Sep 2025). In separate, explicitly interpretive mappings, the same label has also been used to describe varied-threshold utility mining and best-item identification under random utility models, but those usages are not author-defined algorithms in the corresponding papers (Gan et al., 2019, Saha et al., 2020).

1. Formal problem setting in periodic utility mining

In the explicit MU-MIP formulation, the database is $D = \{T_1, T_2, \dots, T_m\}$ over an item universe $I = \{i_1, i_2, \dots, i_n\}$. Each transaction has an identifier, a set of items with quantities, and an implicit timestamp determined by transaction order. Each item $i$ has an external utility $eu(i)$, and its utility in a transaction $Tid$ is

$U(i, Tid) = iu(i, Tid) \times eu(i).$

For an itemset $X$, the transaction-level and database-level utilities are

$$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$

and

$U(X) = \sum_{Tid \supseteq X} U(X, Tid).$

The periodic component is defined through the set of transaction identifiers containing $X$, written $I = \{i_1, i_2, \dots, i_n\}$0, and the gap sequence

$I = \{i_1, i_2, \dots, i_n\}$1

with $I = \{i_1, i_2, \dots, i_n\}$2 and $I = \{i_1, i_2, \dots, i_n\}$3. From this, the framework defines $I = \{i_1, i_2, \dots, i_n\}$4, $I = \{i_1, i_2, \dots, i_n\}$5, and

$I = \{i_1, i_2, \dots, i_n\}$6

where $I = \{i_1, i_2, \dots, i_n\}$7 is the support of $I = \{i_1, i_2, \dots, i_n\}$8. Given thresholds $I = \{i_1, i_2, \dots, i_n\}$9, $i$0, $i$1, $i$2, and a minimum utility threshold $i$3, an itemset $i$4 is a periodic high-utility itemset (PHUI) if it satisfies all of the following: $i$5, $i$6, $i$7, and $i$8 (Zhou et al., 19 Sep 2025).

Sensitive periodic high-utility itemsets (SPIs) are a user-specified subset of the PHUIs. The sanitization objective is to transform $i$9 into a sanitized database $eu(i)$0 such that each sensitive pattern fails at least one of the original discovery constraints. Concretely, a sensitive pattern may be hidden by reducing its utility below $eu(i)$1, by forcing $eu(i)$2 above the allowable bound, or by reducing support sufficiently that the implied $eu(i)$3 exceeds $eu(i)$4 (Zhou et al., 19 Sep 2025).

2. MU-MIP objective and victim-selection rule

MU-MIP is defined by a two-stage selection principle. First, for a sensitive itemset $eu(i)$5, it selects the supporting transaction that contributes the highest sensitive utility:

$eu(i)$6

Second, within that transaction, it selects the victim item in $eu(i)$7 having the smallest item-level $eu(i)$8:

$eu(i)$9

where $Tid$0 denotes the item-level maximum periodic gap (Zhou et al., 19 Sep 2025).

The algorithm tracks three working quantities for each sensitive itemset. The utility surplus to be eliminated is

$Tid$1

The current support is denoted $Tid$2, and the current maximum periodic gap is $Tid$3. Because

$Tid$4

implies

$Tid$5

MU-MIP derives a minimum required support threshold

$Tid$6

Reducing support below this value guarantees violation of the average-periodicity constraint (Zhou et al., 19 Sep 2025).

The modification rule distinguishes deletion from quantity reduction. If the selected item’s utility contribution in the selected transaction is no greater than the remaining sensitive surplus, the item is deleted from that transaction. Otherwise, the quantity is reduced by

$Tid$7

After each modification, the algorithm recalculates the sensitive itemset’s total utility, support, and maximum periodic gap, and stops only when the pattern is no longer a PHUI (Zhou et al., 19 Sep 2025).

This design distinguishes MU-MIP from MU-MAP. MU-MAP uses the same transaction-selection rule but chooses the item with maximum $Tid$8 rather than minimum $Tid$9. The reported consequence is that MU-MAP tends to break periodicity constraints faster and often yields lower Missing Cost, whereas MU-MIP may require more reduction or deletion steps before a sensitive pattern fails the utility or periodicity constraints (Zhou et al., 19 Sep 2025).

3. Data structures, workflow, and complexity

MU-MIP relies on two dedicated data structures. The Sensitive Itemset List (SISL) stores, for each sensitive itemset, the itemset itself, its current total utility $U(i, Tid) = iu(i, Tid) \times eu(i).$0, a list of per-transaction contributions $U(i, Tid) = iu(i, Tid) \times eu(i).$1, its current $U(i, Tid) = iu(i, Tid) \times eu(i).$2 value for $U(i, Tid) = iu(i, Tid) \times eu(i).$3, and its support. The Sensitive Item List (SIL) stores, for each supporting transaction and each sensitive item in that transaction, a 5-tuple of the form $U(i, Tid) = iu(i, Tid) \times eu(i).$4, where $U(i, Tid) = iu(i, Tid) \times eu(i).$5 is the item’s utility in the transaction, $U(i, Tid) = iu(i, Tid) \times eu(i).$6 is the quantity, and $U(i, Tid) = iu(i, Tid) \times eu(i).$7 is item-level $U(i, Tid) = iu(i, Tid) \times eu(i).$8 (Zhou et al., 19 Sep 2025).

The preprocessing stage builds SISL and SIL for all sensitive itemsets, sorts the sensitive set in descending order by itemset length, and computes $U(i, Tid) = iu(i, Tid) \times eu(i).$9. The hiding phase then iterates through the sensitive itemsets. For each one, it initializes $X$0, $X$1, and $X$2, and repeatedly performs three operations: select the highest-contributing supporting transaction, select the minimum-$X$3 victim item, and either delete the item or reduce its quantity. SISL and SIL are updated after every modification, and the process continues until the itemset fails the PHUI constraints (Zhou et al., 19 Sep 2025).

The stated complexity analysis uses the following parameters: $X$4 for the number of sensitive itemsets, $X$5 for average sensitive itemset length, $X$6 for the number of transactions, $X$7 for SISL construction cost, $X$8 for SIL construction cost per item, and $X$9 for the size of the supporting transaction list used in transaction selection. SISL and SIL construction require time and space $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$0; sorting requires $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$1 time and $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$2 space. In the hiding phase, the worst-case time is $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$3, the optimal case is $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$4, and the total space is $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$5 (Zhou et al., 19 Sep 2025).

A notable feature of the stopping logic is that MU-MIP can hide a sensitive pattern by violating any one of three conditions: utility, maximum periodicity, or average periodicity via support. This makes the method a hybrid privacy mechanism rather than a pure utility-reduction heuristic.

4. Evaluation criteria, empirical behavior, and worked example

The evaluation framework uses Hiding Failure (HF), Missing Cost (MC), and Artificial Cost (AC), together with Itemset Utility Similarity (IUS), Database Utility Similarity (DUS), and Database Structure Similarity (DSS). HF measures sensitive PHUIs that remain discoverable after sanitization. MC measures non-sensitive PHUIs lost from the sanitized database. AC measures new PHUIs that did not exist originally. IUS compares the summed utilities of PHUIs before and after sanitization, DUS compares total transaction utility before and after sanitization, and DSS is a cosine-like similarity over the transaction frequency profile (Zhou et al., 19 Sep 2025).

Across the reported experiments on kosarak, retail, mushroom, liquor, chainstore, and foodmart, MU-MIP and MU-MAP achieve $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$6 on all six datasets and tested sensitive percentages. The baselines SMAU, SMIU, and SMSE sometimes produce non-zero AC because they do not model periodicity. The same experiments report that MU-MIP and MU-MAP maintain DUS of over $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$7 after the sensitive itemsets are hidden, and often above $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$8 and frequently above $$U(X, Tid) = \sum_{g \in X \wedge X \subseteq Tid} U(g, Tid),$$9 on some datasets. All algorithms achieve $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$0 by design. On several datasets, MU-MAP yields lower MC than MU-MIP, which is consistent with the rule that selecting the maximum-$U(X) = \sum_{Tid \supseteq X} U(X, Tid).$1 item tends to violate periodic constraints more quickly (Zhou et al., 19 Sep 2025).

The worked example in the source illustrates MU-MIP on a database with ten transactions, thresholds $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$2, $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$3, $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$4, $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$5, and $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$6, with sensitive PHUIs $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$7, $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$8, and $U(X) = \sum_{Tid \supseteq X} U(X, Tid).$9. For $X$0, the initial values are $X$1, $X$2, and $X$3, with $X$4. MU-MIP first selects the transaction with the largest contribution, $X$5, then chooses $X$6 because $X$7 is smaller than $X$8 and $X$9, and deletes $I = \{i_1, i_2, \dots, i_n\}$00 because its utility in $I = \{i_1, i_2, \dots, i_n\}$01 is no greater than $I = \{i_1, i_2, \dots, i_n\}$02. The procedure repeats on $I = \{i_1, i_2, \dots, i_n\}$03, $I = \{i_1, i_2, \dots, i_n\}$04, and $I = \{i_1, i_2, \dots, i_n\}$05, each time deleting $I = \{i_1, i_2, \dots, i_n\}$06, until the support of $I = \{i_1, i_2, \dots, i_n\}$07 becomes $I = \{i_1, i_2, \dots, i_n\}$08, which forces $I = \{i_1, i_2, \dots, i_n\}$09 and hides the pattern through periodicity rather than only through utility reduction (Zhou et al., 19 Sep 2025).

5. MU-MIP as an interpretive lens in varied-threshold utility mining

In a different line of work, MU-MIP is not defined as an algorithm name but can be used interpretively to describe mining with varied item-specific minimum utility thresholds. In the HIMU framework, each item $I = \{i_1, i_2, \dots, i_n\}$10 has an external utility $I = \{i_1, i_2, \dots, i_n\}$11 and transaction-specific quantity $I = \{i_1, i_2, \dots, i_n\}$12, yielding

$I = \{i_1, i_2, \dots, i_n\}$13

For an itemset $I = \{i_1, i_2, \dots, i_n\}$14, the transaction-level utility is

$I = \{i_1, i_2, \dots, i_n\}$15

and the database-level utility is

$I = \{i_1, i_2, \dots, i_n\}$16

Each item has its own minimum utility threshold $I = \{i_1, i_2, \dots, i_n\}$17, and the itemset threshold is

$I = \{i_1, i_2, \dots, i_n\}$18

An itemset is a high-utility itemset iff $I = \{i_1, i_2, \dots, i_n\}$19 (Gan et al., 2019).

Under this interpretation, “Maximum Sensitive Utility” refers to prioritizing rare or sensitive items by assigning them lower item-specific thresholds, and “Minimum maxPer Item” is mapped to enforcing per-item minimum utility constraints while keeping the search tractable through pruning. The HIMU methodology uses the MIU-tree, the global sorted property, the conditional downward closure property, the global downward closure property, utility-lists, and the pruning strategies EUCP and LA-Prune. The one-phase search avoids Apriori-style candidate generation, and the experiments report that HIMU and its enhanced variants are typically one to two orders of magnitude faster than the level-wise baselines HUI-MMU and HUI-MMU_TID, while also reducing memory requirements and addressing the rare item problem (Gan et al., 2019).

This interpretive usage does not describe the same privacy-preserving periodic sanitization procedure as the explicit MU-MIP algorithm. Rather, it recasts varied-threshold utility mining as a framework in which lower thresholds for selected items increase their inclusion probability under the rule $I = \{i_1, i_2, \dots, i_n\}$20.

6. MU-MIP as an interpretive lens in random utility best-item learning

A second interpretive usage places MU-MIP in the framework of best-item identification under Random Utility Models (RUMs). In that setting, each item $I = \{i_1, i_2, \dots, i_n\}$21 has latent utility $I = \{i_1, i_2, \dots, i_n\}$22, and the observed utility is

$I = \{i_1, i_2, \dots, i_n\}$23

with iid additive noise. When a subset $I = \{i_1, i_2, \dots, i_n\}$24 of size $I = \{i_1, i_2, \dots, i_n\}$25 is played, winner feedback returns an item according to

$I = \{i_1, i_2, \dots, i_n\}$26

The framework defines an advantage ratio

$I = \{i_1, i_2, \dots, i_n\}$27

and a minimum advantage ratio

$I = \{i_1, i_2, \dots, i_n\}$28

The key learnability parameter is a distribution-dependent sensitivity constant $I = \{i_1, i_2, \dots, i_n\}$29 such that worst-case pairwise advantage grows at least linearly with the utility gap $I = \{i_1, i_2, \dots, i_n\}$30 (Saha et al., 2020).

Under this mapping, “Maximum Sensitive Utility” means selecting the item with maximal latent utility $I = \{i_1, i_2, \dots, i_n\}$31 when the RUM satisfies the sensitivity condition quantified by $I = \{i_1, i_2, \dots, i_n\}$32, and “Minimum maxPer Item” is interpreted as minimizing the maximum per-item risk or sample budget in hierarchical elimination. The winner-feedback algorithm and its top-$I = \{i_1, i_2, \dots, i_n\}$33 ranking variant use pairwise relative counts, grouping, and iterative elimination, with sample complexity

$I = \{i_1, i_2, \dots, i_n\}$34

for winner feedback and

$I = \{i_1, i_2, \dots, i_n\}$35

for top-$I = \{i_1, i_2, \dots, i_n\}$36 ranking feedback. Matching lower bounds show the dependence on $I = \{i_1, i_2, \dots, i_n\}$37, $I = \{i_1, i_2, \dots, i_n\}$38, and $I = \{i_1, i_2, \dots, i_n\}$39 is order-wise optimal up to logarithmic factors. The analysis does not require independence of irrelevant alternatives, but it does require iid additive noise with density and the stated sensitivity property (Saha et al., 2020).

This usage is explicitly interpretive rather than terminological. It suggests a broad conceptual pattern behind the MU-MIP label: maximize a utility target under a sensitivity condition while controlling the worst-case per-item error or budget.