Submitted:
17 November 2026
Posted:
18 November 2026
You are already at the latest version
Abstract
Background. Retail personalisation must reconcile market propensity, which providers such as NielsenIQ publish for whole generational cohorts, with consumer action, which loyalty programmes record line by line. Without an explicit bridge the two cannot be compared, and recommenders default to echoing past purchases. Objective. This paper specifies and demonstrates the Product Profile Attribute (PPA) taxonomy, a seven-level (L0–L6) instrument that progressively enriches a raw transaction line until account-level behaviour and cohort-level propensity become directly comparable on a common scale, yielding their difference as its terminal output. Method. Following a design-science approach, we specify the taxonomy as a designed artefact—its levels, the L0 crosswalk to ten NielsenIQ categories, the max-normalised action-scoring rule, the gap operator, and a versioned Propensity Scoring Index (PSI) that constructs the market side from published statements under fixed rules—and demonstrate it on a deep single-account omnichannel loyalty record (1,442 product lines on 282 receipts), including a schema-change stress test and sensitivity checks on the observation window, the crosswalk and the thresholds. Results. The taxonomy converts an unstructured receipt stream into a ten-category action vector on a 1–5 scale, aligns it to four cohort propensity vectors and produces a per-category, per-cohort readout that isolates the categories below every cohort's propensity. Every PSI cell is either scored from published statements or reported as needing primary measurement; independent human coding is in progress. The observation window and the crosswalk version are explicit parameters of the instrument. Contribution. The contribution is the instrument: a reproducible, level-by-level taxonomy that makes market propensity and account behaviour comparable from ordinary retail data. Its enrichment is bounded and applied per line, so the pipeline is linear in the number of lines and parallel across accounts; we set out its computational profile and a reference architecture for retail populations of tens of millions of transactions per year.
Keywords:
big data
; scalable data pipeline
; distributed processing
; data variety
; product taxonomy
; attribute enrichment
; propensity–action gap
; recommender systems
; design science
; reproducibility
; retail analytics
; NielsenIQ
1. Introduction
Personalised retail rests on combining what a market is disposed to buy with what a customer does. The difficulty is that these two signals are almost never recorded at the same granularity. Market propensity, when it is available at all, arrives at an aggregate level: data providers such as NielsenIQ publish cohort-level spending statements and projections—Generation Z, Millennials, Generation X, Baby Boomers—across broad retail categories [1,2,3]. Action, by contrast, arrives at the finest possible grain: each loyalty transaction line names a specific stock-keeping unit bought at a specific store at a specific time. Between an aggregate propensity score for “Kitchen” among Millennials and a receipt line for a single saucepan lies a gap of abstraction that no join key can close.
This mismatch has a concrete cost. The intention–behaviour gap, the well-documented divergence between what consumers intend and what they enact [4,5], has a retail counterpart: the distance between what a market is disposed to buy and what an account actually buys. That distance is precisely the signal a forward-looking recommender should exploit, because the categories a market is disposed toward but an account has not yet bought are where assortment, pricing and recommendation can most usefully intervene. Yet computing that distance requires propensity and action to be expressed as comparable quantities. In the absence of a bridge, practitioners abandon the market side altogether and build recommenders on action alone, which structurally limits them to echoing the customer’s purchase history [6].
The setting is also a big-data one. A modern omnichannel retail group of the kind studied here generates transaction records on the order of tens of millions of lines per year, streamed continuously across many store banners and online channels, and—as the demonstration will show—expressed in heterogeneous and changing merchandising schemas. This is the familiar combination of volume, velocity and variety [7,8] that defines big data, and it is exactly what makes the propensity–action bridge non-trivial to build at scale: the enrichment that renders a single line comparable to a cohort benchmark must be applied, consistently and reproducibly, to every line in a very large and schematically messy stream. We therefore treat the PPA taxonomy not only as a conceptual bridge but as a scalable data-enrichment pipeline, and we return in Section 5 to its computational profile and a reference architecture for population-scale deployment.
This paper specifies the bridge as a designed artefact: the Product Profile Attribute (PPA) taxonomy. PPA is a seven-level structure that takes a raw transaction line at one end (L0) and emits a propensity–action gap at the other (L6), enriching the line at each intermediate level with progressively more abstract attributes—product subcategory, functional need, lifestyle signal and value dimension—until the customer’s action can be scored on the same 1–5 scale as the cohort’s propensity. The taxonomy is not a finding about any particular customer; it is reusable infrastructure. Its purpose is to make the propensity–action gap computable from the data retailers already hold. We adopt a design-science framing [9,10], in which the contribution is the artefact together with a demonstration of its utility.
The paper makes three contributions: it specifies the PPA taxonomy as a designed artefact with seven named levels and a terminal gap operator; it demonstrates the artefact end-to-end on a real omnichannel record, including a stress test in which the retailer’s merchandising schema changed; and it establishes the pipeline’s computational profile and a reference architecture, arguing from its per-line design that the instrument can run unchanged from one account to a population of tens of millions of transactions per year; throughput itself is not measured here. After positioning the work against product taxonomies, segmentation, the intention–behaviour gap and big-data enrichment (Section 2), we specify the seven levels, the crosswalk, the action-scoring rule and the gap operator (Section 3), demonstrate the full pipeline on a deep single-account record (Section 4), analyse scalability and deployment architecture (Section 5), and discuss reuse, boundaries and the evaluation agenda (Section 6 and Section 7). The single-account demonstration is reported in fuller behavioural detail in a companion study; here it serves to evidence that the instrument performs its bridging function and scales.
2. Related Work
2.1. Product Taxonomies and Attribute Enrichment
Retailers organise assortments through merchandising hierarchies—department, class, commodity, product group—that are designed for category management rather than for behavioural inference. Such hierarchies describe what a product is but not what need it serves or what value it expresses, and they do not connect to any external propensity benchmark; and, as our demonstration shows, they are not stable, drifting across banners and over time. Attribute-enrichment approaches extend raw product records with derived features, but enrichment is typically pursued for search, catalogue or recommendation-feature purposes and stops at an open-ended feature set. Design-science guidance on taxonomy development [11] stresses that a useful taxonomy is built to a purpose and terminates in a meaningful end; PPA follows this prescription. It differs from generic enrichment in its terminus: enrichment proceeds specifically until a transaction line can be compared with a cohort propensity score, making the gap—not the attribute—the object of the design, and making the taxonomy auditable level by level.
2.2. Segmentation, Personas and the Unit-of-Analysis Problem
Segmentation and persona methods translate customer data into actionable groups and archetypes [12,13,14], and generational-cohort segmentation in particular is widely used to summarise consumer propensity [1]. These methods are powerful but presuppose that the descriptors needed to characterise a segment are available and commensurable. The recurring obstacle in practice is the unit-of-analysis mismatch: behavioural data are account-level and fine-grained, while reference benchmarks—cohort propensity profiles above all—are aggregate and coarse, and the demographic keys that would ordinarily join them are frequently missing or untrustworthy in loyalty data. PPA addresses this obstacle directly by providing a documented, level-by-level path from the individual transaction line to the aggregate benchmark, so that the two can be subtracted rather than merely juxtaposed, and so that an account’s cohort affinity can be read from behaviour when it cannot be read from the registration form.
2.3. From the Intention–Behaviour Gap to a Computable Retail Quantity
The intention–behaviour gap is a robust behavioural construct, consolidated by meta-analytic review [4] and grounded in the reasoned-action tradition [5,15]. It is, however, almost always measured through self-report on both sides, which limits its operational use in a retail setting where the action side is continuously and objectively recorded. The methodological move this paper supports is to ground the action side in observed purchasing and to express it on the same bounded scale as a cohort-level propensity benchmark, turning an abstract psychological construct into an operational, per-category retail signal that a recommender can consume [6]. What has been missing is not the construct but the bridge that makes it computable at the grain retailers actually hold; supplying that bridge, as a reusable and scalable artefact, is the gap this paper fills.
2.4. Enrichment Pipelines at Scale
The engineering counterpart to attribute enrichment is the large-scale feature or ETL pipeline, in which raw records are transformed through successive stages before analysis. Big-data practice favours transformations that are bounded per record and composable, because these map cleanly onto data-parallel execution models such as MapReduce [16] and in-memory cluster computing [17], and onto columnar storage for analytical scans. PPA is deliberately of this form: each level consumes the previous level’s output and adds one bounded transformation, so the whole pipeline is a linear chain of per-line operations plus a small per-account aggregation. This is what allows a taxonomy designed on one account to be executed, unchanged, across a retail population—a property made explicit in Section 5. Where general-purpose enrichment pipelines stop at features for search or modelling, PPA’s terminus is the propensity–action gap, giving the pipeline a defined analytical output rather than an open-ended feature set.
3. The PPA Taxonomy
PPA is specified as a seven-level enrichment pipeline. Each level consumes the output of the level below and adds one layer of abstraction; the levels are ordered so that the final two render action and propensity commensurable and then subtract them. Table 1 states the seven levels, their names and the role each plays in the bridge. The level labels are those of the canonical PPA crosswalk used throughout the wider doctoral project.
Development method. The taxonomy follows the principle of the taxonomy-development method of Nickerson et al. [11] rather than its dimension-and-characteristic form. The meta-characteristic is fixed by the purpose: every attribute must move a transaction line closer to the frame in which cohort propensity is published, so that propensity and action become commensurable. PPA is a hierarchical ladder, each level an abstraction of the one below, not a set of orthogonal dimensions; mutual exclusivity and collective exhaustiveness are therefore required, and tested, at L0, where every line belongs to exactly one of the ten categories or to the explicit out-of-scope residual. The objective ending conditions are that every line in the demonstration window maps to one L0 class or to the residual, that no L0 class is empty, that lines of the same commodity are never assigned to different classes, and that the last iteration neither added nor split a class. The subjective ending conditions are that the ladder is concise (seven levels), robust to a change in the retailer’s schema (Section 4.5), comprehensive for the ten-category frame, extendible through versioned tables, and explanatory, in that every level states its role in the bridge (Table 1). The levels were set conceptual-to-empirical from the market-side frame (ten NielsenIQ categories on a 1–5 scale); the L0 crosswalk was then refined empirical-to-conceptual on the demonstration data in two versioned iterations: version 1 (64 rows), built when the retailer’s hierarchy changed, and version 2 (67 rows), adopted after a check of version 1 against the category scope of the index (Appendix B, Table B1) found mixed departments and seasonal decoration placed inconsistently with it. Version 2 met the objective conditions without changing any level, which ended the iteration.
3.1. The Category Crosswalk (L0)
The base level maps heterogeneous transaction lines onto a fixed set of ten NielsenIQ retail-group categories: Furniture, Electronics & Gadget, Cleaning & Household, Home Improvement, Hobby & Lifestyle, Kids & Toys, Kitchen, Health & Sports, Automotive, and Bed & Bath. The crosswalk is documented rather than inferred, reconciled to the raw data and versioned (version 2, 67 rows; Appendix A), so that the mapping is auditable. Because all downstream levels inherit this category space, the crosswalk is the single most consequential design decision in the taxonomy and is treated as a versioned artefact. Being the load-bearing decision, it is specified to be robust to changes in the retailer’s own merchandising schema: in the demonstration below the source extract replaced the retailer’s top-level ‘world / department-group’ hierarchy with a flat department schema, and the L0 crosswalk was rebuilt from department codes and reconciled to the ten-category space (Appendix A) without disturbing any downstream level. Lines whose products fall outside the ten NielsenIQ categories—here fashion, food-and-beverage, cosmetics and personal-care items, about 7% of item spend—are assigned to an explicit out-of-scope residual rather than force-fitted, keeping the category space exhaustive and its coverage quantified. Home decoration and home fragrance are assigned to Home Improvement, following the provider’s own grouping of decor with home improvement (Appendix A). Two implementation rules are part of the specification: the crosswalk is keyed on the pair (business unit, department code), because a retail group reuses codes across banners (in the demonstration data AA is Cleaning Supplies in one banner and Babies in another); and every share used downstream is computed over in-scope lines and receipts, after the residual has been removed, so that the residual quantifies coverage without diluting the action score.
The lifestyle-signal level (L3) also carries a role on the market side that the demonstration made explicit. Published market statements are frequently about experiences rather than products—‘spending more on at-home experiences’ is the canonical example—and such a statement names no category. In the PPA ladder an experience is a lifestyle signal: it enters at L3, and it reaches L0 only through the product basket that the L3 dictionary associates with that signal (a blanket bought for watching television is neither ‘Bed & Bath’ nor ‘Electronics’ as a disposition; it is a home-entertainment signal whose basket contains both).
3.2. Action Scoring (L5)
The unit of analysis is the loyalty account rather than a verified individual: the account is what the retailer observes and what a recommender addresses, regardless of who in the household presents the card. To express account-level action on the cohort propensity scale, each category’s spend share s(c) and trip penetration t(c) are max-normalised across the account’s own categories to s̃ and t̃, then combined into an action score:
A(c) = 1 + 4 × ½ ( s̃ + t̃ )
Averaging spend and penetration requires a category to be both valuable and frequent to score highly, preventing a single large basket from dominating; max-normalisation expresses action relative to the account’s own behaviour, which is what permits comparison against the externally fixed 1–5 propensity scale. The rule is intentionally simple and transparent so that the gap it feeds is interpretable. Three properties govern how its output may be read. A(c) is anchored on the account’s own largest category, so it is a relative preference profile within the account whereas I(c,k) is an absolute cohort benchmark: the two share a ruler but not an anchor, and the L5 commensuration step is what makes their subtraction meaningful rather than literal. The rule has no time dimension—every line in the observation window counts equally—so the window is a parameter of the instrument and must be reported with every result (Section 4.4). And trip penetration structurally favours frequently bought categories over durable ones, so the rule reads where an account’s habitual attention lies; category-specific replacement cycles are left to the population phase.
3.3. Propensity Construction (L5): the Propensity Scoring Index
The cohort propensity score I(c,k) is not published by any provider in the form the taxonomy needs: NielsenIQ and comparable providers publish net-intent percentages, spending projections and per-generation exhibits [2,3], never a category-by-cohort table on a bounded scale. The score used at L5 is therefore constructed, and the first version of the instrument constructed it by the author’s reading of the reports—legitimate for a demonstration, but not reproducible by another analyst and silent about which cells rest on evidence. A terminological point follows from this and is observed throughout the paper: because the market side is an aggregate published about a cohort rather than a measurement taken from a person, it is reported as propensity, and the term intention is reserved for the construct as the theory of planned behaviour defines it. The Propensity Scoring Index (PSI) replaces the author’s reading with a fixed, versioned procedure that separates three tasks: quoting evidence into an append-only log (one finding per report, with source, page, geography and the cohort named); coding each item under a closed rubric; and computing the score from the codes. Only the coding retains judgement, and it is bounded so that it can be tested for inter-coder agreement [18,19]. The coder decides whether an item names a category directly (D) or a product construct adjacent to it (P), which cohort it concerns (or ALL), and to which paying cohort it accrues; statements that name no product construct—attitude and situational statements, channel and trust statements, spending capacity—are excluded by a closed list, because they moderate conversion from disposition to action and belong to the gap-decomposition model rather than to I. Statements about experiences inside the home are coded as evidence for the in-home lifestyle signals of L3 and reach L0 through each signal’s product basket (Appendix B), so that the market side stays inside the taxonomy. Each row then receives a signal value v ∈ {−1, −0.5, 0, +0.5, +1} by fixed thresholds on the form of the excerpt (proportion, net intent, per-cohort figures, qualitative wording or growth rate), and a weight w = w_geo × w_link × w_coh (target market 1.0, region 0.75, global 0.5; D 1.0, P 0.5; cohort-specific 1.0, ALL 0.5). For each cell the weighted mean direction m(c,k) = Σw·v ÷ Σw is mapped to the propensity scale by
I(c,k) = 3 + 2 · m(c,k), rounded to the nearest 0.5 and clipped to [1, 5]
so that 3 is the neutral point and the index sits on the same 0.5 lattice as the original scoring model. A cell whose total weight falls below w_min = 1.0—two direct global statements about that cohort, or one about the target market—is not scored but reported as INSUFFICIENT, so that its value comes from primary measurement rather than inference; a cell with positive and negative weight of at least 0.5 each is flagged TENSION. The five parameters are fixed before scoring and varied in a sensitivity analysis (Appendix B). The index is versioned as tables read by a fixed routine, so a change to a table is a new index version and a new provider release is a new log version; reproducing the market side is then a matter of applying the same tables to the same log, and refreshing it at population scale is a table update rather than a re-reading. The index is specified so that inter-coder agreement can be established; establishing it with independent human coders is an evaluation item (Section 6.1) and is not claimed here.
3.4. The Gap Operator (L6)
The terminal level subtracts action from propensity for every category–cohort pair, G(c,k) = I(c,k) − A(c). The sign and magnitude carry the meaning: a large positive gap marks a category the cohort is disposed toward but the account under-acts upon—latent demand and a recommendation target—while a negative gap marks an over-served category. Classifying gaps into severely under-served, under-served, aligned and over-served bands turns the continuous matrix into an actionable agenda. In the demonstration instrument reported here those bands are fixed by two constants on the shared 1–5 scale: a category is read as aligned when |G| ≤ ε with ε = 0.30, as severely under-served when G ≥ τ with τ = 2.20, and otherwise as under-served or over-served according to the sign of G. The constants are properties of this instrument and this scale rather than universal thresholds, and a different propensity scale requires them to be re-derived. Because the gap is computed against all cohorts, opportunities that hold regardless of which cohort the account is assumed to belong to can be identified, which is valuable precisely when demographic data are unreliable.
3.5. Worked Illustration
To make the enrichment concrete, Table 2 traces a single line—a cordless power drill—through all seven levels. At L0 it is mapped to Home Improvement; at L1 to the power-tools subcategory; at L2 to the functional need of home maintenance; at L3 to a DIY-enthusiast lifestyle signal; at L4 to a self-reliance value dimension. At L5 the account’s Home-Improvement action score (4.61) is placed beside the cohort propensity for that category (for Gen X, 4.5), and at L6 the two are subtracted to a gap of −0.11—aligned for that cohort (|G| ≤ 0.30). Every line follows the same mechanical path, and nothing in it depends on a demographic field or on the identity of the account holder.
Algorithm 1. PPA enrichment of one account’s lines to a gap matrix
Input: account m’s transaction lines L_m ; reference tables R (L0–L4) ; cohort propensity I(c,k)
for each line ℓ in L_m:
c ← L0_crosswalk(ℓ.department) # hash lookup; unmapped → OUT_OF_SCOPE
attach L1–L4 attributes from R
aggregate over L_m for each category c:
s(c) ← spend share ; t(c) ← trip penetration
s̃, t̃ ← max-normalise s, t across categories
A(c) ← 1 + 4 × ½ (s̃ + t̃) # L5 action score, 1–5
for each category c, cohort k:
G(c,k) ← I(c,k) − A(c) # L6 propensity–action gap
Output: per-account gap matrix G (10 categories × 4 cohorts)
4. Demonstration
We demonstrate the full pipeline on a single, unusually deep omnichannel loyalty record from an Indonesian retail group (276 receipt-level transactions across six banners, enriched by a complete item-level extract of 1,442 product lines across 282 receipts spanning November 2023 – July 2026). The demonstration is an existence proof of utility in the case-demonstration tradition [20], not a behavioural study; its purpose is to show that the artefact transforms raw data into a gap matrix as designed.
4.1. Data and Reconciliation
The item-level extract spans April 2021 to July 2026; restricting it to the loyalty-active window of November 2023 to July 2026 yields exactly 1,442 lines on 282 receipts. Two identifiers coexist in the source and must be reconciled explicitly: the point-of-sale system records 282 receipt identifiers (280 parent receipts, two of which were split at the till), while the loyalty system records 276 transactions in the same window; the difference runs both ways—16 item-level receipts have no loyalty transaction (15 in the toy banner and one return), and 10 loyalty transactions, all at one food-service outlet, carry no item detail—and is a property of the two systems, not an error in either. One line is a return (a dining set, quantity −5) and is retained so that net spend is correct. The 52 departments that occur in the window (a 53rd, a single special-transaction line of September 2023, lies outside it) were mapped as in Appendix A. After the L0 crosswalk (version 2), 1,148 lines on 261 receipts fall inside the ten-category space and carry Rp 245.9 million of net spend; the remaining 294 lines, including 21 receipts that contain nothing in scope, are the 7.0 per cent residual of Appendix A. All shares in Section 4.2 are computed over these in-scope quantities.
4.2. End-to-End Run
Run end to end, the taxonomy mapped each transaction line through L0–L4, scored the account’s action across the ten categories at L5, joined the four cohort propensity vectors of the original scoring model, and emitted the gap matrix at L6. The instrument cleanly separated the account’s action profile—concentrated in Home Improvement, Hobby & Lifestyle, Cleaning & Household and Furniture—from categories that scored low on action, and the gap operator then exposed six categories (Furniture, Electronics & Gadget, Bed & Bath, Kitchen, Health & Sports and Automotive) as under-served against every one of the four cohorts. The point for the present paper is not which categories emerged but that the artefact produced a well-formed, interpretable gap matrix from an unstructured receipt stream: the bridge held.
Table 3 records the demonstration output. It is included to make the artefact’s product concrete—a ten-category action vector aligned to cohort propensity and reduced to an interpretable gap range—rather than to advance a behavioural claim, which is the companion study’s task. Six categories are under-served against every cohort, Electronics & Gadget only narrowly; none is over-served against every cohort, and the remainder are cohort-dependent. Whatever the substantive reading, the point for the present paper is that a well-formed, per-category, per-cohort matrix was produced from an unstructured receipt stream, which is the whole of the claim advanced here.
4.3. Constructing the Market Side with the Index
Table 3 uses the first-version propensity scores, the author’s scoring model, with crosswalk version 2: the action side was re-run on 11 November 2026 when the crosswalk changed (Section 4.4), while the propensity side remains the dated original until the independently coded index replaces it. The index of Section 3.3 is applied to the same 56-item evidence log by coders independent of the authorship, and its result, the cells that reach the evidence threshold, the cells reported as INSUFFICIENT, and the comparison with the original model cell by cell, is reported once that coding is complete; it is not reported here, because an index result produced by the authors would not be the independent construction the index exists to provide. What the index already fixes, before any result, is the boundary the reader will be able to see: every cell is either scored from quoted statements under the rubric or declared unmeasured and handed to primary measurement, and no cell is filled by life-stage inference.
4.4. Sensitivity: Observation Window, Crosswalk and Thresholds
Because the action-scoring rule has no time dimension (Section 3.2), the observation window is a parameter, and the demonstration account makes the consequence concrete: its Hobby & Lifestyle score of 4.40 is carried almost entirely by bulk pet-food purchases in one commodity (Rp 71.3 million of the category’s Rp 78.8 million, on 184 lines) that run from late 2023 to mid-2025 and then all but stop, with Rp 0.4 million in the final twelve months. Restricting the window to those twelve months (August 2025 to July 2026; 626 lines, 122 receipts) re-anchors the account: Home Improvement leads on both axes and scores 5.00, Hobby & Lifestyle falls to 2.30, and the set of categories under-served against every cohort contracts from six to five—Furniture, Kitchen, Health & Sports, Bed & Bath and Automotive—because Electronics & Gadget falls into the aligned band against Boomer (Table 4). The instrument is not wrong in either window; it answers a different question in each, so the window must be stated with every gap matrix, as the versioned crosswalk already is.
The crosswalk is the second parameter. Version 2 of the L0 crosswalk, adopted on 11 November 2026 after a check of version 1 against the index’s category scope, splits Trendy Goods by commodity and assigns all home decoration and home fragrance to Home Improvement. Against version 1 it moves Hobby & Lifestyle from 4.92 to 4.40, Home Improvement from 4.26 to 4.61, Cleaning & Household from 3.18 to 2.85 and Electronics & Gadget from 1.89 to 2.14, and raises the out-of-scope residual from 6.2 to 7.0 per cent. The six categories under-served against every cohort are unchanged, but Hobby & Lifestyle is no longer over-served against all four cohorts, and Electronics & Gadget now clears the aligned band against Boomer by only 0.06. Both versions are reported so that the change stays traceable. The thresholds are the third parameter: with ε varied over [0.25, 1.02], the six-category set of Table 3 holds up to ε = 0.36, beyond which Electronics & Gadget leaves it, and the other five hold up to ε = 1.01. Taken together, the three checks locate the instrument’s sensitivity where its design places it—in the window, the crosswalk and the propensity construction—and not in the arithmetic of the levels.
4.5. Stress Test: Schema Change at L0
The demonstration also stress-tested the artefact. Because the source extract had dropped the retailer’s top-level ‘world’ hierarchy, the L0 crosswalk was rebuilt from department codes and reconciled to the ten-category space, and roughly 7% of item spend fell outside those categories and was reported as an explicit out-of-scope residual—evidence that the instrument tolerates a real change in source schema and quantifies its own coverage. Two design properties frame the remaining sensitivity points: the quality of the L0 mapping governs everything downstream, and the asymmetry between account-level action and cohort-level propensity is a deliberate property, not a defect, that the L5 commensuration step is built to absorb. Each is revisited in the evaluation agenda.
5. Scalability and Deployment Architecture
A methods contribution for big retail data must show not only that the artefact works but that it scales. This section sets out the pipeline’s computational profile, a reference architecture for population-scale deployment, and the path from the single-account demonstration to the full account population.
5.1. Computational Profile
The pipeline’s cost is dominated by per-line operations that are independent across transaction lines and across accounts. Levels L0–L4 are per-line lookups and mappings: the L0 crosswalk is a hash join against a small versioned dictionary (constant time per line), and L1–L4 attach attributes from bounded reference tables. Level L5 is a per-account aggregation—spend share and trip penetration are group-by sums over the account’s lines, a single linear pass—followed by max-normalisation over the ten categories (constant work). Level L6 subtracts two 10×4 matrices (constant). The whole pipeline is therefore O(N) in the number of transaction lines N, with a constant per-account overhead, and it is embarrassingly parallel: accounts share only the read-only reference tables, so lines can be partitioned by account and processed independently. Table 5 summarises the per-level cost and the corresponding data-parallel pattern, where nₐ denotes the number of transaction lines belonging to account a.
5.2. Reference Architecture for Population Scale
At population scale—of the order of tens of millions of transaction lines per year across banners and channels—the pipeline is deployed as a batch (or micro-batch) enrichment over a columnar transaction store. The versioned L0–L4 reference tables are broadcast to every worker; transaction lines are partitioned by account; each partition is enriched independently and reduced to a per-account action vector; cohort propensity vectors—produced by the index of Section 3.3 and refreshed as a versioned table whenever a provider publishes—are joined in at L5; and the L6 gap matrices are written back keyed by account. Because the reference tables are small and read-only and the per-account work is bounded, throughput scales linearly with added workers, and incremental arrivals (new receipts) are folded in by re-reducing only the affected accounts rather than reprocessing the corpus. The crosswalk’s robustness to schema change (Section 3.1 and Section 4) is first-class here: at scale the retailer’s merchandising hierarchy drifts over time and across banners, so the L0 dictionary is versioned and the out-of-scope residual is monitored as a coverage metric rather than assumed away. Figure 1 sketches the architecture. The absolute throughput of such a deployment is not measured here, and no figure is claimed: the argument of this section is structural rather than empirical. What the analysis establishes is the shape of the cost, linear in the number of lines and parallel across accounts, so that wall-clock time is governed by the number of workers applied rather than by the size of the corpus. The corresponding measurement, lines per second and wall-clock over a full population pass, is stated as an open item in Section 6.1. The incremental case is bounded by construction: because only accounts with new receipts are re-reduced, a daily update touches a share of the corpus proportional to the share of accounts transacting that day. Storage follows the same partitioning: a columnar layout keyed by account makes the per-account scans and group-bys sequential, and the small versioned reference tables are cached in memory on every worker. Neither the levels nor the scoring rule change with scale; only the execution substrate does. Velocity is handled by the same partitioning: because an account’s gap matrix depends only on that account’s lines and the read-only benchmarks, gap matrices can be refreshed in near-real time as receipts stream in, updating only the accounts touched—so the bridge supports both periodic population passes and continuous per-account updates.
5.3. From One Account to the Population
The demonstration fixes every design decision the population study will inherit: the category space, the enrichment levels, the action-scoring rule and the gap operator are all defined and exercised here. Scaling to the population changes the inputs, not the instrument. Three extensions are anticipated. First, the market side: the cohort-level propensity benchmark used here is retained as I, while intention measured by survey is a separate construct, written INT, entered alongside it at L5, so that the readout can be computed against either (G = I − A or G_INT = INT − A) without disturbing other levels. Second, segmentation: clustering accounts by their L6 gap vectors yields gap-based hyper-personas, the subject of the population-scale phase of the wider doctoral programme. Third, recommendation: the gap matrix is the native input to a gap-weighted recommender, whose engagement uplift is evaluated against action-only baselines. Each reuses the same pipeline; none requires it to be rebuilt. What the population phase adds is not a new bridge but evidence at scale—throughput, stability of the gap matrix under a moving schema, and downstream predictive utility—identified in the evaluation agenda below.
A fourth extension rides in the same pass, at a cost that is small but not free: an account-composition indicator. Because a loyalty card may be presented by more than one household member, each account can be scored on five per-account probes, physical simultaneity of in-store receipts (subject to a geographic-feasibility check), payment-instrument diversity, merged legacy identifiers, velocity-flag share and multi-store-day share, combined into a weighted indicator H(a) ∈ [0,1]. Four of the five are per-account aggregations of the same form as L5 and add O(nₐ) work to the reduce step already performed there. The simultaneity probe is the exception: it compares an account’s receipts pairwise within a day and across stores, which is O(nₐ log nₐ) once the account’s lines are sorted by timestamp, and it requires a store-coordinate table to be added to the broadcast set so that the geographic-feasibility check can be evaluated locally on each worker. The indicator therefore preserves the partition-by-account structure and the linear-in-accounts scaling of the pipeline, but it is an addition to the reference tables and to the per-account work, not a free rider. The indicator then acts as a serving rule for the L6 output: composition-invariant quantities (the universally under-served categories, computed against all cohort profiles simultaneously) may be served to any account, while composition-sensitive readings (persona narratives, cohort-affinity assignments) are released only with the account’s H(a) band attached. Weights and bands are calibrated at population scale against labelled accounts; the demonstration account scores in the likely-household band under provisional settings, which is precisely the case the serving rule exists to handle.
6. Discussion
The PPA taxonomy is offered as reusable infrastructure rather than as a result. Its value is that it makes a previously incomparable pair of signals mutually computable from ordinary retail data, and does so through an auditable, level-by-level path that any analyst with comparable inputs can reproduce. Three properties make it portable. It is provider-agnostic: any cohort- or segment-level propensity benchmark expressed on a bounded scale can be substituted at L5 for the NielsenIQ profiles used here. It is demographics-robust: because the gap can be evaluated against all cohorts, it yields actionable targets even when the customer’s true cohort is unknown. And it is recommender-ready: the L6 gap matrix is exactly the input a gap-aware ranking function requires.
The artefact also has clear boundaries. The crosswalk at L0 encodes judgement and must be versioned and validated; the action-scoring rule at L5 is one defensible choice among several and invites sensitivity analysis; and the commensuration of account-level action with aggregate propensity, while deliberate, sets a ceiling on precision that only individual-level intention measurement could lift. Two further boundaries were made visible by the demonstration rather than assumed. The market side is constructed: the instrument’s output is only as complete as the evidence base it is fed, and the index’s chief service is to say which cells rest on published statements and which are absent. And the action side is windowed: the rule carries no clock, so the matrix is a statement about a stated period, and the period is part of the result. These are not reasons to withhold the instrument but the agenda for evaluating and improving it.
6.1. Evaluation Agenda
Consistent with design-science practice [10], we identify how the artefact should be evaluated beyond demonstration. Robustness: vary the L0 crosswalk and the L5 scoring rule and measure the stability of the resulting readout matrix (partly evidenced here—the crosswalk survived a change in the retailer’s source hierarchy with a quantified out-of-scope residual). Scalability: benchmark throughput (lines per second) and wall-clock across the full account population, of the order of tens of millions of lines per year, confirming the linear, partition-by-account scaling argued in Section 5 and measuring the stability of the readout under a drifting merchandising schema. Predictive utility: test whether the readout improves a downstream recommender against action-only baselines. Substitutability: replace the propensity benchmark with an alternative provider and confirm the bridge still holds. Coding reliability: have the index applied independently by two human coders external to the authorship and report weighted κ on the signal value and agreement on the direct/proxy decision; re-run the index on each provider release and report the cells that change. This is the outstanding item: the index is specified to make such a test possible, and the test itself has not yet been run. Window stability: at population scale, compute the readout over rolling windows and report the share of accounts whose below-propensity set is invariant. Each turns the instrument from a demonstrated artefact into an evaluated one.
6.2. Boundaries and Threats to Validity
Three boundaries qualify the contribution. The demonstration is a single-account existence proof: it evidences utility and scalability-by-construction, not generalisation, and the population evaluation set out above remains outstanding. The L0 crosswalk encodes human judgement and must be versioned and validated; its coverage is monitored through the out-of-scope residual, but a mis-mapping at L0 propagates to every downstream level. Finally, commensurating account-level action with cohort-level propensity is deliberate but sets a ceiling on precision that only individual-level intention measurement could lift. The propensity scores of Table 3 are the original scoring model, reported for continuity; the independently coded index result of Section 4.3, once complete, supersedes it as the reproducible construction, and the two will be placed side by side so that the reader can see exactly which claims rest on published evidence. None of these is a reason to withhold the instrument; together they define the agenda for evaluating and improving it. A further scope condition: the demonstration account may be operated by more than one household member — a companion paper examines this empirically — but the instrument is indifferent to it, since enrichment, action scoring and the gap operator are defined on the account’s transaction lines regardless of who produced them, and the account is the unit a deployed recommender addresses. The demonstration account is the first author’s own; 23 of its 33 months predate the written intention–action framing of October 2025, while the remaining ten may reflect the holder’s awareness of it and are a limitation of the demonstration; nothing in the instrument itself depends on who the holder is.
7. Conclusion
Recommenders fail to act on what a market is disposed toward because propensity and action are recorded at incompatible levels and are rarely bridged. The Product Profile Attribute taxonomy is a designed bridge: seven levels that carry a raw transaction line up to the frame on which published cohort propensity is reported, so that the two can be differenced per category and per cohort. The worked example shows the bridge carrying real data end to end, and the versioned index shows the market side being constructed by rule rather than by reading, with the cells it cannot support reported as unmeasured rather than filled by inference.
Supplementary Materials
The following supporting information can be downloaded at the website of this paper posted on Preprints.org.
Funding
This research received no external funding.
Use of artificial intelligence
During the preparation of this manuscript the authors used generative AI tools for language editing, for drafting data-processing scripts, and for locating and transcribing candidate passages from the market reports, each of which the authors verified against its source page. The first-version propensity scores of Table 3 were assigned by the authors. During the development of the Propensity Scoring Index, exploratory applications of the rubric with AI assistance were used to find ambiguities in the rules; no result of those applications is reported in this paper, and the index result will be produced by human coders independent of the authorship (Section 6.1). The verification of reported figures against source pages and raw data was AI-assisted and spot-checked by the first author. All content was reviewed and edited by the authors, who take full responsibility for it.
Companion work: The same anonymised record is used, for a disjoint purpose, in a companion single-case study demonstrating the measurement bridge between market-level propensity and account-level transaction behaviour on this account. The two papers share data and cross-reference accordingly.
Conflicts of interest
The first author is employed by the retail group that owns the loyalty and point-of-sale data used in the demonstration. The group had no role in the design of the study, in the analysis or interpretation of the data, in the writing of the manuscript, or in the decision to publish. The remaining authors declare no competing interests.
Appendix A. Department-to-Category Crosswalk
The item-level extract used in the demonstration replaced the retailer’s world and department-group hierarchy with a flat department schema. Each of the 52 departments that occur in the analysis window was mapped to one of the ten NielsenIQ categories, or to an explicit out-of-scope residual, using crosswalk version 2 and the category scope of Table B1; for the four departments whose commodities span more than one category (Travels, Seasonal, Baby & Kids; Outdoor Comfort; Pendopo; Trendy Goods) the assignment was made at commodity level. Home decoration (indoor, outdoor and seasonal) and home fragrance are assigned to Home Improvement, following NielsenIQ’s Home Improvements–Decor grouping [21]. The full extract contains one further department, a single special-transaction line dated September 2023, outside the window. The residual, 7.0 per cent of item spend, comprises fashion, food-and-beverage, cosmetics, personal-care and non-merchandise lines, which were reported as out of scope rather than force-fitted. Table A1 summarises the mapping by category; the full crosswalk is a supplementary table.
Table A1.
Summary of the department-to-category crosswalk, version 2, by share of item-level spend.
| Category | Share of item spend | Contributing departments (code, name) |
| Hobby & Lifestyle | 29.8% | AR Lawn Garden; AS Outdoor Living; AU Aqua Pet; AY Travels, Seasonal, Baby & Kids (AY03 Travels); BH Trendy Goods (BH05 Student Supplies) |
| Home Improvement | 24.0% | AC Paint & Sundries; AD Building Supplies; AE Ladder; AG Tools; AH Electrical; AJ Lighting; AK Plumbing; AL Sanitary Ware; AM Hardware; AN Safes, Office & Security Systems; AT Outdoor Comfort (AT11 Outdoor Decoration); AY Travels, Seasonal, Baby & Kids (AY04 Seasonal); BD Home Decor & Fragrance; BJ Pendopo (BJ17) |
| Furniture | 10.8% | AP Home Storage; AT Outdoor Comfort (AT10 Outdoor Furniture); BN Locker, Cabinet & Racking; CA Sofa; CD Ruang Tamu & Penyimpanan; CE Ruang Makan & Dapur; CG Ruang Tidur; CI Ruang Kerja & Belajar |
| Cleaning & Household | 9.1% | AA Cleaning Supplies; BJ Pendopo (BJ02, BJ06 Household); CH Perlengkapan Rumah |
| Electronics & Gadget | 5.9% | AB Cleaning Appliances; AO Appliances; BH Trendy Goods (BH03 Gadget & Accs); BM Fans & Air Treatment |
| Kitchen | 3.6% | AF Horecaba Supplies; BC Kitchenware; CF Peralatan Makan & Dapur |
| Health & Sports | 3.4% | AV Sports; AZ Health Supports; BJ Pendopo (BJ12 Beauty & Wellness) |
| Kids & Toys | 2.8% | AA Babies; AB Boys; AC Girls; AE Family Fun; AQ Babies Consignment; AR Boys Consignment; AS Girls Consignment; AT Education Consignment; AU Family Fun Consignment; AW International; AY Travels, Seasonal, Baby & Kids (AY09 Baby & Kids) |
| Bed & Bath | 2.6% | BE Home Textile; CB Textile |
| Automotive | 0.9% | AW Automotive |
| Out-of-scope (residual) | 7.0% | AY Travels, Seasonal, Baby & Kids (AY06 Foods & Beverages); BA Non Merchandise; BG Service; BH Trendy Goods (BH04 Beauty, BH13 Fashion & Accs); BJ Pendopo (BJ03, BJ05, BJ13, BJ15, BJ18); BK Promo Event; BL Delivery & Installation |
Appendix B. The Propensity Scoring Index, Version 1.3
The index is a set of six tables read by a fixed scoring routine. Tables B1–B2 reproduce the tables that carry judgement; the signal rubric and parameters are stated in Section 3.3. Version history: v1.0 (9 September 2026) introduced the rubric, weights and threshold; v1.1 added the closed exclusion list, the product-proxy rule and the category scope, after exploratory applications exposed ambiguity about which statements count as evidence; v1.2 replaced an even spread of at-home statements across in-home categories with the L3 basket rule of Table B2; v1.3 (11 November 2026) aligned the category scope with crosswalk version 2, moving home décor and fragrance to Home Improvement, gadget accessories to Electronics & Gadget and artificial plants out of Hobby & Lifestyle (Table B1, rows 1, 3, 5 and 9), with the corresponding change to Table B2. No score changed, because the index has not yet produced scores. The evidence log draws on four NielsenIQ reports [2,3,22,23] and four supporting sources listed in the supplementary files. Parameters were unchanged across versions. The index was renamed from Intention Scoring Index at v1.2 to match the terminological scope stated in Section 3.3; the tables and the routine are unchanged by the rename. Inter-coder reliability has not yet been established: the rubric is closed and the parameters fixed precisely so that an independent coding test is possible, and that test, with human coders external to the authorship, is listed as an evaluation item in Section 6.1.
Table B1.
Category scope (index table T0, version 2). Excluded statement types (T1): attitude and situational statements naming no product construct; channel and trust statements; cohort spending-capacity statements. Product-proxy rule (T2): a statement enters with link P only when it names a product construct adjacent to a category.
Table B1.
Category scope (index table T0, version 2). Excluded statement types (T1): attitude and situational statements naming no product construct; channel and trust statements; cohort spending-capacity statements. Product-proxy rule (T2): a statement enters with link P only when it names a product construct adjacent to a category.
| # | Category (L0) | Scope |
| 1 | Home Improvement | Tools, building materials, paint, electrical, lighting, plumbing, sanitary, hardware, ladders, home security and home-office; small DIY jobs; home décor (indoor, outdoor and seasonal, incl. artificial plants) and home fragrance |
| 2 | Furniture | Tables, chairs, sofas, bed frames, cabinets, racks, home storage, outdoor furniture |
| 3 | Electronics & Gadget | Phones, TVs, laptops, audio, wearables; large and small appliances incl. electric kitchen appliances; smart home; gadget accessories |
| 4 | Kitchen | Cookware, tableware, food storage, non-electric kitchen tools |
| 5 | Cleaning & Household | Cleaning supplies, tissues, soap, laundry, bins, brushes, small household goods |
| 6 | Bed & Bath | Sheets, blankets, pillows, towels, curtains, home textiles |
| 7 | Health & Sports | Sports equipment, bicycles, fitness, health supports, body care and wellness |
| 8 | Kids & Toys | Toys, baby and children’s products, children’s education |
| 9 | Hobby & Lifestyle | Pets, live plants and garden, outdoor recreation, travel goods, stationery and small lifestyle items, out-of-home experiences incl. dining out |
| 10 | Automotive | Vehicle care and accessories, personal transport |
Table B2.
At-home signal index (T3). At-home signal set: LS-01 Nesting Homemaker, LS-04 DIY Builder, LS-05 Aesthetic Curator, LS-06 Tech-Forward Home, LS-13 Premium Entertainer, from the project’s L3 lifestyle-signal dictionary; the outdoor item in LS-13 (barbecue grill) is excluded. The weight multiplies the row weight of at-home evidence items only.
Table B2.
At-home signal index (T3). At-home signal set: LS-01 Nesting Homemaker, LS-04 DIY Builder, LS-05 Aesthetic Curator, LS-06 Tech-Forward Home, LS-13 Premium Entertainer, from the project’s L3 lifestyle-signal dictionary; the outdoor item in LS-13 (barbecue grill) is excluded. The weight multiplies the row weight of at-home evidence items only.
| Category | At-home signals whose basket contains it | Signals / 5 | Weight w_athome |
| Furniture | LS-01 (dining furniture, storage); LS-05 (mirror); LS-13 (sofa set, outdoor furniture) | 3 / 5 | 1.000 |
| Home Improvement | LS-04 (power tools, drill, ladder, paint, plumbing, hardware); LS-01 (decorative items); LS-05 (décor, wall art, indoor artificial plants, candles) | 3 / 5 | 1.000 |
| Electronics & Gadget | LS-06 (smart devices, IoT, smart lighting, gadgets, laptop); LS-13 (large TV, sound system) | 2 / 5 | 0.667 |
| Kitchen | LS-01 (kitchen set); LS-13 (bar tools) | 2 / 5 | 0.667 |
| Bed & Bath | LS-01 (bedding) | 1 / 5 | 0.333 |
| Cleaning & Household | not in any at-home basket since v1.3 (décor items moved to Home Improvement) | 0 | 0 |
| Health & Sports; Kids & Toys; Hobby & Lifestyle; Automotive | not in any at-home basket | 0 | 0 |
Sensitivity of the index to its parameters is reported together with the independently coded result: the five parameters of Section 3.3 are each varied in turn (geographic weights flat at 1.0; proxy weight 0.75; all-consumer cohort weight 1.0; slope 1.5; w_min at 0.5, 0.75 and 1.5) and the number of cells that change, and whether any category’s reading against every cohort changes, are reported for each setting.
References
- Schewe, C.D.; Meredith, G. Segmenting global markets by generational cohorts. J. Consum. Behav. 2004, 4(1), 51–63. [Google Scholar] [CrossRef]
- NielsenIQ. Mid-Year Consumer Outlook: Guide to 2025. NielsenIQ 2024. [Google Scholar]
- NielsenIQ. Indonesia Consumer Outlook: Guide to 2025. In NielsenIQ Indonesia; 2024. [Google Scholar]
- Sheeran, P.; Webb, T.L. The intention–behavior gap. Soc. Personal. Psychol. Compass 2016, 10(9), 503–518. [Google Scholar] [CrossRef]
- Fishbein, M.; Ajzen, I. Predicting and changing behavior: the reasoned action approach; Psychology Press: New York, 2010. [Google Scholar]
- Ricci, F.; Rokach, L.; Shapira, B. (Eds.) Recommender systems handbook, 3rd ed.; Springer: New York, 2022. [Google Scholar]
- Kitchin, R.; McArdle, G. What makes Big Data, Big Data? Exploring the ontological characteristics of 26 datasets. Big Data Soc. 2016, 3(1), 1–10. [Google Scholar] [CrossRef]
- Chen, M.; Mao, S.; Liu, Y. Big data: a survey. Mob. Netw. Appl. 2014, 19(2), 171–209. [Google Scholar] [CrossRef]
- Hevner, A.R.; March, S.T.; Park, J.; Ram, S. Design science in information systems research. MIS Q. 2004, 28(1), 75–105. [Google Scholar] [CrossRef]
- Gregor, S.; Hevner, A.R. Positioning and presenting design science research for maximum impact. MIS Q. 2013, 37(2), 337–355. [Google Scholar] [CrossRef]
- Nickerson, R.C.; Varshney, U.; Muntermann, J. A method for taxonomy development and its application in information systems. Eur. J. Inf. Syst. 2013, 22(3), 336–359. [Google Scholar] [CrossRef]
- Wedel, M.; Kamakura, W.A. Market segmentation: conceptual and methodological foundations, 2nd ed.; Springer: New York, 2000. [Google Scholar]
- Nielsen, L. Personas, user focused design, 2nd ed.; Springer: London, 2019. [Google Scholar]
- Hughes, A.M. Strategic database marketing; Probus Publishing: Chicago, 1994. [Google Scholar]
- Ajzen, I. The theory of planned behavior. Organ. Behav. Hum. Decis. Process. 1991, 50(2), 179–211. [Google Scholar] [CrossRef]
- Dean, J.; Ghemawat, S. MapReduce: simplified data processing on large clusters. Commun. ACM 2008, 51(1), 107–113. [Google Scholar]
- Zaharia, M.; Xin, R.S.; Wendell, P.; Das, T.; Armbrust, M.; Dave, A.; et al. Apache Spark: a unified engine for big data processing. Commun. ACM 2016, 59(11), 56–65. [Google Scholar]
- Cohen, J. Weighted kappa: nominal scale agreement with provision for scaled disagreement or partial credit. Psychol. Bull. 1968, 70(4), 213–220. [Google Scholar] [CrossRef] [PubMed]
- Krippendorff, K. Content analysis: an introduction to its methodology, 4th ed.; Sage: Thousand Oaks, 2018. [Google Scholar]
- Yin, R.K. Case study research and applications: design and methods, 6th ed.; Sage: Thousand Oaks, 2018. [Google Scholar]
- NielsenIQ. Consumer Outlook: Guide to 2027; NielsenIQ, 2026. [Google Scholar]
- NielsenIQ; World Data Lab. Spend Z: a global report. NielsenIQ 2024. [Google Scholar] [CrossRef]
- NielsenIQ; World Data Lab. The X Factor: how Generation X is quietly driving trillions in consumer spending. NielsenIQ 2025. [Google Scholar]
Table 1.
The seven levels of the Product Profile Attribute (PPA) taxonomy.
| Level | Name | Role in the bridge |
| L0 | Product category | Maps each transaction line to one of the ten NielsenIQ retail-group categories via a documented crosswalk. Establishes the common category space, with an explicit out-of-scope residual for lines outside it. |
| L1 | Product subcategory | Resolves the line to a granular product type within the category (e.g., sofa, rice cooker, drill) — the unit at which action is aggregated, smoothing SKU-level sparsity. |
| L2 | Functional need | Abstracts the subcategory to the practical need it satisfies (e.g., comfort, food preparation, security), decoupling behaviour from catalogue specifics. |
| L3 | Lifestyle signal | Reads what the purchase reveals about the buyer’s lifestyle (e.g., health-conscious, DIY enthusiast, minimalist living). |
| L4 | Value dimension | Lifts the lifestyle signal to the underlying value or motivation (e.g., practicality, status, sustainability), comparable to the value framings used in cohort propensity research. |
| L5 | Cohort affinity / propensity score | Carries the cohort propensity score I(c,k) for category c and cohort k on a 1–5 scale, constructed from published market statements by the Propensity Scoring Index of Section 3.3, and computes the account’s action score A(c) on the same scale. Renders the two sides commensurable. |
| L6 | Propensity–action gap | Emits the terminal output G(c,k) = I(c,k) − A(c): the per-category, per-cohort gap. Positive = under-served (latent demand); negative = over-served. |
Blue = the commensuration level; green = the terminal gap output.
Table 2.
Illustrative enrichment of a single transaction line through the seven levels.
| Level | Illustrative value for one line |
| L0 Product category | Home Improvement |
| L1 Product subcategory | Cordless power drill |
| L2 Functional need | Home maintenance and construction |
| L3 Lifestyle signal | DIY enthusiast |
| L4 Value dimension | Self-reliance and competence |
| L5 Cohort affinity / propensity score | Cohort propensity I (Gen X) = 4.5; account action A = 4.61 (same 1–5 scale) |
| L6 Propensity–action gap | G = I − A = −0.11 → aligned for Gen X |
Table 3.
Demonstration output — the account’s ten-category action scores and the gap range across the four cohorts (positive = under-served; crosswalk version 2).
Table 3.
Demonstration output — the account’s ten-category action scores and the gap range across the four cohorts (positive = under-served; crosswalk version 2).
| NielsenIQ category | Action A(c) | Gap range (min…max across cohorts) | Reading (all four cohorts) |
| Home Improvement | 4.61 | −2.61 … −0.11 | Over-served (Gen Z, Millennial, Boomer) / aligned (Gen X) |
| Hobby & Lifestyle | 4.40 | −2.40 … +0.10 | Over-served (Millennial, Gen X, Boomer) / aligned (Gen Z) |
| Cleaning & Household | 2.85 | −0.85 … +0.65 | Mixed |
| Furniture | 2.40 | +1.10 … +2.10 | Under-served |
| Electronics & Gadget | 2.14 | +0.36 … +2.36 | Under-served (narrowly vs Boomer) |
| Kitchen | 1.99 | +1.01 … +2.51 | Under-served |
| Kids & Toys | 1.60 | −0.60 … +2.90 | Mixed |
| Health & Sports | 1.58 | +1.42 … +2.42 | Under-served |
| Bed & Bath | 1.56 | +1.44 … +2.44 | Under-served |
| Automotive | 1.34 | +1.16 … +2.66 | Under-served |
Table 4.
Window sensitivity of the demonstration output — action scores and gap ranges (original propensity scores) for the full 33-month window and for the final twelve months. Crosswalk version 2. The anchor category changes with the window; five categories remain under-served against every cohort in both.
Table 4.
Window sensitivity of the demonstration output — action scores and gap ranges (original propensity scores) for the full 33-month window and for the final twelve months. Crosswalk version 2. The anchor category changes with the window; five categories remain under-served against every cohort in both.
| NielsenIQ category | A(c), 33 months | Gap range, 33 months | A(c), last 12 months | Gap range, last 12 months | Under-served vs all cohorts (33 m / 12 m) |
| Home Improvement | 4.61 | −2.61 … −0.11 | 5.00 | −3.00 … −0.50 | no / no |
| Hobby & Lifestyle | 4.40 | −2.40 … +0.10 | 2.30 | −0.30 … +2.20 | no / no |
| Cleaning & Household | 2.85 | −0.85 … +0.65 | 3.63 | −1.63 … −0.13 | no / no |
| Furniture | 2.40 | +1.10 … +2.10 | 2.46 | +1.04 … +2.04 | yes / yes |
| Electronics & Gadget | 2.14 | +0.36 … +2.36 | 2.71 | −0.21 … +1.79 | yes / no |
| Kitchen | 1.99 | +1.01 … +2.51 | 2.60 | +0.40 … +1.90 | yes / yes |
| Kids & Toys | 1.60 | −0.60 … +2.90 | 1.70 | −0.70 … +2.80 | no / no |
| Health & Sports | 1.58 | +1.42 … +2.42 | 1.52 | +1.48 … +2.48 | yes / yes |
| Bed & Bath | 1.56 | +1.44 … +2.44 | 1.55 | +1.45 … +2.45 | yes / yes |
| Automotive | 1.34 | +1.16 … +2.66 | 1.34 | +1.16 … +2.66 | yes / yes |
Table 5.
Per-level computational cost of the PPA pipeline and its data-parallel execution pattern.
| Level | Operation | Cost | Data-parallel pattern |
| L0 Category | Hash-join to versioned crosswalk dictionary | O(1) / line | Broadcast lookup (map) |
| L1–L4 Enrichment | Attribute lookups from bounded reference tables | O(1) / line | Broadcast lookup (map) |
| L5 Action + propensity | Group-by aggregation + max-normalisation | O(nₐ) / account | Reduce by account |
| L6 Readout | Subtract 10×4 propensity and action matrices | O(1) / account | Map |
| Overall | Raw lines → per-account gap matrix | O(N) total | Map–reduce, partition by account |
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.