Solution case study

IP Collaboration Candidate Screening

A decision-support prototype that narrows a large pool of character IPs to a per-audience shortlist — separating "generally popular" from "a fit for this segment".

Yichuan Zhang,Shiman Yuan,Heng Zhong9th Data Science Contest · 2026

Role: Co-creator — Team とあるパンダ
Period: 2026
Status: Completed
Stack: Python · Entropy Balancing · Sparse Group Regularisation · Nested Cross-Validation · Survey Data

Statistical screening reduces the candidate pool; brand fit, creative feasibility, licensing and commercial terms remain human decisions in stage two.

One-line summary

A decision-support prototype that separates broad popularity from audience-specific fit, narrowing a large IP pool to a shortlist for human due diligence.

At a glance

01Problem: Overall popularity cannot show whether an IP is unusually relevant to a specific target audience.
02Context: A first-stage screening decision before teams spend time on brand, creative, licensing and commercial due diligence.
03My contribution: Co-created the analytical prototype and final analysis report with Team とあるパンダ.
04Outcome: A validated analysis prototype that keeps market potential, audience lift and structural affinity separate, and withholds rankings when support is weak.

99: surveyed IP candidates
240: out-of-fold target audiences
23,760: audience–IP rows

Three-minute walkthrough

Why a popularity ranking is not enough

A brand choosing a collaboration partner is not asking which character IP is popular in general. It is asking which candidates deserve the cost of deeper investigation for a particular audience. A global ranking hides that distinction and can send the same obvious names to every brief.

Overview of target-aware IP screening from candidate pool to shortlist and deeper human review — Figure 1. The screening stage combines a target-audience brief, historical purchase relationships and IP attributes to narrow the pool before brand, creative, licensing and commercial review.

Two-stage IP collaboration screening process — Figure 2. Stage one uses comparable behavioural evidence to form a shortlist. Stage two evaluates brand fit, creative feasibility, licensing availability, contract terms and business viability outside the statistical score.

Three decision views instead of one opaque score

The output deliberately separates target-audience historical purchase level, audience-specific lift against an age- and gender-matched reference group, and structural affinity between consumer characteristics and IP attributes. These views answer different questions, so combining them behind an arbitrary weight would hide the trade-off from the decision maker.

Illustrative comparison between overall popularity and target-specific fit — Figure 3. Illustrative example; values are not project results. The diagram explains why broad popularity, audience fit and purchase behaviour answer different screening questions.

From an audience brief to a reviewable shortlist

Entropy balancing reconstructs a weighted target audience from the survey pool using age, gender and ten behaviour-and-value concept means. Separate awareness and purchase-among-aware models produce three decision views. A pre-defined policy then returns up to ten candidates only when the input has sufficient support; the result remains a screening list rather than an automatic partnership decision.

IP screening system from inputs and audience reconstruction through modelling, decision outputs and human due diligence — Figure 4. The validated configuration keeps historical purchase level, **audience-specific lift** and affinity separate, applies transparent screening rules and hands the shortlist to human reviewers.

Model overview and mathematical formulation

An audience-weighted, two-stage logistic model combines consumer characteristics, IP characteristics and person–IP interactions to estimate historical purchase propensity for a target audience. It first estimates whether a consumer knows an IP, then whether that consumer has purchased it given awareness. Entropy-balanced weights aggregate these person-level predictions into the target audience's purchase score.

Overview · From consumer–IP predictions to an audience score

\begin{aligned} a_{ij} &=\Pr\!\left(\mathrm{Aware}_{ij}=1\mid\mathbf x_i,\mathbf z_j\right),\\[5pt] m_{ij} &=\Pr\!\left(\mathrm{Purchase}_{ij}=1\mid\mathrm{Aware}_{ij}=1,\mathbf x_i,\mathbf z_j\right),\\[5pt] P_{Gj} &=\sum_{i\in\mathcal D_s}w_i(G)\,a_{ij}\,m_{ij},\\[3pt] w_i(G)&\geq0,\qquad\sum_{i\in\mathcal D_s}w_i(G)=1. \end{aligned}

i indexes consumers, j indexes IP candidates, and G denotes the target audience. xᵢ and zⱼ are the consumer and IP feature vectors. aᵢⱼ is awareness probability; mᵢⱼ is historical self-purchase probability conditional on awareness. The normalized audience weights wᵢ(G) are built within the requested age × sex donor segment 𝒟ₛ. P is the audience-weighted historical purchase score.

Two-stage prediction. Both logistic models include consumer and IP main effects and demographic × IP-history interactions. The purchase-among-aware model additionally includes sparse preference/behaviour/value × structural/commercial interactions. This allows the model to express how the same IP can relate differently to different consumers.

Target-audience reconstruction. Age band and sex define the donor segment; entropy balancing assigns weights so that its ten disclosed behaviour-and-value concept means match the target audience. Consumers contribute to the audience score according to these weights.

Decision outputs. The aggregated purchase score is presented alongside audience-specific lift against a matched demographic baseline and structural affinity. These remain three separate decision views, with reliability rules controlling which outputs can be shown.

The four formulations below detail the audience-weight construction, the two logistic predictors, the full-funnel aggregation, and the definitions of audience lift and structural affinity. The scores describe historical purchase relationships in the survey; future sales and causal collaboration outcomes require separate evidence.

1 · Reconstructing the target audience

\begin{aligned} w_i(\lambda^\star) &=\frac{\exp\!\left(\tilde{\mathbf c}_i^{\top}\lambda^\star\right)} {\sum_{k\in\mathcal D_s}\exp\!\left(\tilde{\mathbf c}_k^{\top}\lambda^\star\right)},\\[4pt] \lambda^\star &=\underset{\lambda}{\arg\min}\;\left[ \log\!\sum_{k\in\mathcal D_s}\exp\!\left(\tilde{\mathbf c}_k^{\top}\lambda\right) \right.\\[-2pt] &\hspace{5.8em}\left. -\log|\mathcal D_s|-\tilde{\boldsymbol\mu}_G^{\top}\lambda +\frac{10^{-8}}{2}\lVert\lambda\rVert_2^2 \right]. \end{aligned}

Within the requested age × sex donor segment 𝒟ₛ, c̃ᵢ is respondent i's standardized ten-concept vector and μ̃G is the target vector. The dual ridge is fixed at 10⁻⁸; a solution is valid only when the maximum standardized moment residual is at most 0.02.

2 · Person × IP full-funnel prediction

\begin{aligned} \eta^{\mathrm{aware}}_{ij} &=\alpha_A+\mathbf x_i^{\top}\boldsymbol\beta_A +\mathbf z_j^{\top}\boldsymbol\gamma_A\\[-1pt] &\hspace{1.3em}+\mathbf d_i^{\top}\mathbf W_{DH,A}\mathbf h_j, &a_{ij}&=\sigma\!\left(\eta^{\mathrm{aware}}_{ij}\right),\\[5pt] \eta^{\mathrm{mon}}_{ij} &=\alpha_M+\mathbf x_i^{\top}\boldsymbol\beta_M +\mathbf z_j^{\top}\boldsymbol\gamma_M\\[-1pt] &\hspace{1.3em}+\mathbf d_i^{\top}\mathbf W_{DH,M}\mathbf h_j +\mathbf p_i^{\top}\mathbf W_{PS}\mathbf s_j, &m_{ij}&=\sigma\!\left(\eta^{\mathrm{mon}}_{ij}\right),\\[5pt] b_{ij}&=a_{ij}m_{ij}.&& \end{aligned}

Awareness uses frozen Model C: person and IP main effects plus demographic-PC × historical/lifecycle/opportunity interactions. Monetization given awareness uses frozen Model A and adds the sparse preference/behaviour/value × structural/commercial interaction. The expected historical purchase relation is bᵢⱼ = aᵢⱼmᵢⱼ; it is not a future-sales forecast.

3 · Aggregating the full funnel

\begin{aligned} R_{Gj}&=\sum_i w_i a_{ij}, &P_{Gj}&=\sum_i w_i a_{ij}m_{ij},\\[3pt] M_{Gj}&=\frac{\sum_i w_i a_{ij}m_{ij}}{\sum_i w_i a_{ij}} & &\neq\sum_i w_i m_{ij}. \end{aligned}

R is target-group reach, P is the modeled historical purchase level, and M is monetization conditional on awareness. The awareness term must remain in both the numerator and denominator; averaging mᵢⱼ directly would answer a different question.

4 · Audience lift and structural affinity

\begin{aligned} L_{Gj} &=P_{Gj}-\frac{1}{|\mathcal D_s|}\sum_{i\in\mathcal D_s}a_{ij}m_{ij},\\[3pt] A_{Gj} &=\sum_i w_i\left(\mathbf p_i^{\top}\mathbf W_{PS}\mathbf s_j\right). \end{aligned}

Lift L compares the reconstructed audience with a uniform donor baseline from the same age × sex segment—not with all-Japan popularity. Affinity A aggregates only the sparse P × S interaction component. Market Potential P, Lift L and Affinity A remain three separate rankings.

Validation that keeps the test set out of the workflow

The frozen analysis used nested model selection and three complementary hold-out designs: unseen IPs, unseen respondents, and both at once. The result supports a limited claim: audience-specific matching adds information that is not reproduced by historical popularity alone. It does not establish future sales, causal campaign lift or return on investment.

Three hold-out validation designs — Figure 5. IP, respondent and double hold-outs test whether the matching signal survives beyond remembered identities; preprocessing and tuning remain inside the training folds.

What the held-out evidence actually shows

The result figures keep distinct validation questions separate. The first two audit the target-aware historical-purchase ranking under held-out people, next-year outcomes and unseen IPs. The final figure evaluates the complete target-summary-to-ranking pipeline across 240 out-of-fold audience cases.

Scientific multi-panel result comparing national, sex-only and target-aware Top-3 IP selection under person-held-out evaluation — Figure 6. Person-held-out benchmark across eight age × sex audiences. Target-aware selection improved the median observed historical purchase rate by **2.46 percentage points** over the national Top-3 (95% person-bootstrap interval 1.08–4.19), while the 0.33-point change over the sex-only Top-3 did not meet the pre-set +1-point value threshold. Scope: 99 surveyed IPs; this is historical purchase affinity, not collaboration lift or revenue.

Scientific multi-panel result comparing rank recovery for held-out people, next-year outcomes and unseen IPs — Figure 7. The validation axes are not interchangeable: median rank correlation was 0.879 for known IPs with new people and 0.625 for the same IPs in the following year, while unseen-IP cold start reached a five-fold mean of 0.287 ± 0.082 standard error. The last result defines a capability boundary rather than a headline accuracy score.

Scientific multi-panel result showing candidate recovery and reliability for out-of-fold audience reconstruction — Figure 8. Final R5/R6 same-survey out-of-fold audit across 240 target audiences and 23,760 audience–IP rows. **Audience-Specific Lift** recovered a median **8 of the gold Lift Top-10 candidates**; Absolute Market Potential recovered 0. Reliability was assigned before output across 18 Reliable, 163 Limited and 59 Unsupported cases. This is not external-population or temporal validation.

A shortlist for investigation, not an automatic partner decision

The system is a 2024-survey analysis prototype. It does not validate new IPs outside the survey, future time periods, natural-language conversion of a business brief, licensing availability or commercial outcomes. Its safest role is to reduce an initial pool and document why a candidate entered the next review stage.