Represent
Behavior creates the geometry.
Compress pre-decision activity into an embedding, combine it with ordinary customer attributes, and estimate enrollment propensity for every eligible program.
This is where the main predictive gain appears.
Utilities need to choose the right eligible program for each customer. This project tests whether behavioral history and local prediction errors can improve that choice. It also identifies when the evidence is too thin to trust.
Program recommendation · calibration · reliability
When should local evidence change the recommendation?
Behavioral embeddings reveal what customers are likely to choose. Propensity neighborhoods ask a harder second question: is the nearby evidence strong enough to trust?
Decision pathEvidence aligned
01Represent behaviorhistory → embedding $z(x)$
02Score eligible programsattributes + embedding → $\widetilde p(x)$
03Measure local supportnested peers → mismatch + uncertainty
04Decide honestlycalibrate · retain · abstain
$\text{trust local evidence} \iff \text{match gain} > \text{support cost}$
The best recommendation is not always the most locally refined one. It is the one whose specificity is still supported by evidence.
The story arc
From behavioral signal to an honest decision.
Each step solves one problem and exposes the next. The empirical result gives the three steps distinct jobs.
Refine
Specificity spends support.
Add matching dimensions to find more relevant peers. The remaining neighbors become more alike, but every refinement can only reduce their number.
A better match arrives with greater uncertainty.
Decide
The frontier governs trust.
Balance residual mismatch against finite-sample uncertainty. Correct when supported; otherwise keep the direct ranking or flag low confidence.
The theory makes restraint part of the method.
The mathematical core
One bound captures the trade.
The notation says something simple: local correction is useful only while closer peers compensate for having fewer of them.
The refinement frontier
At neighborhood level $m$, the score uncertainty is bounded by a mismatch term plus a support term:
\[ b_m(x)=\varepsilon_m^\rho(x)+B\sqrt{\frac{\log(2K_x/\alpha)}{2n_m}}. \]
Under the theorem's assumptions, the chosen program's propensity regret is at most $2b_m(x)$. The preferred level minimizes the frontier rather than blindly choosing the most detailed neighborhood.
$\varepsilon_m^\rho(x)$ · local mismatchHow different the remaining model errors may be among nearby customers.
$n_m$ · peer supportHow many reference customers remain after level $m$ of refinement.
$K_x$ · eligible choicesHow many programs compete for this customer's recommendation.
Chronological KuaiRec study
Behavioral embeddings win. Neighborhoods calibrate.
The strongest result is also the simplest operational recommendation: encode pre-decision behavior as customer embeddings, use them to estimate enrollment propensities, then rank eligible programs directly by those scores. KuaiRec is an almost fully observed dataset of viewing and feedback logs from the Kuaishou video app.
Propensity RMSE−31.9%Behavioral embedding versus static features
Top-1 regret−6.4%Lower regret across the prespecified comparison
Exact hit rate+2.58 ppImproved in nine of ten chronological splits
The honest ending: neighborhoods contain real calibration information, but neither they nor the supplemental reranking variants establish a confirmed improvement in the winning recommendation. Better probability estimates do not automatically produce a better top choice.
Primary publication
Start with the accessible paper.
The HTML and PDF are equivalent editions of the main manuscript. The browser edition is designed for the easiest first read; the PDF is ready to download and share.
Featured · HTML + PDF
Behavioral Embeddings and Propensity Neighborhoods for Program Recommendation
The complete story, formal statements, proofs, operating workflow, and KuaiRec evidence in a concise, accessible presentation.
Source edition
LaTeX source
Inspect the manuscript source, theorem statements, notation, and references used to build the accessible edition.
Supplemental materials
The experimental trail remains open.
These artifacts document how the conclusion was tested, challenged, and narrowed. They support the paper; they are not competing entry points.
Experiments & synthesis
Follow-up tests of reranking, preprocessing, embedding neighborhoods, uncertainty gates, and decision boundaries.
Notes, code & reproduction
The longer technical paper, earlier geometric note, executable scripts, and data-layout instructions.
Reproducible by design. Source, specifications, split-level metrics, and negative results are retained alongside the manuscript.
View on GitHub