Auxiliary products such as flight protection and flexible-change services are recommended alongside primary travel purchases. This setting creates a granularity ambiguity: SKUs within one product type share semantics but differ in price and coverage, while rejecting one offer does not imply disinterest in the entire type. We present PEETS (Preference-aware Embeddings with Explicit Tree Structure), a training-time framework for auxiliary product bundling. PEETS regularizes an existing click--conversion predictor through two conditional comparisons. Hierarchical Contrastive Loss uses the pair's common ancestor as a reference; Negative Contrastive Loss conditions structural sampling and comparison shifts on logged rejection events. Both act on shared SKU representations during training. This separates catalog relatedness from individual preference supervision and leaves the predictive serving path intact. PEETS achieves 78.20% AUC, exceeding TWIN V2 by 0.90 percentage points and the matched base model by 1.15 points. In a 14-day online evaluation, gross auxiliary GMV per exposed user-day improves by 8.20% over Policies, 3.06% over Base, and 0.84% over TWIN V2, with all three user-bootstrap 95% intervals above zero. Bundling rate is 0.51 points below Policies. The cancellation interval's upper limit exceeds the prespecified margin. PEETS adds no auxiliary-loss computation to serving and records 42.1ms p95 latency, compared with 42.0ms for Base.
Fig. 2: PEETS framework. (a) Predictor and training-only auxiliary branch; dashed links denote sharing and training dependencies. (b) Simplified hierarchy (five levels in experiments). (c) Ancestor-conditioned comparisons. (d) Rejected and reference buckets; structural pairs are not preference labels. (e) Rejection-conditioned comparisons; “bad bucket” is event-specific. In (c,e), ordinary comparisons are on the left. Here b denotes a bucket and L1 the root. Geometries are schematic, not measured projections.
TABLE II: Offline comparison. N@10/N@15 denote NDCG@10/NDCG@15. Learned-model rows report five-seed means; the bottom rows show variability for three key comparators. AUC is in percent and its SD in percentage points. † Price-rule rows use distinct exposure sets, unlike common-test-set learned models.
Fig. 3: Selected results from Tables II and VI . GMV per exposed user-day change is relative to Policies; BR change is in percentage points (pp). The panels show point estimates under their respective evaluation settings; overall GMV intervals appear in Section V-D .
Variant
N@10
N@15
AUC (%)
Base
0.8774
0.9484
77.05
Base + bucket features
0.8821
0.9538
77.32
Base + ordinary CL
0.8845
0.9556
77.46
Base + NCL
0.8892
0.9601
77.72
Base + HCL
0.8918
0.9625
77.89
PEETS
0.8956
0.9667
78.20
TABLE III: Matched-backbone five-seed means. Top: feature and objective controls. Bottom: NCL component controls, retaining HCL and the original NCL sampling while changing only the shift.
α
Tickets
BR (%)
Revenue
0
8697
34.23
141725
1
8690
34.11
143457
2
8482
32.17
153671
3
8560
32.55
160420
4
8625
32.84
165800
5
8698
33.12
170150
TABLE IV: Price-policy sensitivity. Revenue is gross auxiliary GMV in CNY. The sweep varies the price exponent α .
Fig. 4: Bucket-count sensitivity. PEETS includes validation and test AUC at seven settings; HCL/NCL are shown at five test settings. Base is a fixed reference. Lines connect the evaluated settings.
Subset
Requests
Δ (%)
95% CI
Ambiguous
350000
+3.6
[3.12,4.08]
Low-frequency
420000
+4.2
[3.65,4.75]
High-frequency
980000
+1.1
[0.82,1.38]
TABLE V: Selected online subsets: relative change in GMV per request over TWIN V2, with user-bootstrap 95% intervals. Requests are pooled across the two arms.
Arm
Exposed
Tickets
BR (%)
Revenue
Policies
20382
8561
33.83
159764
WDL
20435
8602
33.27
164378
ESMM
20419
8627
33.25
164635
HAKG
20371
8611
33.22
169922
TWIN V2
20442
8671
33.32
171927
Base
20395
8640
33.28
167850
TABLE VI: Online results: per-arm daily means over 14 days. Revenue is gross auxiliary GMV in CNY before refunds.
Metric
Ctrl.
PEETS
Δ
95% CI ( Δ )
Margin
Complaints
0.12
0.13
+0.01
[-0.009,0.028]
0.03
Cancellations
1.80
1.83
+0.03
[-0.076,0.137]
0.10
Abandonment
18.60
18.55
-0.05
[-0.251,0.152]
0.20
TABLE VII: User-experience guardrails. Rates are in percent; treatment-minus-control differences, intervals, and margins are in percentage points.
Method
Train time
Mean
p50
p95
Policies
–
2
1.8
3.2
GBDT
34m 19s
14
13.2
21.6
WDL
58m 28s
27
26.1
42.8
Base
67m 50s
26
25.4
42.0
PLE
73m 28s
27
26.5
43.7
TWIN V2
77m 33s
26
25.5
42.3
TABLE VIII: Training and serving efficiency on 11M records. Neural models train for 3 epochs; latencies are in ms.