betaFeedback

WorksComputer Science

Numerical Embeddings for CTR Prediction: A Taxonomy, Unified Benchmark, and Open-Source FuxiCTR Extension

Abstract

Numerical embeddings map scalar features into vectors before a click-through rate (CTR) backbone combines the fields. To our knowledge, NumEmbBench is the first open benchmark dedicated to comparing numerical embeddings for CTR prediction. We classify 18 configurations into six mechanism-based families and evaluate them through one FuxiCTR interface. After equal-budget validation tuning, we compare frozen configurations on three datasets, nine backbones, and five seeds under a shared train-only z-score protocol. DEER has the best mean rank, but six methods lead at least one dataset–backbone setting. A matched native-scale test changes the winner in 11 of 27 settings, while paired latency measurements show that accuracy and inference cost need not move together. Code and per-seed results: https://github.com/YH-learning/NumEmbBench

CCS Concepts

• Information systems → Recommender systems; • Computing methodologies → Neural networks.

Introduction

CTR models combine categorical fields, such as user or item identifiers, with numerical fields, such as prices and counts. Categorical identifiers use learned lookup vectors. A numerical value instead has magnitude and order, so nearby values may carry related information. A numerical embedding maps scalar features xjx_j to vector ej=fj(xj)∈Rd\mathbf{e}_j = f_j(x_j) \in\mathbb{R}^{d}. Thus the backbone sees a description of the value, such as its scale, interval, or smooth basis responses, rather than the raw number alone. Some methods also condition this vector on other fields [1, 3, 4, 5, 10].

General tabular studies have compared numerical encoders across backbones [4], and CTR method papers include numerical-encoder baselines [5, 10]. BARS instead benchmarks CTR backbones [16]. These studies do not jointly provide an open, encoder-centered CTR benchmark. Their preprocessing, models, and training pipelines also differ, so an apparent encoder gain may reflect the surrounding setup. A dedicated comparison must isolate the numerical representation and measure accuracy and computational cost.

We ask which encoders remain competitive across backbones, whether their rankings transfer across input scales, and what training and inference cost accompanies their accuracy. A shared z-score protocol defines the primary leaderboard; a native-scale track tests the same frozen configurations without retuning. Our contributions are threefold:

  • We introduce, to our knowledge, the first open CTR benchmark centered on numerical embeddings and organize 18 configurations into six mechanism-based families.

  • We establish an equal-budget, five-seed comparison across three datasets and nine CTR backbones, including input-scale and efficiency analyses.

  • We release a reusable FuxiCTR plugin interface, data-processing scripts, frozen configurations, and per-seed results for evaluating future encoders under the same protocol.

Taxonomy and Benchmark Design

Numerical Embedding Taxonomy

Let xj∈Rx_j \in\mathbb{R} be numerical field jj and ej=fj(xj)∈R16\mathbf{e}_j = f_j(x_j) \in\mathbb{R}^{16} its backbone input. Table 1 classifies methods by the mechanism that constructs ej\mathbf{e}_j, not by the downstream interaction network. The plugin registry exposes the same field-wise interface to DNN, DeepFM, WideDeep, NFM, PNN, DCN, DCNv2, AutoInt, and FiBiNET [2, 6, 8, 9, 11, 13, 14, 15]. Categorical embeddings and each backbone’s native interaction structure remain intact.

Coordinate mechanisms. Pointwise methods use the scalar directly or apply a small learned projection. Quantile and logarithmic buckets turn it into a discrete region; AutoDis instead assigns differentiable weights to learned prototypes [5]. Periodic and spline bases evaluate multiple functions of the same input, while PLE and PLR construct continuous, local coordinates over intervals [4]. DEER learns local numerical geometry and NaryDis uses a symbolic multi-radix representation [1, 3]. DAE adapts to feature distributions, and DAES additionally uses selected contextual fields [10, 12]. Grouping these methods by the source of their coordinates makes similarities and differences testable without treating architectural names as families.

The families differ along three representation axes. First, a hard bucket maps nearby values in the same bin to one code and can break continuity at a boundary, whereas soft assignments and basis expansions can preserve graded changes. Second, a preset basis evaluates a fixed set of functions across the input range, while piecewise methods use interval-local support and adaptive-coordinate methods construct their own geometry or digits. Third, most encoders depend only on the target scalar after fitting; distribution-adaptive methods use training-derived information, and selected DAES variants also condition on other fields. These distinctions describe the numerical information supplied to the backbone, not the architecture of the backbone itself.

FamilyEvaluated configurationsDefining mechanismStatisticsContext
Pointwise projectionField, Linear, Linear-ReLUScale or transform one scalarNoNo
DiscretizationQuantile, LogBucket, AutoDis [5]Assign intervals or learned prototypesOptionalNo
Fixed basisPeriodic, B-splineEvaluate preset basis functionsNoNo
Piecewise basisPLE-A/B/ReLU, PLR [4]Construct local interval coordinatesYesNo
Adaptive coordinatesNaryDis/Lite [1], DEER [3]Learn local geometry or symbolic digitsOptionalNo
Distribution/context adaptiveDAE [12], DAES-Gate/Tran [10]Adapt to distributions or selected contextYesMethod dependent

Table 1. Six families classified by their numerical coordinate construction. Statistics means a mapping uses training-derived numerical statistics; context means it can depend on fields other than the scalar being embedded.

Evaluation Protocol and Reproducibility

Common interface. Every plugin receives field-wise numerical tensors and returns one vector of the same width for each field. A registry selects the implementation, and method-specific adapters supply training-derived statistics or selected context only where required. Thus the numerical branch changes without rewriting each CTR backbone. The release distinguishes direct implementations, wrappers around official dependencies, and paper-based reimplementations; tests cover output shape, field ordering, finite gradients, and serialization. This is an infrastructure contribution, not a claim that the underlying CTR library is new [16].

Data and shared settings. The three datasets are AutoML-A, AutoML-E, and Criteo, with 2.80M, 5.14M, and 36.67M training examples, respectively [7]. All numerical statistics are fitted on training data only. The primary protocol applies the same train-only zz-score transform to every encoder. For each dataset–backbone setting (a cell), we use fixed splits, batch size 4096, Adam at 10−310^{-3}, validation-AUC checkpointing, patience 2, and seeds 2026–2030.

The datasets have 23, 6, and 13 numerical fields, respectively, alongside 51, 25, and 26 categorical fields. Processing scripts record split manifests and train-derived vocabularies so that an encoder cannot infer validation or test statistics. All methods output 16-dimensional numerical vectors. Each backbone’s linear and interaction paths are configured consistently across encoders; the exact numerical-field wiring is recorded with the implementation.

Tuning and evaluation. Each method receives the same staged tuning budget: a bounded screening round on DNN and DCNv2, followed by full-schedule validation of the two best candidates on those development backbones. One configuration per method is frozen by validation rank, with logloss breaking ties. The final nine-backbone evaluation contains 18×3×9×5=2,43018 \times3 \times9 \times5 = 2{,}430 runs; test metrics never select configurations. A second 2,430-run matrix changes only the numerical input scale and its train-derived LogBucket vocabulary, reusing those frozen configurations and seeds. Complete per-seed matrices and preprocessing scripts are in the public code repository.

This design deliberately separates three questions: the primary leaderboard measures representation quality after a common stabilizing transform; the native-scale track measures whether the chosen configurations transfer to observed input ranges; and the efficiency study measures implementation cost under a common measurement protocol. In the tuning screen, candidates are predeclared for each encoder and receive the same evaluation budget. The final global configuration is not customized to each backbone or test split, avoiding a large hidden per-cell search advantage.

Results and Implications

RQ1: Which Encoders Lead Across Backbones?

We first compare the zz-score columns of Table 2, where every encoder receives the same train-only zz-score preprocessing. Across the 27 dataset–backbone cells, DEER has the best mean AUC rank (3.41), 14 wins, and 19 top-three placements; PLE-B follows at 5.30. A Friedman test rejects equal method performance (χ2=157.04\chi^2 = 157.04, p=1.02×10−24p = 1.02 \times10^{-24}), and paired comparisons against DEER remain significant after Holm correction. This supports DEER as a strong default under the primary protocol, not as a universal winner.

The full table shows where the summary breaks down. DEER leads eight of nine Criteo backbones but only one of nine AutoML-E backbones; six methods lead at least one zz-score cell overall. Even on AutoML-A, the DEER–Periodic ordering reverses: DeepFM yields 0.7103 versus 0.6982, while FiBiNET yields 0.7063 versus 0.7110. These are five-seed means, not claims that either pairwise difference is statistically significant. A single-backbone leaderboard would miss such changes.

The taxonomy explains representation mechanisms, not a family-level performance ordering. Within piecewise basis, PLE-B ranks 5.30 on average but PLE-ReLU ranks 12.48; within adaptive coordinates, DEER ranks 3.41 but NaryDis-Lite ranks 13.44. Even Field leads three zz-score cells. Thus the particular encoder and backbone matter more than a family label. We retain every dataset–backbone result instead of replacing the matrix with one aggregate score. The Raw columns address scale transfer; per-seed AUC and logloss are available in the artifact.

FamilyMethodDNN z-scoreDNN RawDeepFM z-scoreDeepFM RawWideDeep z-scoreWideDeep RawNFM z-scoreNFM RawPNN z-scorePNN RawDCN z-scoreDCN RawDCNv2 z-scoreDCNv2 RawAutoInt z-scoreAutoInt RawFiBiNET z-scoreFiBiNET Raw
Pointwise projectionField0.80850.80600.80560.79350.80790.80290.80090.79750.80920.80780.80830.80290.80940.80390.80890.80560.80900.8048
Linear0.80860.80520.80410.79060.80770.78670.80100.78120.80910.80130.80820.79990.80940.80270.80880.80290.80950.7937
Linear-ReLU0.80930.80800.80530.80230.80850.79680.80220.78170.80950.80980.80920.80780.81080.81080.80920.80500.81020.7960
DiscretizationQuantile0.81010.81010.80760.80760.80930.80930.80520.80520.81120.81120.81010.81010.81160.81160.81050.81050.81080.8108
LogBucket0.79360.80970.79130.80690.79300.80900.79000.80440.79450.81080.79370.80970.79470.81130.79390.80990.79370.8103
AutoDis0.81020.80550.80770.80270.80950.80450.80430.80070.81100.80620.81030.80520.81150.80680.81050.80550.81040.8059
Fixed basisPeriodic0.80900.80570.80230.80230.80800.80460.80280.80000.80970.80700.80880.80550.81070.80830.80860.80570.81000.8069
B-spline0.80920.80920.80650.80640.80850.80860.80130.80130.80950.80950.80930.80930.81040.81040.80940.80930.80970.8097
Piecewise basisPLE-A0.81040.81040.80650.80650.80960.80960.80490.80490.81150.81140.81030.81030.81220.81220.81050.81050.81160.8116
PLE-B0.81060.80650.80650.79730.81000.79390.80520.77720.81180.80240.81070.80500.81250.80550.81080.80400.81170.7884
PLE-ReLU0.80850.80520.80400.79240.80790.77680.80080.77810.80910.80160.80830.80170.80960.80340.80880.80320.80950.7837
PLR0.80950.80670.80640.80330.80900.80620.80370.80100.81030.80750.80960.80680.81130.80850.80990.80670.81030.8078
Adaptive coordinatesNaryDis-Lite0.80930.81060.80660.80810.80850.80980.80170.80470.80990.81160.80920.81050.81100.81210.80960.81080.80960.8111
NaryDis0.81040.81080.80800.80830.80970.81010.80530.80590.81120.81180.81040.81070.81190.81230.81060.81110.81100.8114
DEER0.81090.81100.80850.80860.81030.81030.80590.80590.81220.81220.81100.81090.81260.81260.81140.81140.81160.8115
Distribution/context adaptiveDAE0.81030.79840.80780.79480.80950.79750.80500.79420.81080.79730.81030.79930.81160.79690.81040.80060.81040.7982
DAES-Gate0.81060.81060.80760.80750.80980.80980.80570.80580.81170.81170.81050.81060.81220.81230.81070.81070.81140.8114
DAES-Tran0.81050.81040.80700.80720.80970.80970.80580.80580.81110.81100.81020.81020.81210.81200.81070.81010.81130.8113

Table 2. Five-seed mean test AUC under shared train-only z-score preprocessing and original numerical input scale (Raw). The Raw track reuses the configurations selected under z-score preprocessing without retuning. Bold marks the best AUC within each dataset–backbone–scale column. Full per-seed AUC and logloss are in the public code repository.

FamilyMethodDNN z-scoreDNN RawDeepFM z-scoreDeepFM RawWideDeep z-scoreWideDeep RawNFM z-scoreNFM RawPNN z-scorePNN RawDCN z-scoreDCN RawDCNv2 z-scoreDCNv2 RawAutoInt z-scoreAutoInt RawFiBiNET z-scoreFiBiNET Raw
Pointwise projectionField0.70540.66480.70740.57280.70670.65850.68570.66000.70110.65530.70630.65960.70780.62330.71110.64970.69690.6284
Linear0.70520.61570.69660.49490.70450.56500.68810.53420.70510.57060.70510.50280.70560.52640.70640.65080.70970.5677
Linear-ReLU0.70440.64180.62050.50000.70310.58450.68120.56870.70400.67320.70370.54100.70470.68130.70810.64400.70800.5598
DiscretizationQuantile0.70580.70580.70780.70820.70670.70720.69100.69120.70540.70520.70450.70470.70590.70610.70940.70990.70460.7043
LogBucket0.68040.68220.67980.68470.67970.68230.66450.66910.67990.68200.67910.67920.68050.68030.68620.68820.67980.6812
AutoDis0.70610.68950.70650.68570.70670.68470.68870.67050.70650.68920.70390.68890.70780.68700.71320.70750.70470.6891
Fixed basisPeriodic0.70500.68470.69820.66550.70240.68500.69210.68540.70380.68410.70250.68140.70840.68670.70860.68340.71100.6814
B-spline0.70790.70750.70960.70970.70870.70880.69260.69270.70600.70600.70680.70660.70900.70880.71330.71370.71000.7090
Piecewise basisPLE-A0.70780.70750.67610.67620.70110.70140.68980.69000.71000.71010.70650.70620.70510.70550.70720.70750.70360.7037
PLE-B0.70790.65010.70160.49510.70510.56540.69460.55200.70710.60360.70550.49750.71040.53060.70800.62340.70870.5248
PLE-ReLU0.70490.62560.69730.49460.70230.57020.68760.56840.70410.60410.70120.51150.70720.49600.70350.62690.70910.5225
PLR0.70630.68070.70700.68240.70670.68050.68610.67340.70290.67990.70450.68060.70930.68280.71160.68660.70610.6812
Adaptive coordinatesNaryDis-Lite0.69200.70270.69000.70070.69470.70010.67640.68060.69140.70320.68930.70100.68910.70010.70480.71050.68670.6957
NaryDis0.69370.70540.69960.70460.70010.70430.68860.69200.70160.70450.70170.70900.69010.70060.71160.71220.69570.7052
DEER0.70910.70880.71030.70090.70990.70920.69240.69230.70900.70200.70890.70860.70820.70770.71580.71580.70630.7066
Distribution/context adaptiveDAE0.70430.70050.70690.70320.70470.70230.69040.68870.70510.70020.70450.69980.70500.70040.71300.70870.70360.6985
DAES-Gate0.70540.70530.70340.70350.70430.70400.68940.68950.70510.70500.70530.70490.70630.70640.71120.71110.70970.7091
DAES-Tran0.70460.70490.70060.70020.70230.70250.68990.69010.70180.70000.70220.70220.70310.70300.70910.70900.70670.7067

Table 2 (PDF p. 3). AutoML-A.

FamilyMethodDNN z-scoreDNN RawDeepFM z-scoreDeepFM RawWideDeep z-scoreWideDeep RawNFM z-scoreNFM RawPNN z-scorePNN RawDCN z-scoreDCN RawDCNv2 z-scoreDCNv2 RawAutoInt z-scoreAutoInt RawFiBiNET z-scoreFiBiNET Raw
Pointwise projectionField0.83580.63220.83920.50000.83870.50000.83950.50000.83840.50000.83380.50000.83750.50000.84400.62570.83410.5000
Linear0.83780.50000.83780.49860.83720.49970.83840.49070.83830.50000.83660.50160.83730.49890.83880.50000.83430.5035
Linear-ReLU0.83810.50000.83690.50000.83730.50010.83670.50390.83910.49690.83620.50380.83730.50680.83980.50000.83440.4966
DiscretizationQuantile0.83870.83870.83830.83830.83730.83750.83960.83960.83860.83860.83500.83500.83570.84240.84240.84240.83370.8337
LogBucket0.83660.83820.83720.83650.83740.83620.83850.83890.83850.83710.83520.83570.83630.83470.84140.84100.83250.8338
AutoDis0.83640.83510.83750.83700.83780.83680.83970.83860.83830.83710.83670.83570.83700.83540.84310.84100.83310.8325
Fixed basisPeriodic0.83790.83700.83750.83760.83730.83840.83750.83760.83790.83900.83800.83800.83790.83840.83950.83950.83420.8388
B-spline0.83730.83730.83760.83740.83720.83740.83970.83980.83910.83870.83710.83700.83730.83730.84170.84200.83820.8332
Piecewise basisPLE-A0.83750.83770.83850.83850.83790.83760.83750.83750.83950.83960.83780.83820.83930.83940.83950.83930.83640.8364
PLE-B0.83660.50000.83770.49960.83780.50010.83780.49730.83960.50460.83780.49700.83760.49760.84010.50000.83530.4971
PLE-ReLU0.83670.50000.83830.49860.83690.50010.83780.49730.83850.50460.83740.49700.83760.49760.84010.50000.83530.4971
PLR0.83810.83870.83880.83790.83700.83740.83820.84030.83890.83830.83700.83800.83750.83700.84050.84050.83230.8386
Adaptive coordinatesNaryDis-Lite0.83690.83770.83770.83790.83760.83760.83800.83820.83820.83840.83570.83510.83630.83640.84180.84190.83290.8336
NaryDis0.83710.83670.83660.83710.83650.83660.83930.83940.83800.83860.83580.83520.83680.83690.84170.84310.83120.8321
DEER0.83740.83710.83900.83990.83850.83820.84000.83990.83790.83820.83670.83640.83750.83770.84240.84250.83520.8351
Distribution/context adaptiveDAE0.83840.83790.83830.83810.83800.83710.83930.83920.83890.83840.83680.83650.83630.83610.84250.84200.83440.8344
DAES-Gate0.83830.83840.83590.83570.83570.83570.83880.83890.83890.83870.83490.83510.83780.83780.84210.84240.83160.8317
DAES-Tran0.83820.83840.83730.83730.83710.83760.83880.83890.83780.83830.83710.83710.83610.83630.84140.84140.83340.8335

Table 2 (PDF p. 3, second). AutoML-E.

RQ2: Do Rankings Transfer to Raw Input Scale?

Each Raw column in Table 2 is paired with the adjacent z-score column: the split, seed, backbone, and configuration selected under z-score preprocessing remain fixed while the numerical input scale changes (with a train-derived LogBucket vocabulary). The winning method changes in 11 of 27 cells: one on Criteo, three on AutoML-A, and seven on AutoML-E. Scale transfer is therefore dataset dependent, rather than a uniform shift in AUC.3

The differences also vary sharply by encoder within the same backbone. On AutoML-E/AutoInt, PLE-B changes from 0.8411 to 0.5000, whereas B-spline stays near its z-score result (0.8417 versus 0.8420). Across the matrix, 57 of 486 Raw method cells have test AUC below 0.55 (14 on AutoML-A and 43 on AutoML-E, none on Criteo), compared with none under z-score preprocessing. The concentration on AutoML-E cautions against generalizing a single dataset’s apparent scale robustness. This tests transfer, not independently tuned Raw performance; retuning may recover low scores. Yet the same backbone and training budget do not ensure scale robustness, so preprocessing must accompany each encoder comparison.

RQ3: What Does the Accuracy Cost?

FamilyMethodParam.TrainInfer.
Pointwise projectionField0.2k1.00×1.00×
Pointwise projectionLinear0.4k1.00×1.01×
Pointwise projectionLinear-ReLU0.4k1.01×1.02×
DiscretizationQuantile2.1k1.22×1.23×
DiscretizationLogBucket3.0k1.32×1.60×
DiscretizationAutoDis7.4k1.04×1.12×
Fixed basisPeriodic3.9k1.03×1.07×
Fixed basisB-spline7.2k1.07×1.37×
Piecewise basisPLE-A3.8k1.02×1.09×
Piecewise basisPLE-B4.0k1.03×1.10×
Piecewise basisPLE-ReLU4.0k1.03×1.11×
Piecewise basisPLR7.6k1.03×1.08×
Adaptive coordinatesNaryDis-Lite19.0k1.25×2.26×
Adaptive coordinatesNaryDis32.3k1.44×2.43×
Adaptive coordinatesDEER4.0k0.76×1.20×
Distribution/context adaptiveDAE36.7k1.08×1.13×
Distribution/context adaptiveDAES-Gate4.5k1.28×1.31×
Distribution/context adaptiveDAES-Tran60.9k1.32×1.34×

Table 3. Numerical parameters and paired training and inference latency relative to Field. Ratios use the same dataset, backbone, and measurement round.

Table 3 adds cost to the z-score accuracy comparison in Table 2; we did not measure a separate Raw-scale latency matrix. We time real resident batches on the same idle GPU in eager FP32, using three representative backbones and three randomized rounds per method–dataset–backbone case; disk input is excluded. The 486 successful measurements use the same frozen configurations as the z-score accuracy study. DEER, the best method by mean z-score rank, uses about 4.0k numerical parameters and takes 0.76×0.76 \times Field training time but 1.20×1.20 \times Field inference time. The evaluated NaryDis variants exceed 2×2 \times Field inference latency. These ratios describe released implementations, not theoretical lower bounds.

Parameter count alone is an inadequate cost proxy. DAE has 36.7k numerical parameters yet 1.13×1.13 \times Field inference latency, whereas NaryDis-Lite has 19.0k parameters and 2.26×2.26 \times latency. PLE-B, second by mean z-score rank, takes 1.10×1.10 \times Field inference time compared with DEER’s 1.20×1.20 \times; its accuracy–latency tradeoff therefore merits examination in individual cells. Periodic, PLE-A/B, PLR, AutoDis, and DAE each remain within 1.13×1.13 \times Field at inference, but other discretization and adaptive-coordinate implementations are substantially slower. The family label identifies a representation mechanism, not a fixed computational price. Absolute latency, embedding-stage timing, and peak-memory records in the artifact support deployment-specific decisions; these ratios are not hardware-independent limits.

Overall takeaways.

  1. There is no universal accuracy winner. DEER leads the shared z-score matrix by mean rank, yet six methods win at least one of its 27 dataset–backbone cells; even variants in one family differ substantially.

  2. Input scale changes model selection. With configurations held fixed, the winner changes in 11 of 27 paired cells and 57 Raw method cells fall below 0.55 AUC, versus none under z-score preprocessing. These are transfer failures, not proof that Raw-specific retuning cannot help.

  3. Accuracy does not determine computational cost. DEER trains at 0.76×0.76 \times Field time but infers at 1.20×1.20 \times; the evaluated NaryDis variants exceed 2×2 \times inference time. Both training and serving measurements belong beside AUC.

Together, these findings call for reporting the dataset, backbone, preprocessing, per-cell AUC, and paired cost rather than a single aggregate leaderboard.

Open-Source Benchmark Artifact

The release contains dataset converters, train-only preprocessing, encoder implementations, frozen configurations, a common runner, and deterministic table builders. Per-seed tables identify the dataset, backbone, encoder, seed, and validation and test AUC and logloss. Task manifests and configuration snapshots document the run settings, while the builders check completeness before aggregation. Provenance notes distinguish official dependencies, wrappers, and paper-based reimplementations. Thus the summary tables remain auditable without distributing model checkpoints.

The same artifact supports extension: a new encoder declares whether it needs fitted numerical statistics or other-field context and returns the common embedding width in the recorded field order. It can then enter the existing dataset–backbone–seed matrix without modifying the nine CTR backbones. This makes the release a reusable evaluation substrate, not only a snapshot of the present leaderboard.

Conclusion and Limitations

NumEmbBench turns numerical representation into a controlled model-selection problem: the taxonomy identifies distinct mechanisms, the common plugin interface separates them from CTR backbones, and the multi-backbone results expose ranking changes hidden by isolated comparisons. DEER is the strongest default under the shared z-score protocol, but its advantage does not extend uniformly to other members of its family or to every backbone. The benchmark is limited to three datasets, one global tuning budget per method, and the released eager implementations. AutoML-A and AutoML-E also share a source family, so the three datasets are not three independent domains. Native-scale results transfer configurations selected under z-score preprocessing rather than retune each encoder. Latency reflects the released implementations, not a hardware-independent limit.

Ethical Considerations

The benchmark uses existing tabular CTR datasets and aggregate predictive metrics, with no new user study or personal-data collection. Better CTR prediction may reinforce exposure and feedback biases; deployment should separately audit privacy, fairness, and drift.

References

  1. [1]Bo Chen, Huifeng Guo, Weiyen Liu, Yue Ding, Yunzhe Li, Wei Guo, Yichao Wang, Zhicheng He, Ruiming Tang, and Rui Zhang. 2022. Numerical Feature Representation with Hybrid N-ary Encoding. In Proceedings of the 31st ACM International Conference on Recommender Systems. 3882–3887. doi:10.1145/3511808.3557090DOI
  2. [2]Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Anil, Zakaria Haque, Lichan Hong, Vihan Jain, Xiaobing Liu, and Hemal Shah. 2016. Wide & Deep Learning for Recommender Systems. In Proceedings of the 1st Workshop on Deep Learning for Recommender Systems. 7–10. doi:10.1145/2988450.2988454DOI
  3. [3]Yuan Cheng. 2022. Dynamic Explicit Embedding Representation for Numerical Features in Deep CTR Prediction. In Proceedings of the 31st ACM International Conference on Information and Knowledge Management. 3888–3892. doi:10.1145/3511808.3557587DOI
  4. [4]Yury Gorishniy, Ivan Rubachev, and Artem Babenko. 2022. On Embeddings for Numerical Features in Tabular Deep Learning. In Advances in Neural Information Processing Systems, Vol. 35.arxiv.org/abs/2203.05556
  5. [5]Huifeng Guo, Bo Chen, Ruiming Tang, Weinan Zhang, Zhenguo Li, and Xiuqiang He. 2021. An Embedding Learning Framework for Numerical Features in CTR Prediction. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2910–2918. doi:10.1145/3447548.3467077DOI
  6. [6]Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine Based Neural Network for CTR Prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725–1731. doi:10.24963/ijcai.2017/239DOI
  7. [7]Isabelle Guyon, Lionel Hosoya, Marc Boullé, Hugo Jair Escalante, Sergio Escalera, Zhengying Liu, Damir Jereyic, Bisakha Ray, Mehreen Saeed, Michèle Sebag, Alexander Statnikov, Wei-Wei Tu, and Evelyne Viegas. 2019. Analysis of the AutoML Challenge Series 2015–2018. In Automated Machine Learning. Springer.
  8. [8]Xiangnan He and Tat-Seng Chua. 2017. Neural Factorization Machines for Sparse Predictive Analytics. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364. doi:10.1145/3077136.3080777DOI
  9. [9]Tongwen Huang, Ziqi Zhang, and Junlin Zhang. 2019. FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for Click-Through Rate Prediction. In Proceedings of the 13th ACM Conference on Recommender Systems. 169–177. doi:10.1145/3298689.3347043DOI
  10. [10]Jiahai Hu, Han Cai, Kan Ren, Weimin Zhang, Yong Yang, Ying Wen, and Jun Wang. 2016. Product-Based Neural Networks for User Response Prediction. In IEEE 16th International Conference on Data Mining. 1149–1154. doi:10.1109/ICDM.2016.0151
  11. [11]Yanruu Qu, Han Cai, Kan Ren, Weinan Zhang, Yong Yang, Ying Wen, and Jun Wang. 2016. Product-Based Neural Networks for User Response Prediction. In IEEE 16th International Conference on Data Mining. 1149–1154. doi:10.1109/ICDM.2016.0151
  12. [12]Bin Shen, Jingran Xu, Xu Min, Zeyu Ke, Yong He, Liang Zhang, Xin Dong, and Linjian Mo. 2023. DAE: Distribution-Aware Embedding for Numerical Features in Click-Through Rate Prediction. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 4279–4283. doi:10.1145/3583780.3615212
  13. [13]Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management. 1161–1170. doi:10.1145/3357384.3357925DOI
  14. [14]Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & Cross Network for Ad Click Predictions. In Proceedings of the ADKDD’17. doi:10.1145/3124749.3124754DOI
  15. [15]Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed H. Chi. 2021. DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-Scale Learning to Rank Systems. In Proceedings of the Web Conference 2021. 1785–1797. doi:10.1145/3442381.3450078DOI
  16. [16]Jieming Zhu, Jinyang Liu, Shuai Yang, Qi Zhang, and Xiuqiang He. 2021. Open Benchmarking for Click-Through Rate Prediction. In Proceedings of the

Paper details

vlnc:W_e4f8qy64
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents