The code and experimental results of《A Comprehensive Analysis and Evaluation of Feature Refinement Modules for CTR Prediction》.

Click-through rate (CTR) prediction is widely used in academia and industry. Most CTR prediction models follow the same design paradigm: Feature Embedding (FE) layer, Feature Interaction (FI) layer, and Prediction layer. RefineCTR inserts existing feature refinement modules between FE and FI layer to boost basic CTR models performance, which also sumarizes and compares those modules' effectiveness comprehensively. The structure of RefinedCTR is shown in the following:

The implement of FR modules is shown in the following figure : (A) and (B) inserting a single FR module between FR and FI layer for stacked and parallel CTR models; (C) assigning two separate FR modules for different FI sub-network to generate discriminate feature distributions for parallel models.

Feature Refinement Modules

We extrct 14 FR modules from existing works. And FAL equires that the input data not anonymous. Therefore, we do not evaluate FAL in the experiments as Criteo dataset is anonymous. It is worth noting that we would fix the hyper-parameters of the basic CTR models when applying these modules to ensure the fairness of the experiments. Relatively speaking, this also leaves space for improvement for each augmented model.

Year	Module	Literature
2019	FEN	IFM
2019	SENET	FiBiNet
2020	FWN	NON
2020	DFEN	DIFM
2020	DRM	FED
2020	FAL	FaFM
2020	VGate	GateNet
2020	BGate	GateNet
2020	SelfAtt	InterHAt
2021	TCE	ContextNet
2021	PFFN	ContextNet
2022	GFRL	MCRF
2022	FRNet-V	FRNet
2022	FRNet-B	FRNet

Basic CTR Models

	Model	Publication	Patterns
1	FM	ICDM'10	A SIngle
2	DeepFM	IJCAI'17	B SIngle
3	DeepFM	IJCAI'17	C Separate
4	CN (DCN)	ADKDD'17	A SIngle
5	DCN	ADKDD'17	B SIngle
6	DCN	ADKDD'17	C Separate
7	AFN (AFN+)	AAAI'20	A SIngle
8	AFN+	AAAI'20	B SIngle
9	AFN+	AAAI'20	C Separate
10	CN2 (DCNV2)	WWW'21	A SIngle
11	DCNV2	WWW'21	B SIngle
12	DCNV2	WWW'21	C Separate
13	CIN (xDeepFM)	KDD'18	A SIngle
14	xDeepFM	KDD'18	B SIngle
15	xDeepFM	KDD'18	C Separate

16	AFM	IJCAI'17	A SIngle
17	NFM	SIGIR'17	A SIngle
18	FwFM	WWW'18	A SIngle
19	FINT	arXiv' 21	A SIngle
20	PNN	ICDM'16	A SIngle
21	FiBiNET	RecSys'19	C Separate
22	DCAP	CIKM'21	A SIngle
23	AutoInt	CIKM'19	A SIngle
24	AutoIntP	CIKM'19	B SIngle
25	AutoIntP	CIKM'19	C Separate

Generally, stacked models (e.g., FM. CN) only use pattern A (Single FR module); parallel models can adopt both pattern B (single FR module) and Pattern C (two separate FR modules). Including basic model, we can generate N*14(N is the basic model number) augmented models. Meanwhile, for parallel CTR model, we can assigning different FR module for different FI sub-networks. Now we assign two same FR modules for different sub-networks.

Experiment Results

We will continue to update and upload the latest experimental results and analysis.

Criteo

AUC of Criteo

Modules	SKIP	FEN	SENET	FWN	DFEN	DRM	VGate	BGate	SelfAtt	TCE	PFFN	GFRL	FRNet-V	FRNet-B
FM	0.8080	0.8100	0.8102	0.8100	0.8117	0.8107	0.8090	0.8091	0.8099	0.8112	0.8129	0.8134	0.8139	0.8140
DeepFM	0.8121	0.8128	0.8125	0.8125	0.8121	0.8118	0.8125	0.8127	0.8112	0.8123	0.8129	0.8137	0.8140	0.8141
DeepFM(2)	0.8121	0.8130	0.8128	0.8129	0.8123	0.8119	0.8128	0.8131	0.8129	0.8128	0.8132	0.8138	0.8142	0.8142
CN	0.8093	0.8102	0.8095	0.8094	0.8121	0.8109	0.8107	0.8110	0.8102	0.8122	0.8130	0.8139	0.8143	0.8144
DCN	0.8125	0.8130	0.8116	0.8127	0.8127	0.8118	0.8124	0.8127	0.8122	0.8126	0.8131	0.8142	0.8143	0.8145
DCN(2)	0.8125	0.8136	0.8122	0.8126	0.8131	0.8120	0.8124	0.8127	0.8129	0.8133	0.8132	0.8144	0.8144	0.8146
AFN	0.8099	0.8140	0.8104	0.8106	0.8116	0.8110	0.8103	0.8101	0.8110	0.8122	0.8130	0.8130	0.8139	0.8141
AFN+	0.8108	0.8141	0.8111	0.8124	0.8119	0.8125	0.8119	0.8118	0.8129	0.8128	0.8132	0.8141	0.8141	0.8141
AFN+(2)	0.8108	0.8142	0.8116	0.8128	0.8127	0.8129	0.8124	0.8126	0.8131	0.8131	0.8134	0.8142	0.8143	0.8145
CN2	0.8121	0.8119	0.8119	0.8140	0.8131	0.8140	0.8128	0.8133	0.8131	0.8138	0.8130	0.8143	0.8141	0.8143
DCNV2	0.8128	0.8129	0.8122	0.8142	0.8134	0.8141	0.8130	0.8138	0.8135	0.8136	0.8130	0.8143	0.8141	0.8143
DCNV2(2)	0.8128	0.8136	0.8125	0.8140	0.8135	0.8142	0.8139	0.8140	0.8137	0.8138	0.8131	0.8144	0.8143	0.8144
IMP	-	0.18%	0.03%	0.13%	0.15%	0.12%	0.09%	0.11%	0.11%	0.18%	0.22%	0.33%	0.35%	0.37%

Logloss of Criteo

Modules	SKIP	FEN	SENET	FWN	DFEN	DRM	VGate	BGate	SelfAtt	TCE	PFFN	GFRL	FRNet-V	FRNet-B
FM	0.4437	0.4418	0.4416	0.4418	0.4401	0.4411	0.4425	0.4424	0.4408	0.4405	0.4391	0.4384	0.4380	0.4378
DeepFM	0.4398	0.4390	0.4413	0.4393	0.4399	0.4400	0.4392	0.4390	0.4404	0.4396	0.4390	0.4382	0.4379	0.4378
DeepFM(2)	0.4398	0.4389	0.4410	0.4389	0.4396	0.4400	0.4390	0.4386	0.4389	0.4390	0.4387	0.4380	0.4378	0.4377
CN	0.4498	0.4415	0.4432	0.4496	0.4398	0.4409	0.4409	0.4406	0.4417	0.4398	0.4392	0.4380	0.4378	0.4375
DCN	0.4394	0.4389	0.4419	0.4391	0.4393	0.4400	0.4394	0.4391	0.4395	0.4396	0.4392	0.4377	0.4376	0.4375
DCN(2)	0.4394	0.4383	0.4405	0.4391	0.4389	0.4399	0.4393	0.4391	0.4389	0.4386	0.4388	0.4376	0.4374	0.4374
AFN	0.4420	0.4378	0.4412	0.4411	0.4402	0.4409	0.4413	0.4416	0.4408	0.4396	0.4391	0.4388	0.4379	0.4377
AFN+	0.4410	0.4378	0.4407	0.4394	0.4399	0.4394	0.4398	0.4399	0.4389	0.4390	0.4390	0.4378	0.4378	0.4378
AFN+(2)	0.4410	0.4378	0.4401	0.4390	0.4392	0.4388	0.4393	0.4391	0.4388	0.4388	0.4388	0.4377	0.4376	0.4374
CN2	0.4389	0.4402	0.4402	0.4380	0.4393	0.4381	0.4394	0.4386	0.4387	0.4383	0.4393	0.4378	0.4379	0.4378
DCNV2	0.4390	0.4391	0.4399	0.4379	0.4389	0.4380	0.4390	0.4382	0.4383	0.4385	0.4391	0.4378	0.4378	0.4378
DCNV2(2)	0.4390	0.4382	0.4394	0.4380	0.4388	0.4379	0.4380	0.4380	0.4383	0.4383	0.4391	0.4376	0.4378	0.4377
IMP	-	0.44%	0.03%	0.22%	0.35%	0.33%	0.29%	0.35%	0.35%	0.43%	0.45%	0.70%	0.74%	0.77%

Assigning Two Seperate Feature Refinement Modules for Different sub-networks based on DeepFM.

Frappe

AUC of Frappe

Modules	SKIP	FEN	SENET	FWN	DFEN	DRM	VGate	BGate	SelfAtt	TCE	PFFN	GFRL	FRNet-V	FRNet-B
FM	0.9786	0.9789	0.9800	0.9808	0.9799	0.9820	0.9801	0.9803	0.9806	0.9800	0.9822	0.9821	0.9828	0.9831
DeepFM	0.9824	0.9828	0.9827	0.9824	0.9824	0.9827	0.9828	0.9825	0.9831	0.9824	0.9830	0.9828	0.9837	0.9840
DeepFM(2)	0.9824	0.9830	0.9829	0.9829	0.9827	0.9825	0.9835	0.9828	0.9836	0.9839	0.9829	0.9843	0.9848	0.9846
CN	0.9797	0.9829	0.9798	0.9810	0.9810	0.9803	0.9803	0.9803	0.9816	0.9819	0.9826	0.9827	0.9825	0.9826
DCN	0.9825	0.9830	0.9822	0.9826	0.9838	0.9834	0.9829	0.9820	0.9829	0.9827	0.9828	0.9838	0.9838	0.9837
DCN(2)	0.9825	0.9834	0.9829	0.9831	0.9843	0.9843	0.9835	0.9829	0.9832	0.9839	0.9838	0.9840	0.9844	0.9847
AFN	0.9812	0.9826	0.9812	0.9816	0.9822	0.9821	0.9821	0.9814	0.9820	0.9826	0.9815	0.9835	0.9838	0.9838
AFN+	0.9827	0.9838	0.9827	0.9831	0.9840	0.9836	0.9830	0.9826	0.9830	0.9836	0.9827	0.9838	0.9843	0.9844
AFN+(2)	0.9827	0.9840	0.9830	0.9840	0.9846	0.9838	0.9839	0.9827	0.9837	0.9838	0.9834	0.9841	0.9844	0.9847
CN2	0.9810	0.9822	0.9813	0.9826	0.9830	0.9825	0.9827	0.9813	0.9827	0.9821	0.9817	0.9825	0.9826	0.9834
DCNV2	0.9830	0.9833	0.9835	0.9831	0.9839	0.9837	0.9833	0.9826	0.9829	0.9833	0.9831	0.9840	0.9839	0.9845
DCNV2(2)	0.9830	0.9838	0.9838	0.9838	0.9844	0.9838	0.9837	0.9828	0.9832	0.9841	0.9835	0.9845	0.9841	0.9849
IMP		0.10%	0.04%	0.08%	0.12%	0.11%	0.08%	0.02%	0.09%	0.11%	0.10%	0.17%	0.20%	0.22%

Modules	SKIP	FEN	SENET	FWN	DFEN	DRM	VGate	BGate	SelfAtt	TCE	PFFN	GFRL	FRNet-V	FRNet-B
CIN	0.9834	0.9839	0.9835	0.9830	0.9837	0.9836	0.9772	0.9787	0.9813	0.9792	0.9837	0.9837	0.9845	0.9842
xDeepFM	0.9841	0.9846	0.9844	0.9836	0.9836	0.9846	0.9844	0.9835	0.9830	0.9839	0.9837	0.9850	0.9851	0.9851
xDeepFM(2)	0.9841	0.9841	0.9844	0.9840	0.9841	0.9847	0.9850	0.9845	0.9832	0.9833	0.9831	0.9852	0.9853	0.9854
NFM	0.9818	0.9830	0.9816	0.9799	0.9830	0.9825	0.9845	0.9830	0.9808	0.9818	0.9828	0.9839	0.9843	0.9855
AFM	0.9697	0.9814	0.9797	0.9804	0.9808	0.9771	0.9718	0.9724	0.9793	0.9776	0.9801	0.9812	0.9816	0.9830
FwFM	0.9834	0.9809	0.9834	0.9813	0.9834	0.9816	0.9830	0.9819	0.9840	0.9834	0.9823	0.9842	0.9831	0.9837
FINT	0.9832	0.9845	0.9817	0.9828	0.9828	0.9842	0.9838	0.9833	0.9825	0.9822	0.9827	0.9848	0.9834	0.9840
PNN	0.9841	0.9843	0.9846	0.9837	0.9836	0.9845	0.9835	0.9836	0.9828	0.9833	0.9831	0.9837	0.9842	0.9843
FiBiNet	0.9827	0.9825	0.9825	0.9827	0.9829	0.9824	0.9828	0.9828	0.9830	0.9823	0.9819	0.9832	0.9835	0.9846
DCAP	0.9841	0.9836	0.9845	0.9834	0.9843	0.9857	0.9846	0.9838	0.9823	0.9841	0.9824	0.9847	0.9853	0.9848

Get started

Test existing model with existing module. Users can choose the appropriate CTR model and feature refinement module conveniently according their needs.

cd evaluation/mains
CUDA_VISIBLE_DEVICES=0 python main_frappe_base --model 0 --module 0

Adding new model or module.

Our framework RefineCTR is modularized, users can adjust or add basic models and modules easily.

codectr/RefineCTR