DALLRec:An effective data augmentation framework with fine-tuning large language model for recommendation

We introduce a novel framework (DALLRec) that enables the efficient utilizing large language models to alleviate data sparsity issues and enhance recommendation effectiveness

Main results

data				movielens
backbone	Variants	R@10	R@20	R@50	N@10	N@20	N@50	P@20
MF-BPR	Base	0.0172	0.0304	0.0616	0.0556	0.0800	0.1334	0.0245
	DALLRec	0.0176	0.0320	0.0656	0.0576	0.0833	0.1404	0.0264
	Imporve	↑2.33%	↑5.26%	↑6.49%	↑3.59%	↑4.13%	↑6.17%	↑7.76%
NMF	Base	0.0185	0.0322	0.0670	0.0637	0.0910	0.1523	0.0282
	DALLRec	0.0189	0.0335	0.0716	0.0661	0.0954	0.1606	0.0306
	Imporve	↑2.16%	↑4.04%	↑2.43%	↑3.77%	↑4.84%	↑5.45%	↑8.51%
NGCF	Base	0.0264	0.0544	0.0908	0.0755	0.1270	0.1906	0.0207
	DALLRec	0.0272	0.0554	0.0943	0.0801	0.1315	0.1982	0.0225
	Imporve	↑3.03%	↑1.84%	↑3.85%	↑6.09%	↑3.54%	↑3.99%	↑8.70%
LightGCN	Base	0.0281	0.0576	0.0950	0.0288	0.0401	0.0544	0.0206
	DALLRec	0.0295	0.0616	0.1007	0.0299	0.0425	0.0588	0.0221
	Imporve	↑4.98%	↑6.94%	↑6.00%	↑3.82%	↑3.49%	↑8.09%	↑7.28%
MMGCN	Base	0.0322	0.0647	0.1280	0.0834	0.1304	0.1991	0.0292
	DALLRec	0.0333	0.0675	0.1383	0.0857	0.1343	0.2110	0.0311
	Imporve	↑3.41%	↑4.33%	↑8.05%	↑2.75%	↑2.99%	↑5.98%	↑6.51%
GRCN	Base	0.0392	0.0711	0.1255	0.0874	0.1346	0.2016	0.0326
	DALLRec	0.0412	0.0734	0.1276	0.0895	0.1375	0.2102	0.0351
	Imporve	↑5.10%	↑3.23%	↑1.67%	↑2.40%	↑2.15%	↑4.26%	↑7.67%
MMSSL	Base	0.0469	0.0943	0.1960	0.0888	0.1360	0.2309	0.0356
	DALLRec	0.0502	0.1001	0.2064	0.0913	0.1431	0.2436	0.0402
	Imporve	↑7.03%	↑6.15%	↑5.31%	↑2.82%	↑5.22%	↑5.50%	↑12.9%

We fine-tuning LLM on llama-Factory,Please refer to the specific operation for details:https://github.com/hiyouga/LLaMA-Factory

We generate enhanced data by running Generate enhanced data.py