feat: unified Transformers4Rec with CLM/MLM/PLM/RTD training objective

[hieuddo]

Description

Previously, we ported several sequential models from https://github.com/PreferredAI/CoVE. In the context of CoVE, we experimented and found out that a specific setting leads to better performance: CLM training objective + sequence breakdown ([a,b,c,d] was broken down to [a]->b, [a,b]->c, [a,b,c]->d) and we only calculated loss at the last position (e.g., for [a,b,c]->d, we only calculated loss at the predicted output with input [a,b,c].

For general use, I think we should have the options for other training objectives (CLM, MLM, PLM, and RTD), all four derived from https://github.com/NVIDIA-Merlin/Transformers4Rec.

So here, we try to aggregate all Transformers-based models into one unified class, with all four training objectives. We still keep the one setting we believe is the best, mentioned above (CLM + session breakdown + loss at last position).

Checklist:

• I have added tests.
• I have updated the documentation accordingly.
• I have updated README.md (if you are adding a new model).
• I have updated examples/README.md (if you are adding a new example).

[hieuddo]

Quick tuning with two datasets, sorted by val_MRR descendingly.

Dataset: diginetica

combo	val_AUC	val_MRR	val_NDCG@10	val_Recall@10	test_AUC	test_MRR	test_NDCG@10	test_Recall@10	train_s
gpt2-clm-all	0.7892	0.3458	0.3775	0.4855	0.8362	0.3163	0.3502	0.4649	32.2
bert-clm-last	0.7605	0.3442	0.3719	0.463	0.7612	0.3235	0.3517	0.4482	113.9
gpt2-clm-last	0.8055	0.3341	0.3637	0.4662	0.8483	0.3324	0.3638	0.4749	105.3
xlnet-plm	0.8093	0.3157	0.3548	0.4855	0.8185	0.2713	0.3086	0.4415	84.0
xlnet-mlm	0.8064	0.3085	0.3431	0.463	0.8498	0.2601	0.3001	0.4482	51.0
bert-mlm	0.8158	0.232	0.2618	0.3762	0.8329	0.214	0.2453	0.3679	31.8
electra-mlm	0.8203	0.2129	0.2369	0.3408	0.8287	0.2092	0.2348	0.3411	34.9
bert-rtd	0.8279	0.2053	0.2233	0.3055	0.8283	0.1926	0.2234	0.3445	69.4
electra-rtd	0.8293	0.1907	0.2116	0.3087	0.8346	0.1815	0.2034	0.301	63.7

Dataset: ml100k

For ml-100k dataset, we load the data as USIT format with UUIT data (rating data doesn't have session_id)

combo	val_AUC	val_MRR	val_NDCG@10	val_Recall@10	test_AUC	test_MRR	test_NDCG@10	test_Recall@10	train_s
gpt2-clm-last	0.882	0.0673	0.0737	0.1474	0.8687	0.0618	0.0668	0.1347	1471.0
bert-clm-last	0.8871	0.0597	0.063	0.1273	0.8698	0.0514	0.0516	0.1029	1674.0
gpt2-clm-all	0.8355	0.0472	0.0488	0.1007	0.8128	0.0357	0.0327	0.071	50.5
xlnet-mlm	0.8625	0.0471	0.0494	0.1029	0.8344	0.0351	0.0344	0.0795	47.7
xlnet-plm	0.8687	0.0444	0.0423	0.0838	0.8393	0.0335	0.0307	0.0679	75.2
electra-rtd	0.8547	0.0387	0.036	0.071	0.8249	0.0269	0.0222	0.0509	53.4
bert-rtd	0.8517	0.0361	0.0365	0.0774	0.8353	0.0299	0.0279	0.0626	53.8
bert-mlm	0.8504	0.0338	0.0321	0.0668	0.8308	0.0303	0.0268	0.0541	32.4
electra-mlm	0.8575	0.0331	0.0321	0.0721	0.8366	0.0285	0.0244	0.053	35.9