lambda-opt

This repo provides our implementation for λOpt. λOpt is a fine-grained adaptive regularizer for recommender models. The key idea is to specify an individual regularization strength for each user and item. Check our paper for details about λOpt if you are interested!

Yihong Chen, Bei Chen, Xiangnan He, Chen Gao, Yong Li, Jian-Guang Lou and Yue Wang(2019). λOpt: Learn to Regularize Recommender Models in Finer Levels. In KDD'19, Anchorage, Alaska, USA, August 4-8, 2019.

Usage

1. Put the data in src/data/

For example, movielens-1m, put ratings.dat in src/data/ml-1m

cd src
mkdir data
cd data 
mkdir ml-1m

2. Create folders for tmp files in src/tmp/

cd src
mkdir tmp
mkdir tmp/data
mkdir tmp/res
mkdir tmp/res/ml1m/
mkdir tmp/penalty/
mkdir tmp/penalty/ml1m/

3. Specify the hyper-parameters in train_[dataset]_[regularization method].py

For fixed lambda on movielens-1m, the hyper-parameters are in train_ml1m_fixed.py For lambdaopt on movielens-1m, the hyper-parameters are in train_ml1m_alter.py

4. Train

cd src
python train_ml1m_alter.py

Files

• src/regularizer/: lambdaopt
• src/factorizer/: matrix factorization model
• src/utils/: handy modules for data sampling, evaluation etc.
• src/engine.py: training engine
• src/train_[dataset]_[regularization method].py: entry point

What's new?

λOpt features a parameterized regularizer, where very fine-grained regularization is employed over the user and item embeddings. While itself offers a novel technique to solve the bi-level optimization problem, there are also a bunch of cool work from the meta-learning community. Hence, to put readers in a broader context, this repo also offers the implementation of λOpt using Higher, a pytorch meta-learning repo. Higher provides useful tookits, like differentiable optimizers and automatic patch to make pytorch model functional, which facilitates convenient computation of higher-order gradients.

Comparison between the original λOpt and λOpt using higher is shown in the following table.

	λOpt Original	λOpt Higher
Gradient Path	Truncated 2nd-order	Fully 2nd-order
Memory Cost	1X	~1.2X
Computational Cost	1X	~1.7X
Require manual gradient composition?	Yes	No
Require duplicating model forward logic?	Yes	No
Require manual differentiable optimizer?	Yes	No
Concise code	No	Yes

Note that λOpt Higher takes more computational resources but offers a very concise and beautiful code:

       with higher.innerloop_ctx(m, m_optimizer) as (fmodel, diffopt):
            # look-ahead, this is very similar to factorizer update except that lambda is included in the computational graph
            u, i, j = sampler.get_sample('train')
            preference = torch.ones(u.size())[0]
            if self.use_cuda:
                u, i, j = u.cuda(), i.cuda(), j.cuda()
                preference = preference.cuda()
            prob_preference = fmodel.forward_triple(u, i, j)
            l_fit = self.criterion(prob_preference, preference) / (u.size()[0])
            lmbda = self.lambda_network.parse_lmbda(is_detach=False)
            l_reg = fmodel.l2_penalty(lmbda, u, i, j) / (u.size()[0])
            l = l_fit + l_reg
            diffopt.step(l)

            # compute the validation loss
            valid_u, valid_i, valid_j = sampler.get_sample('valid')
            valid_preference = torch.ones(valid_u.size()[0])
            if self.use_cuda:
                valid_preference = valid_preference.cuda()
                valid_u, valid_i, valid_j = valid_u.cuda(), valid_i.cuda(), valid_j.cuda()

            self.lambda_network.train()
            self.optimizer.zero_grad()
            valid_prob_preference = fmodel.forward_triple(valid_u, valid_i, valid_j) / valid_u.size()[0]
            l_val = self.criterion(valid_prob_preference, valid_preference)
            l_val.backward()
            torch.nn.utils.clip_grad_norm_(self.lambda_network.parameters(), self.clip)
            self.optimizer.step()
            self.valid_mf_loss = l_val.item()
            return self.lambda_network.parse_lmbda(is_detach=True)

How to run λOpt in higher?

• simply specify the regularizer as alter_mf_higher

Requirement

• pytorch=1.2.0
• higher

Citation

If you this repo, please kindly cite our paper.

@inproceedings{lambdaopt,
 author = {Chen, Yihong and Chen, Bei and He, Xiangnan and Gao, Chen and Li, Yong and Lou, Jian-Guang and Wang, Yue},
 title = {lambdaOpt: Learn to Regularize Recommender Models in Finer Levels},
 booktitle = {Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \&\#38; Data Mining},
 series = {KDD '19},
 year = {2019},
 isbn = {978-1-4503-6201-6},
 location = {Anchorage, AK, USA},
 pages = {978--986},
 numpages = {9},
 url = {http://doi.acm.org/10.1145/3292500.3330880},
 doi = {10.1145/3292500.3330880},
 acmid = {3330880},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {matrix factorization, regularization hyperparameter, top-k recommendation},
} 

Contact

Any feedback is much appreciated! Drop us a line at yihong.chen@cs.ucl.ac.uk or simply open a new issue.

TODO

• Test LambdaOpt in Higher