关注于评分预测问题
- 利用用户对商品的评论文本进行建模, 获取用户向量以及商品向量表示
- 利用DSN [1] 等Domain Adaptation方法构建用户在源领域与目标领域上的关联
- 利用待推荐用户在目标领域上的向量为用户在源领域上进行推荐
所有的实验基于Amazon公开的数据集 下载地址: http://jmcauley.ucsd.edu/data/amazon/
Good Luck
This is a paragraph. Its' quite short.
├── CDRTR
│ ├── core
│ │ ├── DeepModel: DL代码路径, 包括SentiRec, DSN, DSNRec
│ │ └── LinearModel: 包括LR
│ ├── dataset: 为SentiRec, DSNRec 编写的数据集类
│ ├── preprocess: 文本预处理工具, 包括, 两个领域词典生成,
词向量化, 统计领域用户交叠情况, 冷用户统计等
├── exam: 实验代码路径
│ ├── data: 样例数据集
│ │ ├── preprocess: 预处理数据输出路径
│ │ └── source: 原数据存放路径
│ ├── log
│ │ ├── baseline
│ │ ├── debug_generateVoca_MultiCross
│ │ ├── debug_mergeUI_MultiCross
│ │ ├── DSNRec
│ │ └── sentitrain
│ ├── MultiCross: 多领域相互跨领域推荐实验数据路径
│ ├── Beauty_Clothin_Shoes_and_Jewelry
│ ├── Beauty_Movies_and_TV
│ ├── Kindle_Beauty
│ ├── Kindle_Clothin_Shoes_and_Jewelry
│ ├── Kindle_Movies_and_TV
│ └── Movies_and_TV_Clothin_Shoes_and_Jewelry
├── tests: 测试代码路径
在exam目录下面有一个makefile, 内容如下:
export PYTHONPATH=..
transCSV:
python -m CDRTR.preprocess csv_format --mode $(MODE) --fields reviewerID,asin,overall --data_dir $(DATA)
# make generatevoca MODE=DEBUG DATA=data
generatevoca:
python -m CDRTR.preprocess generatevoca --mode $(MODE) --fields asin,reviewerID,overall,reviewText,unixReviewTime --data_dir $(DATA)
# make extractinfo MODE=DEBUG
extractinfo:
python -m CDRTR.preprocess extractinfo --mode $(MODE)
preprocess: generatevoca extractinfo
# make sentitrain DATA=data DOMAIN=Music
# make sentitrain DATA=data DOMAIN=Auto
sentitrain:
python -m CDRTR.core.DeepModel.SentiRec --dir $(DATA)/preprocess --domain $(DOMAIN) --filter_size 3,5,7,11 --epoches 400
mergeUI:
python -m CDRTR.preprocess mergeui --mode $(MODE) --data_dir $(DATA)
DSNRec:
python -m CDRTR.core.DeepModel.DSNRec --data_dir $(DATA) --src_domain $(SRCDO) --tgt_domain $(TGTDO) --epoches $(EPOCH) --mode $(MODE)
makefile中设置了多个任务, 包括:
- generatevoca: 预处理原json数据(存放在data/source/), 该任务将统计两个领域数据, 生成词典, 冷用户情况等;
- transCSV: 将json格式的原数据处理为csv格式, 方便使用MyMediaLite进行baseline实验
- sentitrain: 对两个领域用户评论评分数据进行sentiRec [2] 训练, 获取CNN层输出作为评论句子的向量表示
- mergeUI: 将sentitrain输出的评论文本聚合为用户, item的向量表示
- DSNRec: 对两个领域数据进行跨领域推荐
exam目录下的pipeline.sh脚本文件, 包含了上述几个步骤的运行
Usage pipeline.sh data_dir src_domain tgt_domain epoch
example:
winchua@CCNLForDL:~/CrossDomainReviewTextRecommendation/CDRTR/exam$ ./pipeline.sh data Auto Musi 400 ... ... # 运行了多个实验之后, 可以通过如下脚本查看各个领域针对完全冷启动用户的评分预测RMSE结果 winchua@CCNLForDL:~/CrossDomainReviewTextRecommendation/CDRTR/exam$ python show_result.py `find . -name "test*.pk" | grep -v MultiCross` +-------------+----------------+ | domain | rmse | +-------------+----------------+ | Cell | 1.02770589547 | | Digital | 0.9646660533 | | Electronics | 1.09582470097 | | Kindle | 1.01310818718 | | Musi | 0.737983482096 | | Tools | 0.94334647007 | | CDs | 1.00803053039 | | Video | 0.987097572114 | | Jewelry | 1.04308169038 | | Movie | 1.00178416516 | | Beauty | 1.09342681075 | | Sports | 0.901738399564 | | Auto | 0.937711151504 | | Office | 0.855115969443 | | Toys | 0.911580666733 | +-------------+----------------+ winchua@CCNLForDL:~/CrossDomainReviewTextRecommendation/CDRTR/exam$ python show_user_record_count.py `ls -F | grep / | grep -v log | grep -v MultiCross | cut -d/ -f1` +----------------------------------------------+--------------+--------------+--------------+------------------+-----------------+ | domains | srcUserCount | tgtUserCount | overlapCount | srcOverlapRate | tgtOverlapRate | +----------------------------------------------+--------------+--------------+--------------+------------------+-----------------+ | Beauty/Clothing_Shoes_and_Jewelry | 18143 | 35167 | 4220 | 0.232596593728 | 0.11999886257 | | CDs_and_Vinyl/Electronics | 68998 | 186143 | 6260 | 0.090727267457 | 0.0336300586109 | | Cell_Phones_and_Accessories/Video_Games | 26480 | 22904 | 1399 | 0.052832326284 | 0.0610810338805 | | Toys_and_Games/Digital_Music | 19233 | 5362 | 179 | 0.00930692039723 | 0.0333830660201 | | Sports_and_Outdoors/Grocery_and_Gourmet_Food | 33563 | 12646 | 2035 | 0.0606322438399 | 0.160920449154 | | Kindle_Store/Movies_and_TV | 65469 | 121206 | 2754 | 0.0420657104889 | 0.0227216474432 | | Office_Products/Tools_and_Home_Improvement | 3012 | 14745 | 1893 | 0.628486055777 | 0.128382502543 | +----------------------------------------------+--------------+--------------+--------------+------------------+-----------------+ +----------------------------------------------+------------+---------------+------------+---------------+ | domains | srcURcount | srcOverRcount | tgtURcount | tgtOverRcount | +----------------------------------------------+------------+---------------+------------+---------------+ | Beauty/Clothing_Shoes_and_Jewelry | 149091 | 49411 | 238420 | 40257 | | CDs_and_Vinyl/Electronics | 968881 | 128711 | 1589951 | 99237 | | Cell_Phones_and_Accessories/Video_Games | 179952 | 14487 | 211931 | 19849 | | Automotive/Musical_Instruments | 20117 | 356 | 9842 | 419 | | Toys_and_Games/Digital_Music | 164876 | 2721 | 61975 | 2731 | | Sports_and_Outdoors/Grocery_and_Gourmet_Food | 273513 | 22824 | 114343 | 36911 | | Kindle_Store/Movies_and_TV | 932521 | 50098 | 1628754 | 68779 | | Office_Products/Tools_and_Home_Improvement | 26170 | 27088 | 111009 | 23467 | +----------------------------------------------+------------+---------------+------------+---------------+ winchua@CCNLForDL:~/CrossDomainReviewTextRecommendation/CDRTR/exam$ python baseline_result_show.py +----------------------------+-------------------------------+---------------------------+----------+-------------+---------+-----------------+---------------+-------------+-----------------------------+ | domain | FactorWiseMatrixFactorization | BiasedMatrixFactorization | SlopeOne | SVDPlusPlus | Random | BiPolarSlopeOne | GlobalAverage | ItemAverage | LatentFeatureLogLinearModel | +----------------------------+-------------------------------+---------------------------+----------+-------------+---------+-----------------+---------------+-------------+-----------------------------+ | Musi | 0.74917 | 0.75819 | 0.74003 | 0.7429 | 2.07642 | 0.74003 | 0.74003 | 0.79938 | 2.5739 | | Toys_and_Games | 0.91575 | 0.90545 | 1.01175 | 0.94514 | 1.96148 | 1.01175 | 1.01175 | 0.91091 | 2.22349 | | Tools_and_Home_Improvement | 0.92429 | 0.93079 | 0.94901 | 0.93271 | 2.02191 | 0.94901 | 0.94901 | 0.98844 | 3.01554 | | Jewelry | 1.07836 | 1.08417 | 1.09712 | 1.08301 | 2.02008 | 1.09712 | 1.09712 | 1.12147 | 2.63191 | | Beauty | 1.1154 | 1.12241 | 1.14271 | 1.11946 | 2.01473 | 1.14271 | 1.14271 | 1.16054 | 2.46513 | | Office_Products | 0.86723 | 0.86452 | 0.91 | 0.88177 | 1.99483 | 0.91 | 0.91 | 0.89912 | 2.39326 | | Auto | 0.91939 | 0.91734 | 0.95858 | 0.93183 | 2.06752 | 0.95858 | 0.95858 | 0.95209 | 2.65961 | | CDs_and_Vinyl | 0.95356 | 0.95376 | 1.00893 | 0.96628 | 2.02005 | 1.00893 | 1.00893 | 0.96833 | 2.60192 | | Cell_Phones | 1.04561 | 1.05164 | 1.09231 | 1.05525 | 1.98757 | 1.09231 | 1.09231 | 1.06846 | 2.19888 | | Grocery | 0.98135 | 0.98238 | 1.0676 | 0.99329 | 1.94582 | 1.0676 | 1.0676 | 0.99631 | 2.15443 | | Sports | 0.90509 | 0.90759 | 0.93544 | 0.9153 | 2.0045 | 0.93544 | 0.93544 | 0.94513 | 2.72289 | | Video_Games | 1.00654 | 1.00706 | 1.08674 | 1.02059 | 1.96103 | 1.08674 | 1.08674 | 1.02006 | 1.98473 | | Digital_Music | 0.99705 | 0.98776 | 1.08185 | 1.01934 | 1.97827 | 1.08185 | 1.08185 | 0.99791 | 2.34118 | | Kindle | 0.99793 | 0.9961 | 1.04402 | 1.01122 | 1.96238 | 1.04402 | 1.04402 | 1.0107 | 2.61299 | | Electronics | 1.07022 | 1.07283 | 1.12268 | 1.08046 | 2.01596 | 1.12268 | 1.12268 | 1.08378 | 1.91997 | | Movies_and_TV | 0.97773 | 0.96989 | 1.09381 | 0.99628 | 1.96544 | 1.09381 | 1.09381 | 0.96867 | 1.73305 | +----------------------------+-------------------------------+---------------------------+----------+-------------+---------+-----------------+---------------+-------------+-----------------------------+
winchua@CCNLForDL:~/CrossDomainReviewTextRecommendation/CDRTR/exam$ cd log/DSNRec/Auto_Musi/ winchua@CCNLForDL:~/CrossDomainReviewTextRecommendation/CDRTR/exam/log/DSNRec/Auto_Musi$ ls test_mses.pk train Untitled.ipynb winchua@CCNLForDL:~/CrossDomainReviewTextRecommendation/CDRTR/exam/log/DSNRec/Auto_Musi$ tensorboard --logdir=train --port 12345
| [1] | Bousmalis, K., Trigeorgis, G., Silberman, N., Krishnan, D., & Erhan, D. (2016). Domain SeparationNetworks, (Nips). Retrieved from http://arxiv.org/abs/1608.06019 |
| [2] | Hyun, D., Park, C., Yang, M.-C., Song, I., Lee, J.-T., & Yu, H. (2018). Review Sentiment-Guided Scalable DeepRecom-mender System. Ann SIGIR, 18, 965–968. https://doi.org/10.1145/3209978.3210111 |
✨🍰✨