# Smart offers project

## Recommendation system
Merchant and banks have always aimed to maintain a vital relation with consumers and costumers. With the emergence of Big data and data science, the hope is to leverage the consumers database and history to better understand his needs and behavior. Such knowledge is helpful to incentify costumers and avoid market stagnation. It also serves in creating a good recommendation and offers that align with consumers needs and choices. MOBI724 is developing a cutting-edge technology in automated offers based on consumers information and previous transactions. The objective of an offer can lie under four main task: Retain, Acquire, Shift and Lift.
## RFMLP based

## Tensor factorization

### Papers

- [Low-Rank Regression with Tensor Responses](http://papers.nips.cc/paper/6302-low-rank-regression-with-tensor-responses.pdf)

- [Probabilistic Matrix Factorization](https://papers.nips.cc/paper/3208-probabilistic-matrix-factorization.pdf)

- [A Spatial-Temporal Probabilistic Matrix Factorization Model for
Point-of-Interest Recommendation](https://pdfs.semanticscholar.org/b8ab/2cc7ea6f621ac4cb2fea6523b72400a733d1.pdf)

## Reinforcement learning

### Papers

- [Residual Policy Learning](https://arxiv.org/pdf/1812.06298.pdf): This paper is useful to learn a policy on top of non RL policies ( e.g., PID controller)

## Deep learning approach

### Papers

- [A review on deep learning for recommender systems: challenges and remedies](https://link.springer.com/content/pdf/10.1007%2Fs10462-018-9654-y.pdf)
- [Sequence-Aware Recommender Systems](https://arxiv.org/pdf/1802.08452.pdf)
- [Deep Learning based Recommender System: A Survey and New Perspectives](https://arxiv.org/pdf/1707.07435.pdf)

- [Applying Deep Learning To Airbnb Search](https://arxiv.org/pdf/1810.09591.pdf)

- [Wide & Deep Learning for Recommender Systems](http://delivery.acm.org/10.1145/2990000/2988454/p7-cheng.pdf?ip=132.207.4.76&id=2988454&acc=OA&key=FD0067F557510FFB%2EC32CC723E17B05B2%2E4D4702B0C3E38B35%2E5945DC2EABF3343C&__acm__=1571236851_65a5ebf503d7d4879226bd19b48dd751)

- [Deep learning for recommender systems](https://www.researchgate.net/profile/Balazs_Hidasi/publication/319277717_Deep_Learning_for_Recommender_Systems/links/59ae6dc10f7e9bdd11627444/Deep-Learning-for-Recommender-Systems.pdf)
- [A Hybrid Collaborative Filtering Model with
Deep Structure for Recommender Systems](https://pdfs.semanticscholar.org/1b1a/d82c3e13d09889d391af095e477d7fcb4be9.pdf)


- Description:
- This paper proposes an autoencoder for the users/items vectors where it is trained to reconstruct the input and encode the signal in a way that is able to reconstruct the vector using dot product. We encode the user vector and the item vector in a way the encoding predict the $R_{ij}$

- [Deep Neural Networks for YouTube Recommendations](http://delivery.acm.org/10.1145/2960000/2959190/p191-covington.pdf?ip=132.207.4.76&id=2959190&acc=OA&key=FD0067F557510FFB%2EC32CC723E17B05B2%2E4D4702B0C3E38B35%2E5945DC2EABF3343C&__acm__=1571237102_47f7fe6190ec0ce9883c603410716aa6)
- [Deep Matrix Factorization Models for Recommender Systems∗](https://pdfs.semanticscholar.org/35e7/4c47cf4b3a1db7c9bfe89966d1c7c0efadd0.pdf?_ga=2.107435396.74165196.1571237184-1869724326.1571237184)
- [https://www.ijcai.org/proceedings/2017/0239.pdf](https://www.ijcai.org/proceedings/2017/0239.pdf)
- [xDeepFM: Combining Explicit and Implicit Feature Interactions
for Recommender Systems](https://arxiv.org/pdf/1803.05170.pdf)
- [Neural Collaborative Filter](https://arxiv.org/pdf/1708.05031.pdf)
- [Gated Attentive-Autoencoder for Content-Aware Recommendation](https://arxiv.org/pdf/1812.02869.pdf)
- [Deep Interest Network for Click-Through Rate Prediction](https://arxiv.org/pdf/1706.06978.pdf)
- [Deep Crossing: Web-Scale Modeling without Manually Crafted Combinatorial Features](https://www.kdd.org/kdd2016/papers/files/adf0975-shanA.pdf)
- [Attentional Factorization Machines:Learning the Weight of Feature Interactions via Attention Networks](https://www.ijcai.org/proceedings/2017/0435.pdf)
- [Wide & Deep Learning for Recommender Systems](https://arxiv.org/pdf/1606.07792.pdf)
- [Deep Learning over Multi-field Categorical Data](https://arxiv.org/pdf/1601.02376.pdf)
- [Product-based Neural Networks for User Response
Prediction](https://arxiv.org/pdf/1611.00144.pdf)
- [Deep Interest Evolution Network for Click-Through Rate Prediction
](https://arxiv.org/pdf/1809.03672.pdf)
## Misc

- [Sequence-Aware Recommender Systems
](https://arxiv.org/pdf/1802.08452.pdf)

## Useful Links

- [Introduction to Recommender System
](https://towardsdatascience.com/intro-to-recommender-system-collaborative-filtering-64a238194a26)

- [OpenLearning4DeepRecsys](https://github.com/Leavingseason/OpenLearning4DeepRecsys)

## Evaluation Methods

### Recommeder system accuracy performance on testing dataset
In general, the performance of a machine learning models on testing subset is computed by different metrics: Accuracy, False Positive Rate, AUC-ROC. In case of recommendation systems, it is a bit tricky to split the the dataset into training and testing subset. Each user and each item apart needs to be trained (update its features) at least once. Thus it is not possible to test on a user/item that have no encounters in the training set. Therefore, We convert of pivot table into a binary table. 0 for NaN and 1 for a positive score. We sum the table along both axis ( rows and columns) in order to check how many positive values we have for each users and items. If a user or item have 2 and more matching with items and users respectively then this chunk is candidate to be a test sample.
Next, we need to choose our loss function. In the case of RFMLP scores we can either choose MSE or cross entropy classification loss.

### Correlation between learned features clustering and computed score for each Item