mariolito / metadata-embeddings-implementation

Metadata User & Item Embeddings by M Kula 2015 implementation & comparison with LightFM, using a Starbucks offers dataset

Metadata User & Item Embeddings - Hybrid Recomender System

In this project, we implement the main algorithm of M Kula 2015 and compare findings with LightFM library, as a sanity check.

Using a Starbucks offers transactions dataset, we firstly transform data to:

• Users' features from several binned features of Starbucks customers
• Items' features from several binned features of Starbucks offers
• Binary transaction Data from Starbucks completed-declined offers

After splitting data to train and test transaction datasets, we train and predict recommendations for all test users, using MetadataEmbeddings algorithm and LightFM library, and compare results.

Dependencies

• Install anaconda3

• Activate virtual enviroment

sudo pip install --upgrade virtualenv
mkdir venvs
virtualenv my_venv
source my_venv/bin/activate

• Install python libraries

pip install -r requirements.txt

Data

Download Starbucks data

• Download data from starbucks-data or

kaggle datasets download -d ihormuliar/starbucks-customer-data
unzip starbucks-customer-data.zip

• Copy all files to data folder of this project

Data Explanation

• porfolio: This dataset contains unique offers that were promoted to customers. Here, every offer is described by its properties.
• profile: This dataset contains unique Customer ids. Here, every customer is described by its properties.
• transcript: This dataset contains events occurred for several customers. In this exercise, we are interested on events that customer interacted with some offers.

Processes

Preprocess (src/preprocessing.py)

This script reads dataset, creates and stores all aforementioned features, and train-test datasets.

cd src/
python  preprocessing.py

Evaluation (src/evaluate.py)

This script reads the preprocessed datasets and feeds in corresponding data to LightFM model and MetadataEmbeddings model.

cd src/
python3 evaluate.py

Next, we report MAP@3 evaluated on test dataset for both models. As we can see both metrics are similar, and the same holds for the recommended product-offers.

Results for test sample

	MAP@3
LightFM	0.362
MetadataEmbeddings	0.361

• LightFM first 10 users-items recommendations:

[1, 7, 6, 8, 2, 4, 0, 5, 3, 9]
[1, 9, 2, 7, 6, 4, 0, 5, 3, 8]
[1, 7, 2, 4, 0, 8, 6, 5, 3, 9]
[1, 9, 8, 6, 4, 5, 3, 0, 2, 7]
[1, 9, 6, 8, 2, 7, 4, 0, 5, 3]
[1, 9, 2, 8, 7, 4, 0, 5, 3, 6]
[9, 1, 7, 6, 8, 2, 5, 3, 0, 4]
[1, 9, 2, 4, 0, 8, 6, 5, 3, 7]
[9, 7, 2, 8, 6, 0, 4, 5, 3, 1]
[9, 1, 7, 2, 8, 6, 4, 5, 3, 0]

• MetaDataEmbeddings first 10 users-items recommendations:

[1, 6, 7, 4, 0, 8, 2, 3, 5, 9]
[1, 9, 6, 7, 4, 0, 2, 3, 5, 8]
[1, 6, 7, 4, 8, 0, 2, 3, 5, 9]
[1, 9, 6, 8, 4, 3, 5, 0, 2, 7]
[1, 9, 6, 8, 2, 7, 4, 0, 3, 5]
[1, 9, 8, 2, 7, 4, 0, 3, 5, 6]
[1, 6, 9, 7, 8, 2, 3, 5, 0, 4]
[1, 9, 6, 4, 8, 0, 2, 3, 5, 7]
[9, 6, 7, 4, 0, 8, 2, 3, 5, 1]
[1, 9, 6, 7, 4, 8, 2, 3, 5, 0]

Finally, we visualize the Items space implied by Items embeddings, along with 3 users that had 2 transactions on train dataset. In order to visualize our features against our target variables, and evaluate the explanatory power of our features, we perform a TSNE dimensionality reduction algorithm. Thus, we project all feature embedding, to a 2-D space

User 5, had bought items [0 9]. Top 2 recos are: [1, 6]
User 7, had bought items [8 9]. Top 2 recos are: [1, 6]
User 18, had bought items [0 4]. Top 2 recos are: [1, 9]