Where To Sell Products

Author: Luis Eduardo Ferro Diez contact@ohtar.mozmail.com

This repository contains all my work for my MsC in Computer Science project.

Where To Sell Products (wtsp) is a project in which I try to solve this very same question. The idea is to characterize geographic areas in terms of its relationship with a selected set of products. The relationship is derived from geotagged texts, i.e., I gathered geotagged text data from Twitter and product reviews from Amazon. For the former I generated spatial clusters from which all the tweets are aggregated to form a single cluster-corpus. For the latter I trained a convolutional neural network to classify product categories given several review texts. Finally each cluster-corpus is submitted to the classifier to emit a relevance score for some categories. The result is displayed on a map of an area of interest, e.g., a city, in which the clusters are shown with their corresponding relationship score with certain products categories.

Demo

Los Angeles (2013-07)

Vancouver (2013-07)

New York (2013-07)

Live Demo

City Date
Los Angeles June 2013
Los Angeles July 2013
Los Angeles August 2013
Los Angeles September 2013
New York June 2013
New York July 2013
New York August 2013
New York September 2013
Vancouver June 2013
Vancouver July 2013
Vancouver August 2013
Vancouver September 2013
Toronto June 2013
Toronto July 2013
Toronto August 2013
Toronto September 2013

Repository Structure

System Requirements

Workflow

1. Gather the data

2. Execute the data preparation pipelines

Both twitter and product review data needs to be pre-processed, for this there are two spark projects under dataprep.

Tweets Transformer

This job takes the twitter data, filters out the tweets that are not geotagged and sinks the result as parquet files.

Amazon Product Reviews Transformer

This job takes the raw amazon product reviews and product metadata and converts them into ‘documents’ where each document has categories and either a review text or a product description.

3. Train the product classifier

Install the cli and follow the instructions to create the embeddings and train the classifier with the transformed product documents.

4. Predict the geographic area categories

Use the cli to predict the detect and classify the geographic areas.

Detailed process

A more detailed process is written in jupyter notebooks here

Results

Artifacts

We generated several artifacts that we made publicly available for further research, which can be downloaded from this link, the content is described as follows:

Usage

Product document embeddings

To load and use the generated product document embeddings:

Data

import numpy as np

embeddings_path = "/path/to/document_embeddings.npz"
embeddings = np.load(embeddings_path)

X_train = embeddings['x_train']
X_test = embeddings['x_test']
Y_train = embeddings['y_train']
Y_test = embeddings['y_test']

Label decoder

import numpy as np
from sklearn.preprocessing import MultiLabelBinarizer
import pickle

path = '/path/to/category_encoder.model'
with open(path, 'rb') as file:
    categories_model = pickle.load(file)

categories_model.inverse_transform(np.array([[1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]))

Doc2Vec model

To load and use the Doc2Vec model to infer new document embeddings:

from gensim.models.doc2vec import Doc2Vec
from nltk import word_tokenize
from gensim.models.doc2vec import TaggedDocument

# Loading the model
model_path = '/path/to/d2v_model.model'
d2v_model = Doc2Vec.load(model_path)

# Inferring a new vector
paragraph = "I want to buy a big TV for my bedroom!"
tokens = word_tokenize(paragraph)
tag_doc = TaggedDocument(words=tokens, tags=['Technology'])
doc_embedding = d2v_model.infer_vector(tag_doc.words)

Product document classifier

To load and use the keras model to predict document categories:

from keras.engine.saving import model_from_yaml

base_path = '/path/to/classifier/files'
ann_def_path = f"{base_path}/prod_classifier-def.yaml"
ann_weights_path = f"{base_path}/prod_classifier-weights.h5"

with open(ann_def_path, 'r') as file:
    prod_predictor_model = model_from_yaml(file.read())

prod_predictor_model.load_weights(ann_weights_path)
prod_predictor_model.summary()

Docker

I have created a docker image with a conda environment and cli pre-installed to execute the modeling part after the data engineering pipelines, and configured to run experments starting from pre-processed data https://hub.docker.com/r/ohtar10/wtsp.

The previously mentioned artifacts can be used with the docker image for prediction and characterization. After downloading the Doc2Vec and the classifier models, place them anywhere in your machine with the following structure:

workspace
└──products
    └── models
        ├── classifier
        │   ├── category_encoder.model
        │   ├── classification_report.png
        │   ├── prod_classifier-def.yaml
        │   ├── prod_classifier-weights.h5
        │   └── training_history.png
        └── embeddings
            ├── d2v_model.model
            ├── d2v_model.model.trainables.syn1neg.npy
            └── d2v_model.model.wv.vectors.npy

Next, download the docker image from docker hub:

docker image push ohtar10/wtsp:0.1.1

Then, you can use the this docker-compose file as template to start the container. Remember to set the environment variable WORK_DIR to the root folder of the workspace folder mentioned above. And start the container:

docker-compose up

Finally, ssh into the container and start using the cli

wtsp --help

wtsp predict clusters --filters "place_name=Toronto,country_code=CA" --params center='43.7;-79.4',eps=0.005,n_neighbors=10,location_column=location_geometry,min_score=0.1 /path/to/preprocessed/twitter/data/

Acknowledgements

This work could have not been done without the help and guidance of my advisors: Dr. Norha M. Villegas, and Dr. Javier Díaz Cely, whom I owe my utmost gratitude.

License

GNU GENERAL PUBLIC LICENSE Version 3, 29 June 2007

See the LICENSE_ file in the root of this project for license details.