This project was developed for the course of Cloud Applications and Systems at the University of Modena and Reggio Emilia during my MSc Degree.
The goal of this project is the implementation of a recommendation system that suggests users with wine varieties based either on a generic description or a list of matching adjectives. The suggestion subsystem is based on the Wine Reviews dataset, containing ~130k wine reviews and is powered by two separate Machine Learning models. The feature considered for suggestions is the variety feature. Since the target of the course is to focus on modern cloud architectures, the components of this project are structured for deployment on Google Cloud Platform.
The recommendation system has three main components:
- Chatbot, to handle interaction with users. Implemented through GCP's Dialogflow service that streamlines the process of building a real-feeling conversation manager.
- Classification models, to classify two kinds of requests coming from the user into the best fitting wine variety for their taste. Data cleaning and model training was done with
scikit-learn, which is conveniently supported by GCP's AI Platform - Online Predictions service. - Fulfillment handler, to connect the chatbot with the ML models. A straightforward python script deployed on GCP's Cloud Functions.
The system relies on two different classification models to provide both a generic desciption based search and a more guided and first-timers friendly path.
The first model, description-classification, takes the description column as input feature. The goal of this model is to match a given description with the wine variety that most likely fits it. The problem is essentially a NLP task, the descriptions in the dataset are discretized with a Term Frequency - Inverse Document Frequency procedure after being cleaned of any unnecessary words.
For the second model, adjective-classification, a finite number of specific adjectives (collected from here) is extrapolated from the descriptions in order to obtain a presence table of every adjective in a description. The model takes a list of same adjectives collected by the chatbot to obtain a classification.
Due to the greatly imbalanced nature of the dataset against the variety feature, only varieties with a number of descriptions above the mean have been considered for training the models. To build the adjective-classification model a supersampling technique was applied to improve the final performance.
A step-by-step description for building the adjective-classification model is available in the adjective-classification.ipynb file. The same detailed description is not yet available for the description-classification model but is in the works.
GrapeMaster.zipis the exported configuration of the Dialogflow agentfulfillmentcontains the python script that connects Dialogflow and the two modelsdescription-classificationcontains the files related to data cleaning, model training and model deployment for the description-based modeladjective-classificationcontains the files related to data cleaning, model training and model deployment for the adjective-based modelreportcontains the material shown in this README as well as the project report that was delivered (at the moment only available in italian)