Paula Gual
Data Part Time Barcelona May 2019
In this project we will create a Machine Leaning Model to predict if a visit of the Google Merchandising store will buy or not.
The dataset I will be using for the project is basically data from the Google Analytics account of the Google Merchandising store. This dataset is available on Kaggle.
[Google Analytics from Google Merchandising Store]https://www.kaggle.com/colinpearse/ga-analytics-with-json-columns
I've imported all the Dataset in a pandas dataframe, using the chunk method, as the original data csv is huge.
I've classified all the variables on my dataset:
CATEGORICAL:
- channelGrouping: Channel of arrival of the visit
- fullVisitorId: Visitor ID
- sessionId: Session ID (contains Visitor ID)
- socialEngagementType: Type of engagement
- deviceBrowser: Name of the Browser
- deviceCategory: Name of the device category
- deviceOS: The OS of the device
- geoCity: City of the visit
- geoContinent: Continent of the visit
- geoCountry: Country of the visit
- geoMetro: Metro of the visit
- geoDomain: Domain of the visit
- geoRegion: Region of the visit
- geoSubContinent: Subcontinent of the visit
- trafficSourceAdContent: Source of the Ads
- adwordsClickAdNetworkType: Ad Source
- adwordsClickGclId: Ad Click ID
- adwordsClickInfo.slot: Place of the ad
- trafficSourceCampaign: Name of the souce campaign
- trafficSourceKeyword: Name of the source Keyword
- trafficSourceMedium: Name of the source Medium
- trafficSourceReferralPath: Referral Path
- trafficSourceSource: Name of the source
BOOLEAN:
- deviceIsMobile: is the device mobile
- totalsBounces: Is bounce
- totalsNewVisits: Is new visit
- adwordsClickIsVideoAd: Is from ad video
- trafficSourceIsTrueDirect: Is it a true direct visit
ORDINAL:
- visitId: Visit ID integer (is the date in unix time)
- visitNumber: The number of visit of one user
NUMERICAL:
INTEGER:
- totalsHits: number of hits
- totalsPageviews: number of pageviews
- totalsTransactionRevenue: (TARGET) logaritm if the revenue
- totalsVisits: Number of visits
- adwordsClickInfoPage: click from ad in page
DATE:
- date: Date of the visit
- visitStartTime: unix time of the visit
I've preprocessed and cleaned all the data vars, manly replace the Nans using different strategies, depending on the data. I have also fixed some uncorrect types, of the variables.
I have divided the EDA in two parts: EDA of the numeric vars and the EDA of the categorical vars.
For the EDA of the numeric vars, I have first analysed the Target. Then the other numeric variables, by themselves (distribution) and comparing them to the target. Also I have visualized all the data, manly with scatterplots and boxplots.
I have also checked the correlation, and deleted a variable (hits) that was highly correlated with another (Pageviews).
For the EDA of the categorical vars I have studied the different values for each one, and also compared those number of values depending on the target var. Also I have visualized all the data, manly with bar charts.
In some cases, as the variable had many posible outcomes, I have created a new var with only the most repeated values, or the most related values to the target.
I have selected the features that based on the EDA I belive are more related to the target.
The categorical vars I have encoded them, and craeted dummies.
In the model selection process I will take into account: 1.- It is a binaty classification problem 2.- The data is Sparse 3.- The target is very unbalanced (1,29% vs 98,71) 4.- There is a lot of data 903455 rows and 48 columns
Model Selection 1.- I'll use a decision tree 2.- I'll try to balance the classes down sampling the data 3.- I'll use Under ROC Curve (AUROC) as metric
Algorithm Ramdom Forest Parameters {'bootstrap': True, 'criterion': 'gini', 'max_depth': 9, 'max_features': 'auto', 'n_estimators': 100}
The final results of the model are:
F1: 0.96 Confusion Matrix: (2144, 159) (30, 2273) AUC: 0.96
Feature Importance Importance / Feature 0.5045 / totalsPageviews 0.1228 / totalsBounces 0.0820 / continet_Americas 0.0374 / source_Googleplex 0.0315 / referral_Home
The conclusion is that it is possible to predict the conversion in an ecommerce, only having the data from the Google Analytics data of the visits.