Azure: Machine Learning Applications in Retail & Marketing
Introduction
This article is the brief overview of typically how machine learning applications can be useful for the retailers/advertisers and marketing folks and target the relevant customers for their business based on the location, marketing needs and social media influence. Following retail/marketing use cases also describes about how to target the retail customers in an algorithmic way.
Targeted Audience
Decision makers, Business Analysts, Decision Influencers, AI/Automation/Big Data Architects/Data Scientists
Technology Overview
This is the brief technology overview of implementing following use cases to get started in the world of machine learning.
Step 1: Data/Feature Engineering, Exploration and Data Cleaning
For this data scientist should co-ordinate with domain expert to understand the data thoroughly.
For example, if problem statement is Price Prediction of the house, then in the Price Prediction dataset,
- All the features (columns of the dataset) which impacts the rise or fall in the Price as output variables should be identified and other features (columns) should be eliminated.
- Data should be standardized for variance & bias & cleaned with pre-processing steps. It can be identified with data distribution graphs. (These graphs should indicate uniform distribution of the data.)
- Data should cleaned for null/missing column values.
To address above pre-requisites, Azure Machine Learning Data Preparation package can be used as one of the key methods in data preparation step. This MSDN documentation can help to get you started.
Step 2: Training the Model
Second step involves the model the data in such a way that, for given input features columns, trained model should predict the target variable accurately using best machine learning algorithms. It is just like teaching a kid for the first time and then kid learns from the past experiences. Similarly, model learns from history dataset described in the previous section.
Model can be trained in Favorite IDE, such as the Visual Studio Code/Azure Notebooks or Azure Machine Learning Designer. This MSDN documentation can help to get started.
Step 3: Evaluating the Model
Third step involves, how accurately trained Model can predict the output/target variable. For example,
After the model is developed as described in previous 2 sections, this model is evaluated against the historical values with the predicted values for the target variable. For house price prediction, model will be evaluated for the predicted price vs actual historical price.
If model is able to predict the prices accurately or close to accurate that means , it will be able to predict the prices for the unseen / future input columns and can be deployed to the production.
We can implement this in IDE of our choice, as described here. This post also describes about how to evaluate against the historical values for the predicted values.
Model can be evaluated based on if it is a Classification (i.e. Problem statements like Customer will buy or not) or Regression Prediction (i.e Problem statements like What/ would be the price of house in numbers)
This MSDN documentation can be helpful to understand how it works in Azure ML studio or SkLearn Python Libraries in Visual Studio code
To monitor the real time performance of the model in production, Azure Machine Learning Service data collection feature can be enabled.
Converting Model Output to Power BI Reports
Model output (for each record in the dataset) can be converted to CSV files and then further it can be uploaded to Power BI and visualized further for specific business need.
For this conversion libraries such as Pandas data frame in python or Azure ML Modules can be implemented. And further output can be customized/visualized in Power BI desktop. Power BI Desktop can be installed locally or it is pre-configured in data science VMs. But to publish the reports it in the Power BI portal Office 365 licences needs to be enabled for long term. To get started with this 1 Month’s trial version should be enough.
Next section describes about practical applications of machine learning in retail and marketing using above core concepts.
Prediction of Successful Campaigns
Description: This use case is about predicting the Success /Performance of the Campaigns & Identifying the important success factors to maximize the ROI & better customer engagement. The output is plotted as the report that describes about historical campaign data analysis and probabilities of success of the campaigns. This also describes about campaign or promote the business to gain maximum out of invested marketing spend.
Video : Prediction of successful campaigns for the targeted customers
Dynamic Price Forecasting
Description: This use case is about price prediction based on Competitive analysis of similar listing & dynamic price prediction, supply and demand, location and other external marketing for given time. The Model output is plotted as the reports that describes Competitive analysis of similar listing & dynamic price prediction. It is useful for hosts to gain maximum profit through optimized price for their listings.
Video : Dynamic price forecasting for the targeted customers
Prediction of the Retail Location
Description: This use case describes about the predictive profitability of a new retail location at a specific state in United States for given Marketing Spend , the Revenue & the Population. This also describes about how to find dynamic price by targeting the specific group of customers and demand analysis.
Video : Prediction of the Retail Location for the targeted Customers
Propensity Modelling
Description: This use case is about how to implement Propensity Modelling on the Browser History data of an E-commerce site to target the online customers. And then provide them info for the products of their interests through up selling and cross selling
Video : Propensity Modelling for the targeted Customers [To be added]