- Abdulrahman Aroworamimo
- Angel Oluwole-Rotimi
- Tashfeen Ahmed
- Xingchen Luo
- Yvan Kammelu
In developing Revenue Radar for Productionization, we considered the potential cybersecurity threats and vulnerabilities crucial to ensuring the security and integrity of the application and its data. We propose two cybersecurity applications that can address attacks against machine learning systems and common web application vulnerabilities listed by OWASP.
- Justification:
- Protection Against OWASP Threats: A WAF can provide rules to defend against OWASP threats like SQL Injection, XSS (Cross-Site Scripting), and broken authentication. It can filter, monitor, and block HTTP traffic to and from a web application, which helps protect against attacks that exploit application vulnerabilities.
- Customizable Rulesets for ML Security: With WAFs, you can create custom rules that specifically protect against abnormal, malicious HTTP requests that might indicate attempts at evasion or impersonation attacks on ML models. By monitoring the nature and pattern of the requests, WAFs can help mitigate potential poisoning attacks by blocking requests that try to upload malicious data.
- Abdulrahman Aroworamimo
- Ammad Sohail
- Angel Oluwole-Rotimi
- Rodrigo Castro
- Xingchen Luo
- Yvan Kammelu
Our dataset contains transaction data from individual visits to the GStore between December 1st, 2018, and January 31st, 2019. Each line in our dataset represents a visit to the store.
The original dataset is available here: https://www.kaggle.com/competitions/ga-customer-revenue-prediction/data?select=train_v2.csv
For our project, we explore two major use cases related to customer conversion and revenue generated from conversions:
- Customer Conversion Analysis: To identify trends in user behavior that lead to conversions, enabling targeted marketing strategies.
- Revenue Optimization: To understand how conversions translate into revenue, providing insights for optimizing ad spend and marketing investments.
Our project aims to explore trends in user conversion, which can drive and improve our investment in marketing and ad spend for GStore, as well as any other retailers looking to leverage customer data to enhance their marketing spend.
Let's dive into the details of each use case for Revenue Radar.
Conversion is a key indicator to understand and optimize, as the number of people purchasing and supporting merchandise maps directly to branding goals.
- Determine which users to target for conversion nudges.
- Increase understanding of the user journey and profile of converting customers.
Model Performance
- Model performance should be measured with F1 Score for an appropriate balance of precision and recall.
Initiative Performance
- The model must outperform the current rule-based process to be piloted.
Explainable models must be explored throughout the data science process.
To facilitate user-centric analysis, the dataset was transformed from session level to user level.
- Page Views -> First Session Page Views, Last Session Page Views
- Device Category -> Number of visits by desktop, mobile, tablet
Columns were excluded based on proportions of missing data and SME (Subject Matter Expert) consultation.
See the code file in the respective labeled folder for more details on preprocessing.
Optimizing revenue is critical for sustaining operations and funding the product innovation required for competitiveness in merchandising.
- Identify which customers are likely to spend more at GStore, focusing on understanding their spending behaviors and patterns.
- Determine the key factors contributing to higher customer spending, enabling targeted marketing strategies and product innovation.
Model Performance
- Adopted Mean Absolute Error (MAE) as our primary metric to evaluate the accuracy of the regression model in predicting customer spending.
To facilitate transaction-level analysis, the dataset was filtered for sessions with transaction revenue == 0.
Columns were excluded based on proportions of missing data and use case relevance.
See the code file in the respective labeled folder for more details on preprocessing.
Originally stored as a JSON file, after transformation, these were the columns available in the dataset:
channelGroupingdatefullVisitorIdsessionIdsocialEngagementTypevisitIdvisitNumbervisitStartTimecontinentsubContinentcountryregionmetrocitycityIdnetworkDomainlatitudelongitudenetworkLocationbrowserbrowserVersionbrowserSizeoperatingSystemoperatingSystemVersionisMobilemobileDeviceBrandingmobileDeviceModelmobileInputSelectormobileDeviceInfomobileDeviceMarketingNameflashVersionlanguagescreenColorsscreenResolutiondeviceCategoryvisitshitspageviewsbouncesnewVisitstransactionRevenuecampaignsourcemediumkeywordadwordsClickInfo.criteriaParametersisTrueDirectreferralPathadwordsClickInfo.pageadwordsClickInfo.slotadwordsClickInfo.gclIdadwordsClickInfo.adNetworkTypeadwordsClickInfo.isVideoAdadContentcampaignCode