Mathficast Software Services Limited

Mathfi-E-Commerce ©
Classifier/Predictive AI

A new breed of next generation AI-powered service helping e-commerce sites or experts to make decisions by performing binary or multi-class predictions and classifications or sanity checking their existing predictions/classifications for example:

1. Predicting future hotel booking cancellations: Butterfly AI can help travel Data Scientists or their AI-powered travel platforms to improve the accuracy of their hotel booking cancellations predictions. Provided that the relevant data features of thousands of past hotel bookings are available including the website customer used for booking (e.g. Expedia, Booking.com, Hotel's own website), Number of simultaneous bookings customers made across different hotels (if available), lead days (number of days between booking and check-in) more than 20 days (Yes or No), lead days between 10 and 20 days (Yes or No), lead days within 10 days (Yes or No), last minute booking (Yes or No), nationality of the guest, the country of residence of the guest, whether or not guest responded to hotel's personalised contact (Yes or No), response to pre-stay email (Yes or No), response to pre-stay survey (Yes or No), arrival month, arrival week, arrival day, stays in weekend nights (Yes or No), past history of staying in that hotel, past history of staying in that chain of hotels, number of adults, number of children, number of babies, whether or not the room details was viewed by that person before booking, whether or not, the room that guest viewed first was the one they booked later, type of meal ordered FB, BB, HB, etc., market segment for example direct, online, corporate, etc., distribution channel for example direct, corporate, etc., travel agency booking (Yes or No), number of previous cancellations, assigned room type within that hotel, reserved room type during most recent stay in that hotel, whether or not current booked room is the same type as the room they had during their previous stay in that hotel (Yes or No), number of times they changed the booking details, deposit type, company booking (Yes or No), number of days in the waiting list, customer type (for example transient, contract, etc.), average daily rate, required car parking spaces, have they booked any tours through hotel website, total number of special requests, and having verified and labelled all those past bookings as (Cancelled or Not Cancelled, the targets of labels of predictions), predict whether a hotel booking is going to be cancelled in future or not (and with what probability and certainty).

2. Fraud detection at self-checkouts in retail and supermarkets: To enable retailers and supermarkets identify cases of fraud within the self-checkout process and minimise revenue loss and disruption to innocent customers, Butterfly AI can help follow-up checks to detect the fraudulent activity seamlessly. Knowing data features of many past self-checkout shopping journeys including customer’s historic individual trust level. 6: Highest trustworthiness (through credit and shopping history, etc.), total time in seconds between the first and last product scanned, grand total of products scanned, number of voided scans, number of attempts to activate the scanner without actually scanning anything, number of modified quantities for one of the scanned products, average number of scanned products per second, average total value of scanned products per second, average number of item voids per total number of all scanned and not cancelled products and any other potential relevant data features including weight, and having all those past journeys correctly labelled as fraud (1) or not fraud (0), decide/classify whether a new self-checkout journey (knowing all the above features for current customer) is fraudulent or not (and with what probability).

3. Prediction of online shoppers’ purchasing intention: Imagine we have collected key behavioural data features of thousands of past customer online journeys across an e-commerce shopping site through Google Analytics and click-stream. Knowing the statistics of each page during the session, including the type of pages visited on that site, the number of time visited home page, and the time spent on the home page, the number of times visited product list page and the time spent on that page, the number of times visited the product details page and the time spent on that page, etc., the number of search attempts, the average value for number of times that a user visited the page before completing an e-commerce transaction (i.e. the page value), the percentage of visitors who enter the site from that page and then leave without triggering any other requests to the analytics server during that session (i.e. the bounce value), the percentage of the page views on any specific page that were the last in the session (i.e. exit rate), the closeness of the site visiting time to a specific special day (e.g. Christmas, Valentine's Day) in which the sessions are more likely end up with transaction, other characteristics such as channel type (mobile, desktop, etc.), the type of the browser, operation system, region, country, customer’s past transaction history, and also knowing whether that past session led to a purchase by the customer or not (0: Not Purchased, 1: Purchased, the targets and labels of prediction), for a current live session in real time predict whether it will end up in a purchase (and with what probability) and if not perhaps offer some incentives to encourage customer to purchase the product.

4. Predict if customer is going to click on an online product advertisement on a webpage or social media site: Butterfly AI can help marketers or marketing platform determine whether a customer is going to click on a product page or banner and to determine the probability of the click. Knowing the key features of thousands of customers’ recent online journeys including product ID, web page ID on which the product has been displayed, social media page ID, whether the product banner displayed on a web page or social media site that customer frequently visits (yes or no), the frequency of visit to that web page by customer, the category of that product, has the customer clicked on similar products from that category of product in past (yes or no) (similar products may be extracted from a knowledge graph), the number of products customer purchased from that category of product within past 6 months, the durability of the product, gender, age, income, employment category, city of location’s development index, customer activity depth score with that website or social media site, product brand or manufacturer, the number of same brand products purchased by customer in past 6 months, is current product on offer (e.g. offers a discount), whether the product is a seasonal product and we are currently in that season (e.g. Christmas, summer beach season), whether the most similar products purchased by customer were on offer (yes or no), search history of customer (frequency of search for that product or its features) within last week, the whereabouts of customer on website right now, the product banner format similarity score (including colours used, frame size, text used, et.) of the past similar products clicked and purchased by customer to the current target product advertisement banner, whether the top ten most similar customers to the target customer in terms on online behaviour (similarity for example in terms of age, gender, demographics, brand of products purchased in past. etc.) have clicked on that product advertisement banner during past couple of days (the similar customers for example may be determined from a knowledge graph), the frequency of visit to that web page by the top ten most similar customers, whether the top ten most similar customers to the target customer in terms on online behaviour have searched for that category of product during past couple of days. whether the product is an accessory to another product that customer purchased just now, and any other relevant data features (including potentially some of the data features described in previous section) and having recorded, verified and labelled whether during those past journeys customers clicked on the product page or not (the label and target of the prediction), predict whether a target customer during a new online journey will click on the target product advertisement or page (and provide the probability of click), provided that those features (listed above) are available for the current targeted customer

5. Package billing classification to help e-commerce companies reduce delivery costs: During past years the carriers have realised that cost models based on actual package weight don’t make sense for lightweight packages that take up cargo space and reduce the carrier’s shipment capacity, therefore they introduced billable weight. Knowing the package data features such as billable weight type, billable weight, weight, length, width, height, volume, dimensional weight, dimensional factor, and other relevant data features and knowing which price zone (The label and target of classification) those past packages ended up in, classify a new package bill into the correct pricing zone (and with what probability).

6. Classify the customers according to the most successful marketing channel they will highly likely respond to: Considering the customers’ data features and behaviour during thousands of past marketing campaigns including social media activities, habits, likes and presence, posts, shopping history, income, credit rating, interests, location, house value, car, social networking profile, mobile type, what type of marketing channel the most similar customers to that specific customer have been engaged to most, during past couple of days (similarity for example in terms of age, gender, demographics, brand of products purchased in past, the similar customers may be determined from a knowledge graph) and any other relevant data features (potentially some of the data features described in the previous two sections) and having recorded, established and labelled which one of the marketing channels, A (Direct Contact), B (Email Marketing), C (LinkedIn Approach), D (Facebook targeting) the customers during those past marketing campaigns have responded to, predict and classify an existing or new customer into the marketing channel A (Direct Contact), B (Email Marketing), C (LinkedIn Approach), D (Facebook targeting) segment (and also present the probability of success) during an upcoming marketing campaign.

7. Product Recommendations: Knowing the past shopper demographics, age, gender, location, job, income, social media preferences, the specific season and time of year, the brand they like most, their price range, the frequency they purchase the products, the channel they usually use for hopping, the engagement with past recommendations, and other relative data features and knowing what type of products they have purchased in past, for a new potential shopper with known data features predict and recommend the type of product they will be interested in most likely and with what probability.

8. Predict if customer will add certain product to an online basket: Knowing the past customers’ data features such as pages visited, search history, click history, basket add and purchase history, preferences, product engagement history, and whether those customers add a certain product to basket, predict/decide whether the new or existing target customer would also add that product to basket or not.

9. Predict a customer or subscriber churn on an E-commerce website or service: Knowing the past subscribers’ details such as pages visited, search history, click history, basket add history, advertisement engagement history, etc, and whether they have churned from website or not, predict/decide whether a new existing or new subscriber will churn or not

and any other binary or Multi-Class E-commerce classification or prediction use cases.