Off late, one of the market segments where data scientists have been in huge demand has been e-commerce. By each day, big companies are becoming bigger and innovative tech startups are cropping up. And each of these firms wants to expand its utilization of machine learning potential to build unparalleled competitiveness in the market. To understand this evolution better, it is imperative to familiarize oneself with all the potential applications of machine learning in the e-commerce business spectrum. In this article, I’ve focused on highlighting that. This list should be of high utility for stakeholders in the e-commerce ecosystem as well as aspiring data scientists planning to find career options in this field.
Machine learning has empowered retailers to analyze all queries, whether searched or abandoned, from all the users. By analyzing past click-through behavior, purchases, preferences and history in real-time, targeted predictive analytics results can increase sale probability. Targeted problems include ranking query (search results), query expansion (QE), related queries (relevant content based on current search) and de-douping (optimizing data storage by removing duplicates).
Merchandising, Recommendation and Promotions
Predictive analytics combines expert knowledge and mined knowledge from consumer behavior, including the purchase history of that consumer and the performance of different products on the site, to determine relevant recommendations for the consumer with higher probability of sale conversion. The intent is to convert browsing-into-buying and increase cross-selling. Common recommendation system methods include attribute-based, item-to-item correlation and user-to-user correlation. Techniques such as algorithmic and context merchandising combines the elements of active consumer behavior (personalization), passive consumer behavior (accounting for user response and repetitiveness) and inventory performances (optimization for outperforming products, in stock products, seasonal products, etc.) to create an effective merchandising system.
Image recognition and understanding
E-commerce has several unique opportunity areas regarding image recognition using machine learning. For example, startups are building tools using using which customers can click a product image and then the software will identify similar products for sale available online (from product catalogs). This visual search technology in e-commerce is also appended by the historical search data of the consumer in identifying the visual match.
In the e-commerce context, concept extraction is used to secure the key content of web content page. This information is then used for applications such as search relevance (opinion mining) and product matching. Common methods used in e-commerce include the automatic concept extractor (ACE) and automatic keyphrase extraction (KEA). Research is underway to expand the abilities of these methods to perform similar analysis on part-of-speech patterns (voice) as well.
Sentiment and trend analysis
Sentiment analysis is employed to make sense of user reviews. Since user reviews on products is multi-dimension on aspects (quality, price, service, etc.), its analysis system mandates thrust on multi-dimensional sentimental analysis with lexicon expanding mechanism. Employing techniques such as natural language processing (NLP), online retailers can analyze consumer sentiments and apply inferences drawn in their planning and forecasting processes.
Supply chain management and demand forecasting
In e-commerce, logistics optimization is required for warehouse space optimization, cash payable management, grouping orders, vendor selection, optimizing route path, planning and forecasting, sourcing optimization, fulfillment, delivery and returns. With e-commerce product demand often suffering from bullwhip effect, machine learning techniques such as recurrent neural networks and SVMs are gradually finding ground in supply chain demand forecasting. Common algorithms deployed for forecasting include the likes of Nearest Neighbor, Lazy Learning and Recnoisy.
Fraud detection and prevention
Fraudulent behavior has been observed in e-commerce in payments (card not present), fraudulent accounts, fraud product listings, malvertizing, transaction chargebacks, astroturfing (fake ratings) and account takeovers. To correctly predict instances of fraud, the system needs to analyze massive data sets, including online presence, location and device-level data of each customer and merchant. Machine learning based predictive analytics can help predict such frauds. Most commonly, time series anomaly detection algorithms are employed to detect frauds. Inclusion of larger set of ML techniques in fraud management is also being initiated in e-commerce under the umbrella of Risk Management.
Predictive modelling can address the traditional issues of non-standardization, duplication and incorrect values in entity description that result in record linkage issues. Entity resolution is an effective tool to improve the efficiency of online purchasing. Al-based algorithms, ranging from simple attributions and value-tagging to employing entity resolution frameworks are deployed to address the problem. Commonly used frameworks include MARLIN, FEBRL and STEM.
Factors such as ever increasing size of data for classification and constraints in real-world expectations on response time has made classification a challenging task in the e-commerce realm. Classification algorithms such as Naive Bayes, K-Nearest-Neighborhood and Maximum Entropy are among the most commonly used. Some research work is also focusing on employing Dense Subgraphs to scale the capabilities of the classifiers to even larger data.
Predictive analytics analyses pricing trends in correlation with sales trends to set the price of the product to maximize revenue and profit. This is achieved by analyzing historical data for products sales, minimum allocated price, customer data, product competition, inventory, time of the day, etc.. To achieve this, most larger e-commerce firms employ dynamic pricing algorithms that emphasize of fundamental concepts such as yield management, competitive pricing or even algorithmic trading.
Most e-commerce firms have realized the negative impact of barraging the customer with excessive levels of notifications and emails. To keep the customer interested and not irritated, it is important that targeted marketing on each customer should be limited to products which she might be interested in buying. Predictive models are employed to identify products that have a higher likelihood of being purchased by the customer (propensity modelling). Propensity score matching (PSM) is the most common machine learning technique used.
Customer Churn Prediction
A customer who abandons an e-commerce portal or uninstalls its mobile app is a costly loss to the business. Gaining a new customer is umpteen times more expensive than retaining older ones. Therefore, customer churn prediction is vital both for churn prevention (individual or group level) and also to assess long-term effectiveness of marketing strategies. Churn prediction algorithms are used in other industries as well, including telecom. In the e-commerce business, well-known churn prediction techniques include C4.5, RIPPER and ANFIS (neuro fuzzy classifiers).
Search Engine Marketing
It is a fact now that Google regularly keeps updating it machine learning based search algorithms. For better search engine optimization (SEO) marketing, it is vital that Google’s algorithm behavior be decoded to the extent possible. For accomplishing this, e-commerce firms need to employ predictive modelling, analyzing search engine results, to measure and plan the efficacy of their search engine marketing campaigns.
What is anticipatory shopping? Amazon recently filed a related patent. It essentially means your product is being packaged even before you buy it – in anticipation of it being shopped. Benefit? In the event of a purchase, your product will be delivered faster to you because it is already en-route. The idea can certainly win laurels from happy customers. But for it to succeed and not cost unwanted wastage of resources, it is important that the predictions made by the algorithms (based on personalized data of customers) is extremely accurate. Anticipatory machine learning algorithms are employed for this purpose.
All this can change when we move into the era of connected devices and virtual personal assistants. Personal assistants might execute a lot of decision-making for us before an e-commerce player gets to interact with us. Personal assistants, integrated with robotics, might also be able to detect and record emotional signals via voice or facial expressions. Companies that control the personal assistants will become the marketer’s central focus. Also, marketers will have to find and innovate ways to lure the personal assistants. Also, connected devices will considerably alter the landscape of customer relationship management (CRM) and its decision-making will reduce customer’s attention on individual devices. Marketers will then have to find ways to breach into this auto communication mechanism to pitch their new products.