November 16, 2017
Predictive Intelligence in Ecommerce, Part II
November 16, 2017
Predictive Intelligence in Ecommerce, Part II
In PI we make a computer program learn (or “train” itself) from its experience with respect to some task and a given performance measure. The objective of such learning is to improve the experience in performing these tasks, as measured by performance rules. For example, if the goal is to issue targeted product recommendations to a specified consumer demographic, then with the correct machine learning model it will learn over time to make better matches between the recommendations and the particular consumer segment. More specifically, through a series of attempts, and fine-tuning, it will learn to improve its recommendation precision. The objective of such learning is generalization, meaning the machine learning model accurately applies the patterns from the test and training data to the new incoming data.
There are different machine learning methods: supervised, unsupervised, and semi-supervised learning. Each of these correlates to the best use of a particular machine learning model (e.g., classification, ranking, clustering, segmentation, recommendation system).
When you’re using a supervised learning algorithm, you’re building a classifier that has both a specific input and output; there are definitive correct and incorrect classifications and correlations already known by the person building the model. Your goal is to “teach” the model to accurately classify each piece of data in the data set. A simple example is if you have a set of data where your objective is to teach the model to classify “surf board purchases” vs. “no surf board purchases.” Once the model has reached an accurate level of performance, meaning it’s correctly classified a threshold percentage of the data, then it’s deployed on a larger scale as it has learned the parameters for your specific classification goal.
Unsupervised learning occurs when you have input values, your data, but the goal is for the machine learning model to uncover the underlying patterns present. In this way, the model will determine whether there are associations between your input values. Continuing with the surfboard example, an unsupervised learning model isn’t presented with the correct answer as to “surf board purchase” vs. “no surfboard purchase.” However, the model will provide clustering or association rules that uncover possible correlations between surfboard purchasers also buying other surf board-related items. This can be extended further to delineating possible connections between surf board purchasers and their movie preferences. Over time, and with additional adjustments to the model, unsupervised learners can provide increased accuracy for multifaceted yet still targeted recommendations.
Semi-supervised learning is a merging of both supervised and unsupervised learning methods. Generally, the scenario is that you have a mixture of definitive “answers” (called labelled data), and non-definitive answers (unlabelled data). Labelling data is tedious and time-consuming. This is particularly true if yours is an e-commerce site with millions of users demonstrating difference behaviours within varying time frames. So, the primary function of the semi-supervised model is to learn from the labelled data, and then transfer that learning to the unlabelled data. You’re presenting the model with correct and incorrect answers for a certain portion of the data, then you’re loading the unlabelled data to determine if the model can accurately generalize from what it has learned. If we’re using the surfboard example, you train the model to classify “surfboard purchase” vs “no surfboard purchase” first and based on examples you know to be true. From that learning scenario, ideally the model will accurately induce the same classification from the unlabelled data.
The short answer is yes, you are learning about the behaviours and preferences of your consumers by using PI. Depending on the machine learning model used, you can confirm connections that are already present (supervised learning) or discover new correlations that are hard to parse out due to the sheer size of your dataset (unsupervised or semi-supervised learning). Perhaps there are smaller segments within the surfboard purchasers who like rock music and documentary movies.
Logically, this isn’t directly transferrable to all consumers who purchase surfboards, and one cannot assume that the consumer is purchasing the item for themselves. These are trickier details that machine learning experts are still in the process of resolving. The current widespread belief is that AI will come to the rescue and have the capability to make inferences, create their own heuristics, understand human intent and improve predictive accuracy. However, as previously stated, this is yet to come to fruition.
There are still numerous shortcomings to be overcome in PI as it evolves. The most common ones are:
We are still in the early days of the path leading to true Artificial Intelligence and understanding true human intent. However, retailers could already use numerous practices as described above to improve the quality of their product recommendations.
We’ve established that personalization is fundamental to attracting new customers while retaining ongoing consumer loyalty. We’ve also discussed how PI is vital to boosting the accuracy of personalization. However, you may still be wondering how, specifically, PI can be used to garner a solid return on your investment.
Beyond promotions and marketing campaign management, PI assists in developing real time pricing models through identifying the correct price points for maximizing sales. Given the amount of data that machine learning algorithms can parse through and analyze, PI can incorporate competitor pricing as well as consumer behaviour data to predict the best price point to attract consumers and reveal the likely impact regarding competitive pricing.
Inventory management is yet another area where PI facilitates operational efficiency and a healthy ROI. Retailers are all too familiar with deciding what to do with the inventory that is collecting dust on their shelves. Some items are seasonal, while others might take up space only to be in demand three months later. PI can be used to forecast product demand with greater accuracy. Additionally, PI metrics provide information for current and future product allocation, thus increasing inventory management precision. Given the large volume of data required to make these decisions (real time consumer behaviour analysis, constant changes in competitor pricing, differences in consumer preferences based on regional considerations, etc.), PI is optimal for aggregating, analyzing, and consistently producing spot-on predictions.
[i] Busath, M. (2017). UserVoice increases MQLs by 37% using predictive marketing. Infer.com. Available at: https://www.infer.com/category/use-cases/
[ii] Wettemann, R. (2015). Salesforce marketing cloud: Room & Board. Available at:
[iii] IBM Marketing Cloud. (2016). Outdoor retailer uses segmentation and targeting to increase open rates by 80%. Available at: https://www-01.ibm.com/common/ssi/cgi-bin/ssialias?subtype=AB&infotype=PM&htmlfid=ZZC03314USEN&attachment=ZZC03314USEN.PDF