Most of us can identify with this particular case, when Netflix knows which movies we would like to watch next and we spend the next 3 hours watching Netflix movies.

Or a case where we spend too much money on online shopping, even though we only wanted to buy one thing. Or a time when we start listening to songs on YouTube, and the playlist leads us to a really cool band we like right away. So it begs the question - how have these machines become so smart that they magically know what we want? The answer is - all the magic that happens behind the scenes is done by machine learning, or more specifically, by recommender systems that use algorithms to find similar items and similar customers, based on their behavior, and recommend items that that particular customer should like.

Recommendation systems are a very popular and effective paradigm in retail businesses. With a recommendation system, shoppers can find items they like with less effort. In addition, they are presented with items that they had never thought about buying, but that really fit their needs.

Customer behavior analysis focuses on understanding the type of customers; what they like, what they don't like, what is their pattern of interaction with products, the value of customers, etc.\ If we are able to model these aspects of a customer, we can predict their future needs.

The main benefits of a customer behaviour analysis system are:

1. Enhancing sales,
2. Better understanding of the customer; and
3. Enhancing customer loyalty.

The main benefit can be expressed in three words - increased sales. According to research, 35% of all Amazon purchases and 70% of Netflix purchases are driven by their recommendation systems, and starting to use recommendations significantly boosted their sales. (Figure 1) In addition, during the COVID-19 pandemic, many retailers went online, digitized their businesses and changed their business culture to adapt to new and ever-changing conditions. According to reports, in 2020, e-commerce sales growth in the US alone exceeded 30% (Figure 1). This provides a huge amount of online data for potential exploration and use in the construction of machine learning systems.


Figure 1: Comparison of US e-commerce sales growth in 2019-2020

The second benefit comes from a better understanding of customers. This is where the customer profile comes in. By profiling customers, we can better understand their behaviour and, consequently, better understand their needs or, in other words, meet their needs, which can ultimately be rewarded with greater customer satisfaction and loyalty. In addition to increased customer satisfaction, we can easily create automated marketing campaigns and personalize them based on customer analysis.

The next advantage is a much better strategy for long-tail items. The term long tail item, refers to niche and hard to find items that are very specific and unique and usually only have a small group of people looking for them. From the customer's perspective, tools such as recommendation systems allow them to find products outside their immediate area and items that they would otherwise not have access to. From a supplier's perspective, if they keep items in a warehouse, hidden from customers who would want them, this strategy could become very profitable. (Figure 1)

To achieve these results, we built the infrastructure and pipeline for data analysis and modeling with machine learning algorithms. Briefly, the pipeline consists of an input module that connects to a data source and sends the data to a data analysis and human behavior modeling module.

In this section, data transformations, such as cleaning and preprocessing, are performed, and data segmentation and recommendation models are constructed. The modeling results are sent to the output, which is presented as multiple dashboards in the control panel. The high-level pipeline is shown in Figure 2.


Figure 2: High-level pipeline of the customer behaviour analysis platform

Customer profiling is done by segmenting customers into groups that exhibit similar behavior based on different parameters, which are derived from data such as number of items purchased, value of items purchased, number of items returned, types of items purchased, etc. Segmentation is essentially clustering based on behavioural similarity.

The input to segmentation is the parameters for which we would like to find similar groups. The result is the number of similar groups, and the rules under which a particular property falls into each group. In addition, we can assign values to each segmentation per segmentation and use them as the numerical weight (score) of the segment. Different segmentations can be combined as equals, or according to assigned importance weights, to provide value to the customer. More segmentations per attribute will give us a more detailed profile of the attribute, since we will have different views of the attribute and, consequently, more insight into the attribute (customer, type, brand, store, etc.).

Therefore, the customer profile is the most important parameter for further recommendations.

An example of a segmentation flow is shown in Figure 3. There we see customers by parameters with a number of purchased items (x-axis) and a profit margin from the purchases (y-axis).
Customers can be divided into 3 distinct groups. We have three segments for which we assign importance and score:\

1. High profit margin & average number of items purchased - The best - rating is 1.00
2. Low profit margin & higher number of items purchased - Medium - the score is 0.66
3. Low profit margin & low number of items purchased - Low - score 0.33

Figure 3: Example of customer segmentation

There are various methods for how to implement recommendation systems and, in this case, we used a hybrid model:

• Cooperative filtering model
• Content-based model

Cooperative filtering is an approach that uses the assumption that users who bought similar items in the past will agree on new items.\
Let's look at an example case with 2 customers - Jack and Jill (Figure 4). If Jill bought items A and B, and Jack bought items A, B and C, it means that if Jack and Jill had already agreed on 2 items, there is a high probability that Jill will also like item C. So, according to the collaborative filtering approach, we will recommend item C to Jill next.

On the other hand, cooperative filtering has some known problems, one of the main ones being the cold start problem. When a new element appears, it has no interaction. This means that it would never appear in the context of recommendations.

Figure 4: Example of collaborative filtering

Another common approach, which mitigates the weaknesses of collaborative filtering, is the content-based model. The content-based model works on the assumption that what the customer liked/bought in the past, they will most likely like/buy in the future.

Uses meta-information of the items and a profile of the user's preferred choices. Let's consider the example of Jelena, who often buys her clothes online (Figure 5). In the last few months, Jelena has bought several items online.

First, she bought a pink skirt, then a few days later a pink top, then pink heels and then a pink hat. It is obvious that Jelena likes pink clothes, a common characteristic shared by all the items. It is quite possible that Jelena likes a pink dress more than a black or a blue one, or even an object that is not a garment. So, following the context-based approach, we will then recommend a pink dress to Jelena. On the other hand, the content-based model also has the problem of a cold start. When a new user appears, she has no previous purchases.


Figure 5: Example of a content-based model

• We created a prototype recommendation system using a hybrid approach and adjusted the parameters that yield the best results.
• We built it step by step, feeding the model with more features at each step, and tested the results.
• We split the data into training and testing packages.
• The training data was used to build a model.
• We then tested the performance of the model on test data.

The Mobiplus Shopping Recommendation Platform is part of Software as a Service (SaaS) services.
All the components needed to make it work are located online, and more specifically all stored in a cloud.

• So, when you start your subscription, you will receive your passwords directly and can start using it.
• More specifically, you will need to download on the main system of your e-shop, a specific software through which the data collection takes place.
• Then, through a technical interface (API), you will have connected your e-shop to the Mobiplus Shopping Recommendation Platform.
• At the same time, the Presentation Software will be activated, which takes care of sending your recommendations to new and existing customers.
• The goal is to attract them to either your e-shop or your physical store!
• Finally, you will have access to detailed statistics, which you can use to optimize your marketing actions!

Contact us by filling out the contact form for a consultant to come and meet with you to show you what your business will look like, learn how other relevant businesses are already benefiting from the platform and see the immediate results it can have for you!