The techniques of cross-selling and upselling have been used in various industries for a long time. Various studies show that when these techniques are implemented correctly, they can reach conversion rates of up to 70%. Generally, multichannel techniques are used, such as offering products when the customer calls the call centre for support (passive offer), SMS directed to clients, calls from our call centres to our clients (active) and/or offers sent directly via traditional mail or email. The main driver in deciding one type of channel or another in the majority of these campaigns is usually cost; for example, sending thousands of SMS is very cheap, however, on many occasions the effectiveness of the campaign if not taken into account, when due to the excessive number of messages sent to potential clients, these potential clients ignore all type of contact.

It is important to be able to choose the target group correctly, those who are the most likely to buy the products on offer. Here is where the new automatic learning techniques and technologies (machine learning), like recommendation systems, frequently used in cross-selling to find profiles similar to our clients and to find new products which might interest them.

Current Machine Learning: Bayesian Networks and Classifiers

These two techniques have been used with a good degree of success over the past 10 years. The Bayesian networks are founded on the construction of an ontology (or business knowledge) in the form of a network that has been refined in order to make recommendations for clients. The commonly used classifiers are Support Vector Machine, naïve Bayesian classifiers, and decision trees. However, all these models are currently considered outdated since they do not taken into account much information hidden within the data and the technique with the best results these days is combined collaborative and content-based filtering.

machine learning classifiers

New Machine Learning Techniques: Collaborative filtering and content-based filtering

Collaborative filtering is based on identifying a group of clients with similar tastes. For each of our clients, a method of resolving collaborative filtering is to use a matricide factorisation from a matrix where we have clients (X) and the products they have bought (Y). The two matrices resulting from the factorisation usually keeps the most relevant information about our clients’ preferences but on a smaller scale. Finally these two matrices are multiplied by each other to generate a new matrix whereby spaces that were previously empty now give us a value that suggests the preference or propensity towards similar products for a group of users with similar consumer preferences.

Collaborative content-based filtering is the recommendation technique that offers the best results #machinelearning Click To Tweet

Another technique is content-based filtering, where the main idea is to recommend products similar to those purchased previously by the client or which have been given a high rating. This seems quite obvious, but when our product catalogue is huge, it is not an easy task. Firstly, a profile must be generated for each product; this profile can be represented in the shape of a vector for example, where all the different characteristics are defined as a Boolean value, integer or real. All our product “vectors” or profiles must have all the characteristics (columns) there are in our catalogue.

Once the product profiles have been created, the profile of each client must be calculated by obtaining an average of the product profiles purchased or valued positively by our client. Finally, the client profile may be compared with all our product profiles to identify recommendations.

All these techniques for recommendation systems with cross-selling and upselling can be easily and efficiently set up in the public cloud, as we are able to adjust demand for GPU or CPU depending on our batch processes for calculation of recommendations.

como funciona el filtrado colaborativo

Author

  • Software company specialized in the design, construction and operation of digital data products based on cloud computing platforms.