Product recommendation method

A product recommendation method implemented by an electronic computer provided with a memory configured to store data, said product recommendation method being associated to a search engine and comprising:

• an indexing step wherein a plurality of products of a catalog are initially loaded in a domain searchable by the search engine, said indexing step for each product of the catalog comprising steps of:
• describing a given product among the plurality of products of the catalog according to a first descriptor vector having as components a plurality of product characteristics;
• determining user-characteristics corresponding to the product;
• describing a user corresponding to the product according to a second descriptor vector having as components the user-characteristics corresponding to the product and probabilities associated with a plurality of the user-characteristics;
• inputting the first descriptor vector to a Bayesian network defined by a certain dataset including a plurality of nodes associated with conditional probabilities, and one or more arcs associated with causal relationships and which define a relationship of related nodes among the plurality of nodes;
• calculating the probabilities associated with the plurality of the user-characteristics using the Bayesian network to generate user-characteristics of a most likely suitable user that is most likely to be suitable for the given product;
• storing in the domain searchable by the search engine product characteristic data of the given product in association with the user-characteristics most likely to be suitable for the given product, the product characteristic data including the first descriptor vector in association with the second descriptor vector;
• a recommendation step including steps of:
• obtaining the user-characteristics of the user who is carrying out the search and defining a third vector descriptor including components corresponding to the user-characteristics;
• calculating vector distances between the third descriptor vector and each of the second descriptor vectors;
• determining an ordered list of the second descriptor vectors starting from the second descriptor vector having a minimum vector distance from the third descriptor vector up to the second descriptor vector having a maximum vector distance from the third descriptor vector according to an increasing distance criterion;
• obtaining a corresponding ordered list of first descriptive vectors that is an ordered list of products from a most likely product suitable for the user to a least likely product suitable for the user;
• and
• displaying the ordered list of products to the user,
• wherein the Bayesian network is trained and refined over a plurality of cycle of iterations using a refining process that identifies determinants of the certain dataset and the conditional probabilities associated with the nodes, and modifying for each cycle of the iterations one or both of the causal relationships to modify the nodes and the conditional probabilities to modify the arcs, thereby minimizing errors at the nodes to which are pointed to by the arcs associated with the causal relationships,
• wherein the refining process comprises:
• storing historical data relating to cases of matching between product-characteristics and most likely suitable user-characteristics, wherein said matching has had a positive outcome measured according to one or more predetermined indicators;
• executing an automatic self-learning algorithm that continuously changes a topology of the Bayesian network to reduce errors and increase an accuracy of the Bayesian network by modifying the conditional probabilities or at least part of the topology and conditional probabilities of the Bayesian network on the basis of the stored historical data;
• performing a test estimate on known cases using the previously created Bayesian network by inputting the first descriptor vector of the products of the known cases in the Bayesian network and determining the second probably suitable user descriptor vector by means of the Bayesian network;
• comparing the test estimate obtained from the Bayesian network with the real data of the known cases to determine whether the test estimate corresponds to the real data of the known cases within a predetermined tolerance interval;
• validating the Bayesian network in response to the test estimate obtained from the Bayesian network corresponding to the real data within the predetermined tolerance interval;
• in response to the test estimate not corresponding to the real data within the predetermined tolerance interval, cyclically repeating refining steps, until the test estimate obtained from the Bayesian network corresponds to the real data within the predetermined tolerance interval, the refining steps comprising:
• executing the automatic self-learning algorithm, wherein the automatic self-learning algorithm modifies, at each cycle among the plurality of cycles, at least part of the topology by changing one or more of the causal relationships between the related nodes, and modifies the conditional probabilities of the Bayesian network on the basis of the historical data stored and the difference between the estimate and real data calculated in the previous cycle;
• performing the test estimate;
• comparing the test estimate obtained from the Bayesian network with the real data of the known cases;
• and
• continuously modifying the at least part of the topology of the Bayesian network and the associated conditional probabilities until the test estimate corresponds to the real data within the predetermined tolerance interval.