FR3017734A1 - PRODUCT RECOMMENDATION OF A CATALOG FROM A SELECTED PRODUCT AND A SEARCH ENGINE - Google Patents

Description

The present invention relates to a method and device for recommending products from a catalog of products presented and available for consultation on a communication network. internet type. The products can be of any type, for example mail-order products, multimedia products for download or viewing, etc. The current recommendation systems for products from a selected product are of the type "those who liked the selected product also liked these other products ". These referral systems are based on the existence of a database that lists the purchases or product consultations that have been previously made by multiple users. It is therefore necessary for this type of recommendation to build a history database of behaviors or purchases before making a recommendation. This database is not always present, especially when the product catalog is new and no purchase or consultation has yet been made on this product catalog. Other systems are based on similarities between products to make recommendations. A product similar to the selected product, by comparison of certain characteristics, can then be recommended. For that, it is necessary to know the characteristics of the products of the catalog to be able to make the comparisons. However, it is common that the characteristics or metadata of products are not known or are insufficient to make this type of recommendation.

Thus, there is a need for a product recommendation system of a catalog that functions even when there are no databases of online purchase history or online consultation of these products or feature data. these products. The present invention improves the situation.

To this end, it proposes a method for recommending products contained in a catalog of products that can be viewed on a communication network from a first product selected. The method is such that it comprises the following steps: determining a plurality search queries from the first selected product and a set of second catalog products and sending this query to at least one search engine; obtaining the number of results obtained from the search of the at least one search engine and calculating, for the set of second products in the catalog, a similarity coefficient between the first and the second product; - recommending at least one second product in the catalog based on the calculated similarity coefficients. Thus, it is possible to make a recommendation of products from the beginning of the online catalog. It is not necessary with this method to have historical data or even to have product characteristic data. In fact, the search result performed on the queries relating to the selected product makes it possible, as soon as this result is obtained, to make immediate recommendations to be made. This method of recommendation is moreover simple to implement and is not expensive in memory space.

The various particular embodiments mentioned below may be added independently or in combination with each other, to the steps of the recommendation method defined above. In a particular embodiment, the search queries comprise a search request for web pages comprising the first product and a search request for web pages comprising a second product or a search query comprising both the first product and a second product. product. These types of basic queries allow results to be obtained even if the search engines solicited can not take into account "AND" and "OR" operators (and perform implicit "OR" or "AND" when a query has several keywords) or can only take into account operators of type "AND". The results of these queries make it possible to obtain a similarity coefficient, at least by estimation. In the case where the search engines take into account the operator "OR" search queries further include a search request for web pages with either the first product or the second product.

The results of these queries make it possible to obtain a similarity coefficient, by a more precise estimate. In a particular embodiment, the similarity coefficient is an estimate of a jacquard coefficient defined as the ratio between the number of pages comprising both the first and the second product over the number of pages comprising either the first product, the second. The jacquard coefficient is particularly suitable for determining a similarity between products. It gives indeed excellent performance of recommendation by similarities. In the case where the search engines solicited do not allow to have the direct calculation elements of this coefficient, an estimate of this coefficient is possible from the responses of search queries. In another embodiment, the search queries include a web page search request comprising either the first product or a second product and a web page search request including both the first product and the second product. These types of queries are then adapted to search engines that take into account both the "AND" and "OR" operators.

With such types of queries, it is then possible to directly calculate a jacquard coefficient defined as the ratio between the number of pages comprising both the first and the second product on the number of pages comprising either the first product or the first product. second. The coefficient of similarity is then more reliable.

In another embodiment, the similarity coefficient between the first product and a second product takes into account a co-occurrence number of the first and second product on the resulting web pages of the search. Thus, the similarity coefficient and the associated recommendation is based on the coexistence of two products on a large number of web pages or on a large number of coexistence on the same web page. This type of criterion is sufficient to make an interesting recommendation. The present invention also provides a device for recommending products contained in a catalog from a first selected product. The device is such that it comprises: means for determining a plurality of search requests from the first selected product and a set of second catalog products and sending these requests to at least one engine of research; means for obtaining the number of results obtained from the search of the at least one search engine and calculation means, for the set of second products of the catalog, of a coefficient of similarity between the first and the second product; ; means for recommending at least one second product of the catalog according to the calculated similarity coefficients. This device has the same advantages as the method described above, which it implements.

The present invention also relates to a recommendation server comprising a recommendation device as described. It relates to a computer program comprising code instructions for implementing the steps of the recommendation method as described, when these instructions are executed by a processor.

Finally, the invention relates to a storage medium, readable by a processor, integrated or not to the recommendation device, possibly removable, storing a computer program implementing a recommendation method as described above.

Other features and advantages of the invention will emerge more clearly on reading the following description, given solely by way of nonlimiting example, and with reference to the appended drawings, in which: FIG. 1 illustrates a communication system comprising a recommendation server according to one embodiment of the invention; FIG. 2 illustrates in flowchart form the steps of a recommendation method according to an embodiment of the invention; and FIG. 3 illustrates a hardware representation of a recommendation device according to one embodiment of the invention.

FIG. 1 thus illustrates a possible context for implementing the invention. In this context, an internet network type communication network 100 enables different entities to communicate. These entities are, for example, a user 105 communicating via a communication terminal, a product server 106 offering different products contained in a product catalog 107, a recommendation server 101 implementing a recommendation method of products as described later with reference to Figure 2, search servers 102 and 103 which make it possible to search the web pages on the communication network according to predefined requests. Thus, in this context, the user who has selected a first product in the product catalog 107 of a product server 106, receives following this selection a recommendation for one or more other products in the catalog. For this, the recommendation server determines from the first selected product and a set of second potentially recommendable products contained in the catalog 107, search queries that it sends to one or more search engines Ml, M2. In another embodiment, these search engines may form an integral part of the recommendation server 104. With the search results obtained from the search engine or search engines, the recommendation server obtains the number of these results and calculates a search coefficient. similarity for each of the second products of the set of second products for which search queries have been determined. The similarity coefficients obtained are then compared and the product or products associated with the greatest similarity coefficients are proposed to the user 105 as recommended products. This process can be used to make the recommendation on a product list without having specific knowledge of the product characteristics and without knowing any purchase or consulting history associated with these products. Figure 2 now illustrates the main steps of a recommendation method according to one embodiment of the invention. In step E202, a first product Pj of a catalog (Cg) or list of products that can be consulted for example on the web is selected by a user. This may apply to any online shopping site for various physical products but also downloads products such as audio, video or electronic documents. This can also apply to products to be viewed as images, photos, messages. These products can be consulted on the web but also on other networks such as social networks. In step E202, a plurality of search queries are determined from the first selected product (PJ) and a second product (P,) from a set of second products of the catalog (Cg). The set of second products can be the complete catalog without the first product selected or a subset of the catalog. This set comprises in the embodiment described here, a number N of second products (Pi) potentially recommendable. These search queries, as defined later, are sent to one or more search engines. The results are then obtained in step E203 and a calculation of a similarity coefficient between the first and the second product is performed. At this stage, the system can retrieve the total estimated result number of each request if the search engine provides this data in its programming interface (API for "Application Programming Interface"). The system can otherwise estimate this number of results, by rapid traverse of the results pages. In a possible embodiment (in background processing) the system can also post-analyze the results of queries to retain only those that are deemed relevant (for example using semantic analysis techniques known to the state art). Post-processing can be performed on query results by a simple sequential scan, by statistical paths, or by semantic analysis. Once the number of results for each determined request, several methods of calculating the coefficient of similarity are possible, as described later.

A similarity coefficient is thus calculated and stored for each of the second products of all the second products. Depending on the search engines solicited, search operators "AND" and "OR" are not necessarily taken into account. In this case, the requests are adapted to the search engines available on the network or in the recommendation server and are determined automatically by the server. In one possible embodiment, a list of search engines and their associated properties is defined at the time of installation of the system. An update of this list can also be done dynamically.

In this list, it is indicated for each searchable search engine, whether or not it supports AND and OR requests explicitly via operators. Thus, several situations can occur. In the case where the search operators "AND" and "OR" are not taken into account by the search engines, the search requests include a search request for web pages referencing the first product P, and a search query web pages referencing the second product P. The requests are thus automatically defined according to the properties of the search engines available on the network.

The results of these investigations are then counted in step E203 taking into account the following inequalities: max (I / 3 I, I) <113 OR Pi '<1/31 + 1P11 (1) min (IP31.1Pil) 113 AND Pill131 + - 113 OR Pi '(2) with IP I, the number of web pages containing P.

Indeed, the maximum between the number of pages containing at least P, and the number of pages containing at least P, is less than or equal to the number of pages containing either P ,, or P, which is also less than or equal to the number of pages containing at least P, plus the number of pages containing at least P. Similarly, the minimum between the number of pages containing at least P, and the number of pages containing at least P, is greater than or equal to the number of pages containing both P ,, and P, which is also greater than or equal to the number of pages containing at least P, plus the number of pages containing at least P, minus the number of pages containing either P ,, or PJ. These inequalities make it possible to calculate a similarity coefficient by estimating a jacquard coefficient defined as the ratio between the number of pages comprising both the first product and the second product and the number of pages comprising either the first product or the first product. second product. The corresponding equation is as follows: J (P1Pi) jET Pi y 1Pi OR PL I (3) Indeed, from the inequalities defined above, an interval of possible values of the jacquard coefficient can be defined between the following values : o and max (IP11,1p11) (4) The value 0 being in case of complete disjunction of sets Pi and P. The determination of a median value of this interval can thus represent an estimate of the coefficient of jacquard 1 (P; , Pi) and thus define the similarity coefficient.

In the case where the "AND" operator is taken into account by the search engine or search engines, a search request for web pages comprising both the first product (P) and the second product (P,) replaces the request. search of web pages comprising the second product (P,).

Thus, from the inequalities defined in (1) and (2), an interval of possible values of the jacquard coefficient can be defined between the following values: VI and PiyetP ~ and Piy (5) 1131+ IPiI max (IPJI, La determination of a median value of this interval can thus represent an estimate of the jacquard coefficient J (PJ, Pi) and thus define the coefficient of similarity.

The estimate here is more accurate than in the first case. In the case where the "OR" operator is taken into account (and not the "AND" operator) by the search engine or search engines, a search request for web pages comprising either the first product (P,) or the second product (P,) can be associated with the requests as defined above.

With the results of this additional search, it is then possible to refine the estimate of the jacquard coefficient as defined above in equation (3). We obtain here, from the inequalities defined in (1) and (2), a more precise range of possible values of the jacquard coefficient, defined between the following values: 1/31 + - IP] OR PI / minaPj I, and (6) / IP] OUM, '/ IP] OUM. Here again, for example by taking the median value of this interval, an estimate of the jacquard coefficient J (pi, Pi) thus defining the similarity coefficient. The estimate here is more accurate than in the first case. Finally, in the case where search operators "AND" and "OR" are taken into account by the search engines, then it is possible to directly obtain the calculation of the jacquard coefficient as defined by equation (3). ) above. Indeed, in this case, search queries include both a search request for web pages comprising either the first product (P) or the second product (P,) and a search request for web pages comprising both the first product (PJ) and the second product The similarity coefficient J (PJ, Pi) is then obtained directly by the calculation of the jacquard coefficient according to equation (3) above, from the results obtained of the request. The jacquard coefficient gives indeed excellent performance of recommendation by similarities. These performances are for example mentioned in the following references: Laurent Candillier, Frank Meyer, Françoise Fessant: "Designing Specific Weighted Similarities Measures to Improve Collaborative Filtering Systems" in ICDM 2008: 242-255 Frank Meyer: "Recommending systems in industrial contexts" in CoRR abs / 1203.4487 (2012). In another embodiment, it is possible to obtain a similarity coefficient different from the calculation or estimate of the Jacquard coefficient. This coefficient can for example be calculated from the co-occurrence number of the first and the second product on the web pages of the result of the queries comprising a search request for web pages comprising both the first product (PJ) and the second product ( P,). This number of co-occurrence or coexistence of the two products can be calculated on the same resulting web page of the search or on different web pages. The coefficient can also contain both types of number. Thus, in step E203 of FIG. 2, different similarity coefficients have been calculated directly or by estimation from the search results and for the N second products of the set or subset of products of the catalog. These coefficients are stored so that in step E204, a comparison is made to find the largest coefficient. The product or products corresponding to the largest coefficients are thus selected to be recommended to the user having selected the first product.

A similarity threshold can of course be set to recommend products that have a similarity coefficient above this threshold. A maximum number of products to recommend can also be set, for example between 1 and 5, in order to offer only the most similar products. The recommendation can be made by simply sending the name of the product, sending a corresponding image or other means of identifying the product to recommend. This method is performed to recommend a product from a first selected product. It can of course be implemented for other selected products and similarities already calculated and stored can be used when selecting another product catalog. The method can be implemented in real time, during the online consultation of a product on the internet. It may just as well have been implemented beforehand for a specific product catalog. The catalog then including an association between the proposed products and those to be recommended according to the result of the process implemented.

Figure 3 now illustrates a hardware embodiment of a recommendation device 300 according to an embodiment of the invention. This device can be integrated with a recommendation server in a communication network or any other server or equipment having access to product catalogs and search engines. This device implements the recommendation method described with reference to FIG. 2 by the main steps E202 to E204.

It has input communication means 301 and output 302 for communicating with other entities of the communication network, for example of the Internet type, in particular to receive from a user a selection of a first product d 'a catalog of products available on the network. Materially, this device 300 comprises a processor 306 cooperating with a memory block 307 having a storage and / or working memory MEM. The processor controls processing modules able to implement the method according to the invention. Thus, this device comprises a module 303 for determining a plurality of search requests from the first selected product (PJ) and a set of second catalog products and 302 sending means of these requests (Req. ) to at least one search engine. It comprises a obtaining module 304 of the number of results obtained from the search (Res.Req.) Of the at least one search and calculation engine, for the set of second products of the catalog, of a similarity coefficient between the first and the second product. It also comprises a recommendation module 305 able to recommend at least a second product (Pk) of the catalog according to the calculated similarity coefficients. Typically, a comparator may be part of the recommendation module for comparing the calculated similarity coefficients and selecting the largest similarity coefficient of all of the second products. The associated product is then the one that is recommended.

This recommended product is sent via the communication means 302 to the user having selected the first product. At initialization, the instructions of the computer program (prog.) Are for example loaded into a RAM memory (Random Access Memory in English) of the memory block 307 before being executed by the processor 306. The processor executes the code instructions the computer program (prog.) for implementing the steps of the recommendation method, and in particular the steps of determining a plurality of search requests from the first selected product and from a set of second products of the catalog and sending these requests to at least one search engine, obtaining the number of results obtained from the search of the at least one search engine, calculating, for the set of second products of the catalog, a coefficient similarity between the first and the second product and recommending at least one second product of the catalog based on the calculated similarity coefficients.

Typically, the description of FIG. 2 repeats the steps of an algorithm of such a computer program. The computer program can also be stored on a memory medium readable by a reader of the device or downloadable in the memory space thereof.

The memory MEM generally records all the data necessary for implementing the recommendation method and in particular the different similarity coefficients calculated for the set of second products in the catalog.

Claims (11)

REVENDICATIONS1. Method for recommending products contained in a catalog of products that can be consulted on a communication network, from the reception of information for selecting a first product from a user terminal, characterized in that it comprises the following steps: - determining (E202) a plurality of search requests referencing the first product and a set of second catalog products and sending these requests to at least one search engine; obtaining (E203) the number of results obtained by the at least one search engine for each of the requests; - calculation (E203) and storing, for the set of second catalog products, a similarity coefficient between the first and the second product as a function of the number of results obtained; sending recommendation information (E204) of at least one second product from the catalog to the user terminal based on a comparison between the calculated similarity coefficients. 2. Method according to claim 1, characterized in that the search requests comprise a search request for web pages comprising the first product and a search request for web pages comprising a second product or a search query comprising both the first product and a second product. 3. Method according to claim 2, characterized in that the search requests further comprise a search request for web pages comprising either the first product or the second product. 4. Method according to one of claims 1 to 3, characterized in that the similarity coefficient is an estimate of a jacquard coefficient defined as the ratio between the number of pages comprising both the first and the second product on the number of pages containing either the first product or the second. 5. Method according to claim 1, characterized in that the search queries comprise a search request for web pages comprising either the first product or a second product and a search request for web pages comprising both the first product and the first product. second product. 6. Method according to claim 5, characterized in that the similarity coefficient is a jacquard coefficient defined as the ratio between the number of pages comprising both the first and the second product on the number of pages comprising either the first product. , the second. 7. Method according to claim 5, characterized in that the similarity coefficient between the first product and a second product takes into account a number of co-occurrence of the first and second product on the resulting web pages of the search. 8. Device for recommending products contained in a catalog that can be consulted on a communication network, the recommendation being made from the receipt of information for selecting a first product from a user terminal, characterized in that it comprises: a module for determining (303) a plurality of search requests referencing the first selected product and a set of second products of the catalog and a module for sending these requests to at least one search engine; ; a module for obtaining (304) the number of results obtained by the at least one search engine for each of the requests and for calculating and memorizing means, for the set of second products in the catalog, of a coefficient of similarity between the first and the second product according to the number of results obtained; a module for sending a recommendation information (305) of at least one second product from the catalog to the user terminal, according to a comparison between the calculated similarity coefficients & 9. Recommendation server comprising a device according to claim 8. Computer program comprising code instructions for carrying out the steps of the recommendation method according to one of claims 1 to 7, when these instructions are executed by a processor. 11. Information carrier, readable by a processor, on which is recorded a computer program comprising instructions for performing the steps of the method of recommendation according to one of claims 1 to 7.