Autonomous self-learning advertisement generating method

Abstract

The present invention relates to a method for autonomous self-learning digital advertisements generation, comprising the steps of: a. generating at least one digital advertisement using a generative model implemented on a computing system, by providing as input at least one source data point context about a content of the digital advertisement and at least one destination data point context about at least one user to which the digital advertisement is destined; b. presenting said digital advertisement to the at least one user through a communication channel; c. collecting an advertisement feedback from the at least one user; d. using the advertisement feedback from at least one user, to learn the generative model to optimizing an impact parameter depending on the at least one source data point context, the at least one destination data point context and the advertisement feedback of the at least one user; e. generating at least one new digital advertisement specifically for the at least one user using the learned generative model.

Description

Autonomous self-learning advertisement generating methodTechnical domain

[0001] The present invention concerns a method for autonomous selflearning digital advertisements generation using generative models.Related art

[0002] Digital ads, such as banners, native ads, video ads, interstitial ads, sponsored posts or content, etc, are widely used on web sites, social networks, mobile apps, etc. for enhancing brand visibility or driving user engagement. The digital ads may include texts, images, videos, audio, hypertext element, etc.

[0003] Most digital ads are generated by publicist using software such as text and image editors, and then saved into a format such as html that can be rendered by a browser on the recipient side. This is a very timeconsuming and empirical process; a lot of time is often spent on elaborating ads that are not effective and fail to generate much user engagement.

[0004] Beside human ad generation, generative Al is more and more widely used for generating ad contents in a faster and more effective way. In particular, the application of Large Language Models (LLMs) in the generation of digital advertisements represents a transformative step in the field of digital marketing. LLMs are widely used for digital advertising, due to their ability to generate creative, diverse, and contextually appropriate content. Unlike traditional ad creation methods that rely heavily on human input, Generative Al using LLMs can autonomously generate textual and visual ad content, including headlines, product descriptions, and persuasive narratives. The model can be trained to adapt its content generation tovarious advertising platforms, such as social media, search engines, and email marketing, ensuring platform-specific optimization.

[0005] The text content of the ads can for example be generated with generative Al tools such as ChatGPT, Gemini, and many others, while the image content can be generated with DALL-E Midjourney, etc. Other generative Al tools have also been developed for generating ads in a faster and cheaper way.

[0006] Other similar and more recent tools, often referred to as "Large Document Models" (LDM) allow multimodal content generation such as documents containing images, images with executable content or more specifically in the context of digital advertisement, banners with clickable content, animated images, etc.

[0007] Customization is another key feature, where the generative model tailors content based on specific client requirements. This level of customization ensures that each ad is not only unique and creative but also aligned with the specific branding and marketing objectives of the advertiser.

[0008] Once the ad has been generated from human and / or Al input, it is often desired to test how effective it is. To this effect, the concept of a split run is often used in digital marketing for determining the efficiency of various marketing elements by comparing different versions of an ad, differing in one or more variables. This variable could range from the phrasing of the ad, the choice of an accompanying image, or design elements like the colour of a call-to-action button, to name a few. The resulting performance data, such as engagement rates, conversion ratios, or click-through frequencies, are then meticulously analyzed to deduce which version yields superior outcomes.

[0009] Incorporating a split run into digital marketing strategies signifies a leap towards data-driven decision-making, substantially reducingreliance on speculative or intuition-based approaches. This knowledge not only bolsters campaign effectiveness but also paves the way for more engaging and successful consumer interactions.

[0010] However, the process of generating and testing multiple variants of a digital ad is time-consuming, even if the variants are generated by an Al-based system. It is highly inefficient because a set of multiple ads must be generated and tested before the ad that performs best among those ads is selected; the other, unselected ads are typically wasted.

[0011] Furthermore, the selection of the different variants to generate and test is largely empirical; for example, in most cases an advertiser will create two different variants of an ad, test them, and select the best of the two. It is usually not feasible to create and test a very large number of variants. Therefore, the advertiser only learns that variant A is better than variant B, but does not know whether other variants C, D, etc. could perform even better than A.

[0012] Moreover, creating an ad that takes into account not only information from the advertiser (type, message, budget, etc.) but also from the user / customer context (location, age, gender, etc.) requires a complex centralisation effort even before the ad creation process begins. In addition, some of these data may change through time so that the data gathering has to be performed multiple times to ensures the creation of an add which is relevant according to the advertiser and customer data.Short disclosure of the invention

[0013] It is therefore an aim of the present invention to overcome the limits of the state-of-the-art.

[0014] It is therefore an aim of the present invention to propose a new Al-based method for generating personalized and effective digital ads.

[0015] It is also an aim of the invention to provide a method allowing an autonomous management of digital advertisements in real-time.

[0016] This aim is attained by a method for autonomous self-learning digital advertisements generation, comprising the steps of: a. generating at least one digital advertisement using a generative model implemented on a computing system, by providing as input at least one source data point context about a content of the digital advertisement and at least one destination data point context about at least one user to which the digital advertisement is destined; b. presenting said digital advertisement to the at least one user through a communication channel; c. collecting an advertisement feedback from the at least one user; d. using the advertisement feedback from at least one user, to learn the generative model to optimizing an impact parameter depending on the at least one source data point context, the at least one destination data point context and the advertisement feedback of the at least one user; e. generating at least one new digital advertisement specifically for the at least one user using the learned generative model.

[0017] The term learning in the context of generative models refers to the degree to which the models responses are not only coherent, but also in harmony with human values, ethical considerations and the specific intentions of the users.

[0018] The method is thus based on a novel learning of the generative models used to generate the digital advertisement. This learning is based, at least in part, on the feedback collected from each user, thereby improving the ability of the model to produce advertisements that are personalised for each user and efficient for that user.

[0019] According to one aspect, the user's feedback is not (or not only) taken into account for selecting the most efficient ad from a set of previously generated ads, but during ad generation, at the level of thegenerative model. This avoids the waste of test ads that won't be run after the test.

[0020] According to one aspect, the generative models are userdependent, so that the digital ad that is generated will be different for each user. If, for example, a first user prefers a colloquial language style and reacts less positively to digital ads written in an overly formal style, the generative model used for this user will be aligned so as to generally generate ads in a colloquial style, for example by favouring first-name terms.

[0021] Each user's interaction with ads can be used by the generative model associated with that user to learn.

[0022] Furthermore, generative models can continuously learn and evolve based on feedback loops, allowing the advertising content to remain dynamic and responsive to changing market trends and consumer behaviours. This adaptability ensures that digital ads remain effective and engaging over time. It therefore allows an autonomous management in real-time of an advertisement campaign that would be impossible for a human to carry out efficiently.

[0023] In one embodiment, the steps b. to e. above are executed iteratively until an optimal impact parameter is obtained.

[0024] The method may further comprise a step of: generating a diffusion plan of the new digital advertisement to the at least one user using the learned generative model.

[0025] The learning of the generative model can therefore be used to further generate not only a new digital advertisement, but also a plan orchestrating the diffusion of the advertisement to one or more users.

[0026] The diffusion plan may comprise: a time-scheduling determining a time at which the new digital advertisement will be presented to the at least one user and / or a number of times that the new digital advertisement is to be presented to the at least one user.

[0027] The time-scheduling may determine a time at which the new digital advertisement will be presented to the at least one user, said timescheduling being based on a data history of an internet browser data history of the at least one user.

[0028] The feedback of the at least one user can comprise standard metrics such as an awareness rate, a conversion rate and / or a sale rate.

[0029] The step of collecting the advertisement feedback can be performed user interaction, e.g. by means of a chatbot interacting with the user.

[0030] The information represented by the source data point context can comprise a time of the year, an hour of the day, a number of repetitions of the digital advertisement, a size of the digital advertisement, a communication channel and / or a budget of an advertiser.

[0031] The information represented by the destination data point context can comprise a geographical situation of the at least one user, a meteorological situation at the at least one user's location, a sport's game result and / or an actuality situation.

[0032] Therefore, the generative model can be trained by taking advantage of contextual information about both the advertiser / content of advertisement (source data point context) and about the target user (destination data point context).

[0033] In order to reduce the number of iterations of the method, the learning of the generative model can include a human-in-the-loop (HITL) sub-step.

[0034] The method may further comprise a step of digital advertisement testing by providing the digital advertisement as an input into a dedicated verification system for determining whether the digital advertisement corresponds to an appropriate content.

[0035] The creativity level of the generative model can be adjusted using a temperature hyperparameter for regulating the randomness of the output of the model.

[0036] The temperature hyperparameter can be determined based on the source data point context and / or on the destination data point context.Short description of the drawings

[0037] Exemplar embodiments of the invention are disclosed in the description and illustrated by the drawings in which:Figure 1 illustrates schematically a method for generating a digital advertisement using a feedback of a user.Figure 2 illustrates schematically various input parameters provided into a LDM for generating a digital advertisement.Figure 3 illustrates schematically a method for generating a digital advertisement and a diffusion plan of the digital advertisement.Examples of embodiments of the present invention

[0038] With reference to Figure 1, the present method comprises a first step of generating a digital advertisement 2 using a generative model implemented on a computing system 1 taking as input both source data point context and destination data point context.

[0039] Source data point context relates to the data describing what the digital advertisement is supposed to communicate, while destination data point context relates to data describing the user (or customer) and his context.

[0040] In the context of the present disclosure, generative model refers to any multimodal generative Al model adapted to generate multimodal content (text, image, video, executable file, files with specific extensions of existing software, etc.), in particular attention-based transformer models. Therefore, the source and destination data point context are typically encoded as vectors or vector embeddings to be provided as inputs to the model. It is understood that large language models are particular cases of generative models.

[0041] Generative models can be implemented on a wide range of computing systems from personal computers to high-performance computing clusters depending on factors such as the complexity of the model, the size of the dataset, the computational resources required for training and inference, and the desired speed and efficiency of the system. The following list provides non-limitative examples of such computing systems:• Standard desktop or laptop computers for small-scale experiments and prototyping.• Workstations equipped with powerful CPUs and GPUs are suitable for training more complex generative models on larger datasets. Workstations often provide better performance and memorycapacity compared to standard desktop computers, allowing for faster training and experimentation.• Cloud computing platforms such as Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure offer scalable computing resources that can be used to train generative models on large datasets. These platforms provide access to virtual machines with high-performance GPUs, which are well-suited for deep learning tasks.• For significant computational requirements, dedicated GPU servers can be deployed to train generative models. These servers are equipped with multiple high-performance GPUs and are optimized for deep learning workloads, allowing for faster training times and larger batch sizes.• HPC clusters provide access to large-scale computing resources for running computationally intensive simulations and analyses. Generative models can be implemented on HPC clusters to leverage parallel computing capabilities and accelerate training on massive datasets.

[0042] The source and destination data point context typically encode information coming from different data streams and / or from different modalities such as text, image, video, sound, code or any combination thereof.

[0043] As illustrated on Figure 1, at least one source data point context 10 representing information about a content of the digital advertisement 2 and at least one destination data point context 11 representing information about at least one user 3 to which the digital advertisement is destinated are inputted into a generative model 1. Based on those data point context, the model generates as an output a digital advertisement 2.

[0044] In a second step of the method, the digital advertisement 2 is presented to a user 3 through a communication channel.

[0045] A communication channel refers to any mean of communicating a digital advertisement to a user. In particular this includes the Internet, any social media platform, email, mobile apps, digital TV, etc.

[0046] In a third step, an advertisement feedback of the user 3 is collected. The advertisement feedback defines any action of a person made in response to being exposed to the digital advertisement.

[0047] In an embodiment, the advertisement feedback is collected in a direct manner, e.g. by explicitly asking the user 3 to fill in a form with questions about the relevance of the digital advertisement or by making the user rate the digital advertisement.

[0048] In one embodiment the user is displayed an auto-generated landing webpage in which one or more feedback options are provided. For example, the advertisement feedback can be collected by means of a chatbot interacting with the user and asking questions about the displayed advertisement. A reward mechanism can be implemented to encourage the user to provide an advertisement feedback.

[0049] Alternatively or complementarily, data related to the behaviour of the user 3 after having been exposed to the digital advertisement can be gathered, e.g. using proxies allowing measuring a click-through rate, accessing a web browsing data history, an increasing internet traffic, an increasing sale level, an increase in the sharing of a post on social media, etc.

[0050] The collected feedback is then provided as an input to the generative model for the purpose of aligning the model so as to generate a new digital advertisement whose impact on the at least one user is more effective. This can be achieved by optimizing an impact parameter which depends on the source data point context, the destination data point context and the advertisement feedback 3.

[0051] The impact parameter is dependent on case-by-case as measuring the impact of an advertisement strongly depends on what the advertiser wants to measure.

[0052] This alignment can be for example performed by reinforcement learning algorithms such as Reinforcement Learning from Human Feedback (RLHF), or Direct Preference Optimization (DPO). Alternatively or complementarily, the generative model can also be aligned through incontext learning (i.e. prompt engineering) and / or fine-tuned the impact parameter using parameter-efficient fine-tuning (PEFT) methods such as LoRA (Low-rank adaptation), QLoRA (Quantized low-rank adaptation), SFT (supervised fine-tuning), etc..

[0053] In particular, the generative model can be a pre-trained generative model for reducing storage and computational costs. In the latter case, the alignment of the model may comprise only fine-tuning optimization of the impact parameter.

[0054] The impact parameter is typically implemented as a loss function, meaning that the loss function of the machine learning model depends on the impact parameter, depending on the particular form of the advertisement feedback that has been collected. The expression impact parameter is not to be confused with the set of parameters of the neural network implementing the generative model. The impact parameter depends on the parameters of the model which are to be optimized to align the model. The impact parameter typically represents an impact of the digital advertisement by the user.

[0055] For example, optimizing the impact parameter may include maximizing a conversion rate, a sale level, etc. measured during collecting the advertisement feedback.

[0056] As a fourth step of the method, at least one new digital advertisement is generated using the learned generative model.

[0057] In order to enhance the learning of the generative model, the presentation of the digital advertisement(s) to users can be split into separate runs with specific characteristics.

[0058] In one embodiment, the runs are time sequentially scheduled. This may allow the generative model to learn at which time (e.g. of the year, of the day, etc.) the digital advertisement is the most efficient, or to create a new digital advertisement which is more relevant at a given time (e.g. creating an advertisement related to Christmas at the end of the year).

[0059] In one embodiment, the runs are conducted on batches of different sizes. The batch size of a run defines the number of identical advertisements included in one single run. Indeed, the same digital advertisement may be presented a plurality of time in one run, e.g. during a particular event such as a sport / e-sport game, a given time slot such as an evening, a time of the year, etc.

[0060] In one embodiment, the runs may overlap each other or may be conducted in parallel. This defines the number of different runs that are overlapping / simultaneously launched

[0061] While all these split-runs parameters may be manually tuned the generative model can further be used to learn and generate a diffusion plan of the new digital advertisement to the at least one user. Indeed, through the optimization of the impact parameter, the generative model also gains knowledge, after each iteration, of the run parameters that optimize this impact parameter. Hence a diffusion plan determining as before the scheduling, the batch sizes and / or the parallelization / overlapping of the runs can be obtained as output of the generative. In addition, the impact parameter can be optimized with respect to the communication channel so that the diffusion plan comprises a communication channel maximizing the impact of the presentation of the new digital advertisement. It may also take into account a budget of anadvertiser to determine an optimal scheduling strategy as well as optimal communication channels with respect to this constraint.

[0062] The run parameters can typically be assigned weights in the neural network model of the generative model and therefore be optimized during the execution of the present method.

[0063] As such, the learned generative model can serve as planification agent capable of both providing a digital advertisement maximizing the impact parameter, and a diffusion plan maximizing the impact parameter. Advantageously, this greatly improves the efficiency of any traditional marketing process in which not only all those tasks are long and costly, but also hardly reusable.

[0064] In order to strengthen generative model, one or more iteration of the steps of the method can be performed. The digital advertisement can therefore be generated and presented to the users a certain amount of times until the impact parameter is considered optimal. This iterative learning is typically performed by means of a reinforcement learning algorithm as described above.

[0065] The threshold for the impact parameter to be considered optimal varies from case-to-case, in particular depending for example on the used communication channel, on the type of the feedback provided by the user, and / or on the specific interests of the advertiser, etc.

[0066] However, in order to avoid users to be unnecessarily exposed to new digital advertisements with little changes for the purpose of slightly improving the learning of the generative model, such a threshold for the impact parameter may be specified.

[0067] Advantageously, the time-scheduling determining the time at which the new digital advertisement will be presented to the user(s) can bedetermined based on data gathered on a device on which the user is presented the digital advertisement. Such data may typically include a browsing data history, purchasing habits stored by online websites. This can be collected as part of the feedback on the advertisement or separately by means of standard analytic and tracking tools.

[0068] It is then possible to optimize the impact parameter according to these data by also providing them as input into the generative model in order to learn it accordingly. Not only the new digital advertisement can have more impact to the user, but also an eventual diffusion plan can be more accurate by proposing a relevant time-schedule for the next run.

[0069] As already mentioned, the new digital advertisement generated by the generative model typically represents a multimodal document, i.e. a document comprising at least a text, an image, a video, an executable file, a file adapted to the use of a given software (i.e. with the right format extension such as .docx, .xlsx, .ppt, .pdf, .csv, .jpg, .svg, .gif, .mpeg, .webm, etc.), computer code, or any combination of all those modalities.

[0070] A digital advertisement will indeed often comprise several different modalities as an animated picture with a clickable element or a text banner with an embedded hyperlink.

[0071] Additionally, a generated diffusion plan may include all elements necessary to perform the split-runs, such as dates or hours at which a run has to be performed, batch sizes, organization of parallel runs, etc.Therefore the diffusion plan typically comprises simple text elements or images as well as more involved formats compatible with time managing software, document processing software and / or database.

[0072] As mentioned before, collecting the feedback of the user(s) can take various forms according to the communication channel through which the digital advertisement is presented.

[0073] In some embodiments, the feedback is collected by using a feedback form displayed to the user during or at the end of the presentation of the digital advertisement, for example asking to rate, to answer questions and / or to describe the experience of the digital advertisement. The feedback form may be a classical form or a chatbot asking directly the question to the user on an auto-generated webpage.

[0074] Alternatively or complementarily, a feedback is collected indirectly by measuring data related to the behaviour of the user(s) during or after the presentation of the digital advertisement. For example, data of the user can be tracked to determine, after being presented the digital advertisement, if he visits a particular website, how much time he spends on a website, how many interactions he performs on a website, how much he spends on a website, if he shows interest for a product or a brand on a social media by posting, sharing or liking content related to the advertised product, etc.

[0075] More generally, any conversion rate metric measuring a percentage of the users that perform a predetermined action (such as purchasing, sharing an article, signing up for a newsletter, downloading a resource, etc.) on a website after being presented the digital advertisement, can be considered as part of the feedback of the user. When the predetermined action is the purchase of a product / service, the conversion rate is also often referred to as sale conversion rate.

[0076] Alternatively or complementarily, the feedback can also comprise an awareness rate related to a given product, service or brand, i.e. a measuring of whether the user recognize the product / service / brand in a given context.

[0077] The source data point context 10 represents information about the content of the digital advertisement to be generated. This can include various contextual information about the advertiser and / or his product / service which is to be advertised by the digital advertisement. Thismay include a budget of the advertiser conditioning the format of the digital advertisement as a video spot broadcasted in prime time is much more expensive than a clickable banner on a website. This may also include information about the type of product / service that is to be advertised in order to calibrate the characteristics (visual, textual, etc.) of the digital advertisement to the encompassed target customer. For example, an advertisement for a luxurious product doesn't have the same esthetical codes than advertisement for child products, etc.

[0078] The destination data point context 11 representing information about the user to which the digital advertisement is to be presented may comprise contextual information that may influence his behaviour in response to the advertisement. In particular, the source vector can comprise information such as a geographical situation of the user (e.g. gathered by tracking the IP location), the meteorological situation at the user's location (e.g. by matching the IP location of the user with a dedicated meteorological website), the result of a sport's game from a user's favourite team (e.g. by gathering history data of the user's computer), an actuality or political situation (e.g. using the IP location of the user and / or his data history).

[0079] All that information may be used to generate a digital advertisement that match the particular context of the user so as to maximize the impact of the advertisement.

[0080] Figure 2 schematically illustrates source data point context 10 and destination data point context 11 provided as inputs to generative model 1 which outputs a digital advertisement 2. This digital advertisement is presented to a user 3 whose feedback is collected and provided as input into the LDM for learning purpose. As illustrated, the source data point context 10 can represent financial information (e.g. an available budget of the advertiser), product / service information, scheduling considerations (e.g. sales period, holiday periods, etc.). The destination data point context 11 can include gender information (gender of a user or group of users),language related information (e.g. formal / unformal language), meteorological conditions at the user's location, results of sport / actuality in at the location of the user, etc.

[0081] In one embodiment, the learning of the generative model is further strengthened by a human-in-the-loop (HITL) process. This refers to a human improving the machine learning over random sampling by selecting the most critical data needed to refine the model. This results in a significant reduced number of iterations of the method until obtaining a digital advertisement having a maximized impact with the user. Hence it prevents the user from being presented to many advertisements before the impact level of the advertisement becomes sufficient.

[0082] As the content of the generated (new) digital advertisement relies on the generative model, it may happen that its content is inappropriate in respect of one or more factors such as the age of the user, moral or ethical norms in a given area or for particular groups of users, the law of the jurisdiction where the user or the advertiser sits. Therefore the method can advantageously comprises a step of testing the digital advertisement, either with human control and / or with a self-learning system for determining if the content of the advertisement is appropriate with respect to a chosen metric. Over several runs of testing, the selflearning system can learn to more easily detect if a content is inappropriate and prevent its presentation to one or more users.

[0083] Various supervised / unsupervised self-learning or reinforcement learning algorithms can be used to execute this testing step, taking as input the generated digital advertisement and outputting a score or a probability of the advertisement's content to be appropriate or not.

[0084] In one embodiment, the amount of creativity involved in the generation of the digital advertisement can be parametrized. This feature is often defined by a temperature hyperparameter regulating the randomness of the output if the generative model. A high temperatureresults in more randomness in the generation of the output, i.e. the model is "more creative" and a low temperature results in less randomness in the generation of the output, i.e. the model is "less creative". This temperature hyperparameter typically takes a value between 0 corresponding to a model with low creativity and 1 corresponding to a model with high creativity.

[0085] This can help to maximize the impact of the advertisement as it allows to further tailor the generated advertisement based on knowledge of the user and / or of the content of the advertisement. Indeed, an advertisement for a bank service or for a funeral service may beneficiate from a model wherein the creativity is limited to prevent inappropriate content, while advertisement for video games or targeted to a young audience can on the contrary beneficiate from a high temperature of the model.

[0086] The temperature can be determined by a human based and / or it can be automatically tuned within the generative model advertisement generation typically through a reinforcement learning algorithm.

[0087] In one embodiment, the temperature hyperparameter is automatically tuned by the generative model based on the source and destination data point context (10,11). The user's feedback may also be used to tune the temperature on the next iteration.

[0088] In one embodiment, a threshold for the impact parameter can be determined so that whenever a digital advertisement is generated by the generative model based on an impact parameter meeting this threshold, it is considered that the digital advertisement has a sufficient impact to the user. Such digital advertisements can then be stored in a database on a memory module to allow their reuse as they are considered sufficient. For each sufficient advertisement, an impact score based on the impact parameter can be computed and stored in the database.

[0089] This threshold can be determined by the advertiser according to standard advertisement metrics measuring the impact as listed above, and / or it can be tuned by the generative model itself through a reinforcement learning algorithm.

[0090] The stored digital advertisements can then be reused, possibly together with characteristics of the users and / or products based on which it has been generated.

[0091] The digital advertisements can be stored per se, i.e. as files or raw data, but the learned generative model can also be stored as a way of storing the generated advertisements. This facilitates transfer learning, i.e. the use of what has been learned by the generative model to another model.

[0092] This may allow the storing of specific learned generative model each having been trained on specific source / destination data point context. A database of styles can therefore be obtained by storing learned models in specific conditions and only needing fine-tuning to be reused in another similar context. As an example, generative models can be learned according to a specific time of the year (e.g. Christmas, Easter, holidays, Black Friday, etc.). Each of these models can then be reused without having to conduct the split runs.

[0093] The method can further comprise an initial step of verifying in a database of stored digital advertisements if there exists a digital advertisement considered as sufficient, e.g. having a sufficient impact score, so that execution of steps a. to e. can be saved. The verification is typically based on knowledge about the user, e.g. the information encoded by the destination vector.

[0094] This allows to take advantage of already generated digital advertisement and therefore to save time and reduce the costs of a splitrun.

[0095] The method of the present invention is advantageously implemented in a system comprising a local server executing the generation steps. An advertiser can connect remotely through a microservice to a platform on which he can describe his needs and constraints: product, budget, market, targeted customer, communication channel(s) etc. The digital advertisement is then generated on a local private server to ensure privacy of the data and presented to the user(s) through the determined communication channel(s).

[0096] The method allows the creation of a multi-agent system comprising a plurality of generative sub-models. Each generative sub-model manages the life-cycle of its advertisement and communicates its impact parameter to the multi-agent system which is able to define in real-time the number of agents running simultaneously and to automatically determine which of the digital advertisements have to be kept alive based on the impact parameters. In other words, the method allows the creation of an autonomous multi-agent system.

[0097] The most performing agent save its generative model and stores it for later agents to use it. The aim is here to further exploit the learned generative model.Reference numerals Generative model Source data point context Destination data point context Digital advertisement User Diffusion plan

Claims

Claims1. A method for autonomous self-learning digital advertisements generation, comprising the steps of: a. generating at least one digital advertisement (2) using a generative model (1) implemented on a computing system, by providing as input at least one source data point context (10) about a content of the digital advertisement and at least one destination data point context (11) about at least one user (3) to which the digital advertisement (2) is destined; b. presenting said digital advertisement (2) to the at least one user through a communication channel; c. collecting an advertisement feedback from the at least one user (3); d. using the advertisement feedback from at least one user (3), to learn the generative model (1) to optimizing an impact parameter depending on the at least one source data point context (10), the at least one destination data point context (11) and the advertisement feedback of the at least one user (3); e. generating at least one new digital advertisement (2) specifically for the at least one user (3) using the learned generative model (1).

2. Method according to claim 1, further comprising a step of f. communicating the impact parameter to a multi-agent system throughout a life-cycle of the digital advertisement in order for the multiagent system to automatically determine if the digital advertisement has to be kept alive based on the impact parameter.

3. Method according to claim 1, wherein the steps b. to e. are executed iteratively until an optimal impact parameter is obtained.

4. Method according to any of the claims 1 to 3, further comprising the step of:generating a diffusion plan (4) of the new digital advertisement (2) to the at least one user (3) using the learned generative model (1).

5. Method according to claim 4, wherein the diffusion plan (4) comprises: a time-scheduling determining a time at which the new digital advertisement will be presented to the at least one user (3).

6. Method according to claim 5, wherein the time-scheduling is based on a data history of an internet browser data history of the at least one user (3).

7. Method according to any of the claims 4 to 6, wherein the diffusion plan (4) comprises: a number of times that the new digital advertisement is to be presented to the at least one user (3).

8. Method according to any of the preceding claims, wherein the advertisement feedback of the at least one user (3) comprises an awareness rate, a conversion rate and / or a sale rate.

9. Method according to any of the preceding claims, wherein the step of collecting the advertisement feedback of the at least one user (3) is performed by user interaction.

10. Method according to claim 8, wherein the user interaction comprises displaying an auto-generated webpage to the at least one user (3) allowing the collection of an improved feedback.

11. Method according to any of the preceding claims, wherein the source data point context (10) comprises a time of the year, an hour of the day, a number of repetitions of the digital advertisement, a size of the digital advertisement, a communication channel and / or a budget of an advertiser.

12. Method according to the preceding claim, wherein the destination data point context (11) comprises a geographical situation of the at least one user, a meteorological situation at the at least one user's location, a sport's game result and / or an actuality situation.

13. Method according to any of the preceding claims, wherein the step of learning the generative model (1) to optimize the impact parameter comprises a human-in-the-loop sub-step.

14. Method according to any of the preceding claims, further comprising a step of digital advertisement testing by providing the digital advertisement (2) as an input into a dedicated verification system for determining whether the digital advertisement corresponds to an appropriate content.

15. Method according to any of the preceding claims, wherein the step of learning the generative model (1) to optimize the impact parameter comprises a sub-step of determining a temperature hyperparameter of the generative model for regulating the randomness of the output of the model.

16. Method according to the preceding claim wherein the temperature hyperparameter is determined based on the source data point context (10) and / or on the destination data point context (11).

17. A multi-agent system comprising a plurality of agents, each agent being able to carry out the method of any of the claims 1 to 16 by means of a generative sub-model, each generative sub-model managing the life-cycle of an advertisement and communicating an impact parameter of the advertisement to the multi-agent system which is able to define in realtime the number of agents running simultaneously and to automatically determine which of the digital advertisements have to be kept alive based on the impact parameters.

Images