RAD AI today released client due diligence designed to test the veracity and effectiveness of its AI EQ optimization platform. Jeremy Barnett, RAD AI CEO, said, “Countless companies in the marketing industry use AI as a marketing buzzword to differentiate themselves from their competition. It was because of this justified skepticism that we welcomed one client’s technical challenge. We’re proud to release the results.” The following questions were asked and answered as part of the process: Q: “Objective function for the more ‘nuanced’ types of model — e.g., for something like an NLP (natural language processing) sentiment model, are you using something off-the-shelf, and fine-tuning with your own labeled data to ensure it captures the type of language you intend?” A: Choosing a single model for sentiment analysis is challenging due to the nature of the research, the data involved, the labeling methodology, the models applied and, lastly, the results from the analysis.  Plutchik’s wheel of emotion was found to be the most expressive and descriptive, with eight different emotions and “levels” within each emotion. Using an off-the-shelf solution, the dataset from social media was trained via classification to give a distributed probability of results very close to human judgment. Due to concerns about the labeling methodology being automated (distant supervision), RAD AI did not want to fine-tune the model to its specific dataset as it had intended to keep the “knowledge” of the model as “generalized” as possible. In some cases, it has built proprietary solutions since limited to no research exists for this specific NLP question. For example, the RAD AI EQ writing style model is a unique solution. For content teams that use uncommon data sets (i.e., email, ads, blogs), the ability to fine-tune performance models based on channel and objective is critical — this is not an “out-of-the-box” solution, as content performance models need customization based on their optimization requirements. Q: “Regarding objective functions — are you typically making recommendations based on a single objective (e.g., engagement)? Or are you optimizing models for a set of outcomes?” A: The solutions are capable of optimizing toward multiple outcomes (improved KPIs relevant to the provided data type) based on the client’s request, however, adding too many outcomes causes limitations on the number of proposed optimizations that it can provide. In general, this analysis is performed during the optimization stage of post-processing and is based on the RAD AI EQ proprietary ranking methodology. RAD AI has found that there are many different variables that impact performance on social posts, articles, emails, and paid ads. It uses a proprietary RAD AI EQ engagement analysis as a catalyst for a scoring engine that weights optimizations for audience type. Depending on the client’s “engagement objectives,” it reveals the “best” suggestion mapped to that “engagement objective” (ex: a client wants to optimize towards page views; it then suggests the result most likely to give the most page views). Q: “For individual models trying to optimize a certain outcome (e.g., impressions), how do you think about your training set? For example, with computer vision, and determining what types of photos create better engagement for a given influencer, how do you think about the ‘right’ set of images to train that model?” A: The models for computer vision problems are based on large datasets of mixed data from the identified social media channels (outliers are filtered out). This data is used for the initial training of the DNN (Deep Neural Networks) model, then fine-tuned by each influencer. At this stage, variation analysis between the original dataset and new incoming data is performed to eliminate the model being overtrained. RAD AI recognizes that there are common ways to describe popular images as a group, however, when applied to a certain audience — such descriptions can and should be modified to achieve the best possible outcome. The methodology is designed with this flexibility in mind. Q: “For ongoing campaigns, using RAD AI EQ insights, what is the baseline comparison for performance tracking? Previous campaigns for the same client? Are you rolling any of that data back into the models for retraining? Any types of experiments you run (A/B tests or multi-arm bandits) within that campaign to ensure you’re right?” A: RAD AI EQ breaks down and baselines the historical performance on the actual digital channels the client plans to distribute the influencer-created content on. The metrics it benchmarks are ultimately dictated by the client, but the models allow for data variability.  RAD AI EQ establishes a client’s KPIs through their performance analytics (i.e., Google Analytics, Adobe for click-through rates, conversion rates, time on site, etc.). From there, it creates AI EQ-informed content that its technology compares to content that is not AI EQ-informed. The client always knows the performance delta between AI EQ-informed content and content that is not using the AI EQ suggestion. To further validate the efficacy of its models, RAD AI EQ uses statistical methods that measure the positive impact and significance (a null hypothesis that the model has 0 weight). It generally optimizes any media its clients provide alongside any new media created. Additionally, any new media created that goes live is also used to further fine-tune the model, thus creating an AI EQ feedback loop endemic to the needs of the client. Q. “How does RAD’s AI EQ understand each client’s audience?” To understand the client’s audience, data is collected from their respective user channels. The customer’s audience is then categorized using their interests, interest levels, emotion, and sentiment analysis.  As part of the extraction process, it uses off-the-shelf pre-trained models to perform interest extraction and sentence embedding. This information is then fed into the RAD AI EQ proprietary software to be optimized using a combination of dimension reduction and clustering for segments of each audience. Once analyzed, it is able to provide the most important interests per group segment and label them. The compilation of relevant interests for each segment of the audience can be scaled across millions of users. Each group has relevance with correlating user engagement to provide priority to different audience interests. Additionally, it connects sentiment analysis components to help marketers better understand which audience should be targeted and why.  Barnett concludes, “In a perfect marketing world, millions of individual and diverse personas need to be communicated with in effective and customized ways at scale. Our AI EQ helps us create authentic organic marketing, then slices it and dices it across the marketing mix to help our clients create economies of scale, unify and integrate their brands and, most importantly of all, capture their customers’ hearts.” About Rad Intelligence (RAD AI) Rad Intelligence (RAD AI) specializes in diversity and inclusion and has developed a marketing AI with emotional intelligence (EQ) that delivers a better way to create authentic, highly engaging, and inclusive influencer marketing programs. This results in unified and integrated marketing at economies of scale that help strengthen brands while generating the frequency needed to crush KPIs. Brands use RAD AI-informed content across the entire marketing mix, including paid advertising, blogs, emails, and brand-owned properties. Media ContactRAD IntelligenceJennifer [email protected] Source: RAD Intelligence