AI for Marketers  ## Metadata - Author: [[Christopher Penn]] | [[3. Resources/Authors/Christopher Penn|Christopher Penn]] - Full Title: AI for Marketers - Category: #books - Kindle: [Open in Kindle](https://www.amazon.com/dp/B08Y82R149) - Document Tags: [[Marketing]], [[AI]] ## Document Notes ## Highlights - AI is the gatekeeper to the eyes and ears of our audience, in the form of filters, recommendation engines, and algorithms prioritizing some content over others. ([Location 191](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=191)) - According to the Merriam-Webster dictionary, AI is: 1 : a branch of computer science dealing with the simulation of intelligent behavior in computers 2 : the capability of a machine to imitate intelligent human behavior ([Location 260](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=260)) - “Machine learning (ML) is the scientific study of algorithms and statistical models that computer systems use to progressively improve their performance on a specific task. Machine learning algorithms build a mathematical model of sample data, known as ‘training data’, in order to make predictions or decisions without being explicitly programmed to perform the task.” ([Location 276](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=276)) - All algorithms used in machine learning and artificial intelligence are fundamentally rooted in statistics and mathematics, from the simplest predictive algorithms like linear regression - such as determining if one variable changes proportionally with another variable - to the most complex, sophisticated neural networks. ([Location 343](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=343)) - In machine learning, you’ll hear these two types of data referred to as continuous (numeric) or categorical (non-numeric). Many types of data have blends of both; ([Location 371](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=371)) - What are some common marketing tasks that are ideally suited for unsupervised machine learning? Sorting images and videos Understanding social media conversations about us, competitors, or our industry Analyzing company coverage to determine brand strength Clustering and sorting customers to find our best customers Digesting the day’s headlines to find trending topics to write about ([Location 402](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=402)) - When we know the outcome and we’re working with continuous data, we’re using regression and prediction methods. The key questions marketers are most likely to ask which would be solved by regression and prediction include: What contributing variables led to this outcome, such as increased or decreased ROI? What was the sentiment of our social media coverage in the last year? Who is most influential in our industry? What’s likely to happen in the next 12 months in our search traffic? ([Location 421](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=421)) - When we know the outcome and we’re working with categorical data, we’re using classification methods. The key questions marketers are most likely to ask which would be solved by classification and categorization methods include: What topics are most prominent in news coverage about us? What have customers mentioned in their calls to our call center? What are our top influencers talking about overall? What are the subjects of the top million photos on Instagram? What are the major job titles of our customers in our CRM? ([Location 428](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=428)) - When we don’t know the outcome and we’re working with continuous data, we’re using clustering methods. The key questions marketers are most likely to ask which would be solved by clustering include: What do our most profitable customers have in common? When customers buy a certain product, what else are they likely to buy? How frequently do negative complaints co-occur? What combinations of metrics lead to our best conversions? ([Location 435](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=435)) - Clustering helps us make order out of disorderly numeric data, showing patterns where we might not otherwise be able to see them. ([Location 440](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=440)) - Good data should be: • Clean • Complete • Comprehensive • Chosen • Credible • Calculable ([Location 514](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=514)) - Clean. Clean data is data that is prepared and free of errors, free of “dirt”. Dirty data would be data with anomalies, or contaminated. ([Location 520](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=520)) - Complete. Complete data has no critical missing information. Some machine learning algorithms do very poorly or outright break if data is missing, so a major part of preparing data for machine learning includes ensuring its completeness. ([Location 525](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=525)) - Comprehensive. One of the biggest challenges of many datasets is that they answer a narrow question but may not answer bigger picture questions that are logical next steps, so data must be comprehensive. ([Location 530](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=530)) - Chosen. Just as completeness mandates we have enough data, chosen data mandates that we not have too much data. ([Location 534](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=534)) - Credible. The process of collecting the data is just as important as the data itself. How credible is the data? What’s its lineage, its provenance? Was it collected in a manner that is free of bias? ([Location 540](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=540)) - Calculable. Finally, data must be in a calculable, usable format that’s accessible to machines and business users alike. Data stored in an obscure or difficult format is less likely to be used than data that is readily accessible or transformable for non-data scientists. ([Location 545](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=545)) - dimension reduction, and it uses the power of statistics to help us remove stuff that simply doesn’t matter. ([Location 665](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=665)) - feature selection, which is part of running advanced regression analysis. ([Location 667](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=667)) - Turning unstructured data into structured data and then quantitative, numeric data is extremely difficult for most marketers. ([Location 729](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=729)) - technology to unlock the data in each of these different formats may not be readily accessible to everyone - yet. ([Location 736](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=736)) - The challenge for us as marketers is bringing order and structure to these very disorderly interactions so that we can make informed choices about where to focus our time and effort. ([Location 909](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=909)) - A machine learning technology known as network graphing does exactly this. Given a list of interactions, network graphing helps draw the literal network connecting people together. Here’s a simplified example of how it works. ([Location 912](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=912)) - This means that instead of being a cost center, companies which have digitally transformed have figured out how to make money on their data, the analysis of their data, and critically, machine learning models which are built on their data. ([Location 990](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=990)) - This is what it means to be digitally transformed: you’ve figured out how to build new products and services based on your data. ([Location 997](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=997)) - What’s paramount to machine learning, and to our journey to AI, is that we be able to speak to our data capabilities thoroughly. In the technology world, this is known as systems architecture, and it’s something marketers have not had to learn until now. ([Location 1032](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1032)) - Who has responsibility for the data? What data do we have within our organization? Where does all our data live? When was our data collected? Why do we have or lack certain data? How do we collect and store data? ([Location 1035](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1035)) - The more complex the data, the more difficult it is to answer why, unless you have qualitative data capabilities, which can include: Customer advisory boards Focus groups Surveys Interviews Shop-alongs In-place video observation Lifestyle immersion Journals/diaries ([Location 1066](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1066)) - What is process automation? It’s exactly as it sounds in name - a concerted effort to automate as many tasks or processes as possible to find efficiencies and free up time and resources in your organization. ([Location 1079](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1079)) - Process automation always begins with process documentation and governance. We can’t automate what we don’t know. By documenting processes using many of the same methods as used in qualitative research, we build a deep understanding of how work is done within our organization. ([Location 1081](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1081)) - “Data science is an interdisciplinary field that uses scientific methods, processes, algorithms and systems to extract knowledge and insights from data in various forms, both structured and unstructured, similar to data mining.” ([Location 1117](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1117)) - Data architects are the people who work alongside the developers to capture the raw data as they extract it from systems and store it somewhere for further use. ([Location 1210](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1210)) - data scientist. We’ll need data scientists to refine and transform the data into data products: analysis, insights, models, dashboards, visualizations, etc. - everything we’d need to make use of the data. ([Location 1214](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1214)) - marketing technologist, the person who speaks both the language of data products and the language of marketing, who takes the data products and helps business users make the most of them. ([Location 1219](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1219)) - a blueprint for how all the components and pieces of the journey to AI are monitored, measured, and managed. ([Location 1232](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1232)) - In the governance of marketing technology, and more specifically within machine learning, we audit to understand how well aligned people, processes, and technologies are. If things are out of alignment, our journey to AI will be slowed or may stall entirely. ([Location 1317](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1317)) - Machines can simulate the language of empathy, but empathy requires you to understand and share someone else’s emotions. ([Location 1373](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1373)) - Second, machine learning is incapable of judgement. By judgement, we mean examples and situations where we step outside the rules. ([Location 1384](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1384)) - machine learning is incapable of emulating or assessing broad life experience. ([Location 1396](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1396)) - machines generally are a poor substitute for human-to-human interactions. For the most part, all things being equal, people like to interact with people most of the time. ([Location 1408](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1408)) - someone who’s good at cloud computing AND sales AND business analysis is going to be very difficult to automate. ([Location 1462](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1462)) - If we expand our capabilities as professionals across disciplines and areas of study, we significantly reduce the likelihood that part or all of our roles can be automated out of existence. ([Location 1464](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1464)) - human oversight of the machines and their outputs. ([Location 1553](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1553)) - human bias, according to the Oxford English Dictionary, is “prejudice in favor of or against something compared to another, usually in a way considered to be unfair”. ([Location 1605](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1605)) - statistical bias. “A statistic is biased if it is calculated in such a way that it is systematically different from the population parameter being estimated”. ([Location 1609](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1609)) - Intentional bias is bias in a model introduced intentionally, by design. When someone creates an algorithm, they introduce a specific outcome they want to achieve, and the machine learning software creates models which generate the intended result. ([Location 1627](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1627)) - source/sample bias. This is a bias that stems from inaccurate data being fed to the machine learning model. ([Location 1640](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1640)) - Marketers most often make source/sample bias errors by not using data that’s representative of the population as a whole. ([Location 1646](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1646)) - third type of unwanted bias is target bias. This is a bias caused by macro conditions which make all data about a particular target population unreliable. In some ways, it’s similar to source/sample bias, but because it’s so much bigger a problem, there may not be ways to clean or compensate for the bias in our data. ([Location 1651](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1651)) - tool or instrument bias. This is a bias caused by limitations in our measurement tools and instruments. When a tool or instrument is unable to capture all the data needed to build a model, it can create unintentional biases or distortions. ([Location 1668](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1668)) - Explainability is what we do with a model’s results - we attempt to explain them after the model has run. Many prominent machine learning and AI companies focus heavily on explainability because their models are so gigantic and complex that it may be the only feasible way for them to demonstrate that a model isn’t biased. ([Location 1815](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1815)) - interpretability. This is when we take a model apart, reducing it down to its source code, down to its individual components, and then walk through how the model functions at each step, so that we can show why a particular step is or is not biased. ([Location 1819](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1819)) - For technical people, interpretability is the standard we should be requiring - the ability to step through every piece of a model and show how it prevents or mitigates bias. ([Location 1829](https://readwise.io/to_kindle?action=open&asin=B08Y82R149&location=1829)) ## Related by Author - [[3. Resources/Readwise/Books/Marketing Blue Belt|Marketing Blue Belt]] - [[3. Resources/Readwise/Books/Marketing Red Belt|Marketing Red Belt]] ## Related - [[3. Resources/Readwise/Articles/Almost Timely News Protecting Your AI Prompts|Almost Timely News Protecting Your AI Prompts]] - [[3. Resources/Readwise/Articles/Almost Timely News Resistance to AI|Almost Timely News Resistance to AI]] - [[3. Resources/Readwise/Articles/Almost Timely News The Importance of Open Source in AI|Almost Timely News The Importance of Open Source in AI]] - [[3. Resources/Readwise/Articles/Almost Timely News The Intelligence Revolution|Almost Timely News The Intelligence Revolution]] - [[3. Resources/Readwise/Articles/Almost Timely News What ChatGPT Is Really Good At, Measurement Strategies for Agencies|Almost Timely News What ChatGPT Is Really Good At, Measurement Strategies for Agencies]] - [[3. Resources/Readwise/Articles/Almost Timely News When Should You Use Generative AI|Almost Timely News When Should You Use Generative AI]] - [[3. Resources/Readwise/Articles/Almost Timely News, April 30, 2023 A Marketing Antidote for Large Language Models|Almost Timely News, April 30, 2023 A Marketing Antidote for Large Language Models]] - [[3. Resources/Readwise/Articles/Almost Timely News, July 23, 2023 AI, South Park, and LLaMas|Almost Timely News, July 23, 2023 AI, South Park, and LLaMas]] - [[3. Resources/Readwise/Articles/Mind Readings The Future of Social Listening in the Age of AI|Mind Readings The Future of Social Listening in the Age of AI]] - [[3. Resources/Readwise/Articles/Mind Readings What Makes a Good ConferenceEvent|Mind Readings What Makes a Good ConferenceEvent]] - [[3. Resources/Readwise/Articles/So What How to Do Prompt Engineering|So What How to Do Prompt Engineering]] - [[3. Resources/Readwise/Articles/You Ask, I Answer Ecological Impact of AI|You Ask, I Answer Ecological Impact of AI]] - [[3. Resources/Readwise/Articles/You Ask, I Answer Why Do Open Source Large Language Models Matter|You Ask, I Answer Why Do Open Source Large Language Models Matter]]