The Story of Body Odor Protection - Three ways to tell a Data Story

Good afternoon, good morning, or good evening, depending what time you are watching this. I am a senior scientist at Proctor and Gamble, and I am delighted to be here. Our mission in the company is to provide branded products and services of superior quality and value to improve the lives of our consumers. Our role in research and development is crucial in driving innovation and developing new products and technologies.

Today, I want to talk with you specifically about one aspect of R&D. I work closely with formulators, marketing and sales teams to discover consumer insights. One of the aspects of my job is to gain a deep understanding of behavior, preferences and unmet needs. This helps in identifying opportunities at developing products that resonate with consumers. In this session, I will show you some of the tools I use in the JMP platform to uncover these insights. We constantly do market and consumer research, and gather data to understand consumers shopping habits, products satisfaction and trends.

Some approaches we use are reading product reviews, emails or posts about our products, and other times we ask people to give us their opinion through surveys. JMP has a suite of tools that facilitate an analysis of non-structured data like text from product reviews and for structured data like those you get from scores on key performance indicators. I will show you some, but not all of the cool features for text analysis, categorical analysis and will share with you some machine learning techniques from JMP. I can also show you some in-depth exploration that I typically with my data.

Let me start with non-structured data analysis. Typical sources of non-structured data are online reviews from retailers like Amazon or Walmart. These reviews are an indicator of how well the products they purchase, meet or know their expectations. Other sources of non-structured data might be comments that consumers leave us after using our products. This type of data helps us answer questions like what are the consumers talking about? What are the different topics people mention on reviews with highest scores and what topics they mention on reviews with low scores.

Like all analysis, the overall flow of it should be cleaning and quality checking, exploring and then analyzing. To analyze non-structured data, first you need to import and clean the data in JMP. I recommend you to change the modeling type of your column of text to non-structured data. For any text analysis, this is how you do it in JMP.

First, open the file that has the data you want to analyze. When you import your data in jump, the preview is always useful to ensure that the column with data you want is there, so it gets imported. After you import your data and confirm the column with the texture, one is there. In my example, it is an Amazon reviews column. You then look at the column information and if needed, change the modeling type to instructor test as I mentioned. I recommend you saving your file once you have made those changes and before start cleaning any and exploring, so you always have your data as a jumping file. After importing the data, you can use the text Explorer suite to clean and start your exploration.

Let me do a demonstration. From the analysis menu, choose text Explorer and then place the column with text in the text columns space as I want to understand what are the things that people like and do not like. I will analyze by star rating. Then as these reviews are in English, I just want to cut my analysis to those words that have at least three characters. You can steam drop your menu or just do basic words for exploration. Depending on the experience and the knowledge that you have on the subject, JMP makes it easy to get insights immediately. The terms and phrases list output gives you an idea of what are the terms most mentioned on the analyzed text. Because I did my analysis by the star rating, I can easily see difference between the reviews with low scores and reviews with high scores. You can continue cleaning your data as you explore in JMP very easy. I like to use the word labs.

You can select the word that you want to exclude from the analysis and then click Add Stop Word on the menu for exploration. Or you can right-click after you have selected the word you want to eliminate and select Add Stop Word. Once you have spent some time cleaning and exploring your data, you can continue doing further analysis. As an example, in the deodorants category, I identify some themes of interest that were mentioned in low and high stat ratings. Let me show you how easily you can see the impact of identified topics with the text Explorer tool.

From the analyze text in the column space, you can choose to change the word of your cloud. The colony tool gives you the option to format your clock byte by volume on key columns. In this case, I wanted to analyze it by star ratings. I like to change the color on my graphic to see impact on my text on star ratings, green to see comments mentioned on high star ratings, and red to see comments seen on the lower star ratings. The redder or greener the word gets, the more likely this topic might be impacting my key performance indicator. To make it more friendly I like to understand, I can then format folder and use the layout options to polish my workload. Hovering over each word shows you not only the volume of that topic, but also any cleaning history you have done. You can then use this to peer-review and explain your analysis to others.

The other feature in JMP text analyzer I use often is the indicator column creation. You can easily select any topic of interest for future analysis and transform non-structured data into structured data. Let's explore structured data analysis next.

When I am interested in getting reactions to prototypes or ideas from people, I deploy surveys after consumers share their opinions. I then rely on the categorical platform in the consumer research analysis. Let me show you. It is an intuitive setup. Depending on the comparisons I wanted to make, I select the column and add to the structure tab top menu. Then, depending on the questions I have or the hypothesis I wanted to test, I select the variables to compare. In this example, I am interested in the key performance indicators across different stages of the use cycle.

I wanted to understand if there are differences between products when first seen on shelf, then after you set at home for the first time, and then after a few weeks, you can select multiple columns holding the Ctrl key when selecting columns and then put them on the side of the structured tab. The result? It is a very informative table. Let me show you what it does. For each product using the distribution of all the responses and the means and overall standard deviation. When I want to start summarizing my data, it's very easy to clean the table in JMP. Going to the red triangle, I can select the metrics I want to share with my team and peers. Since I'm interested on means and key responses, I can select the standard deviation and the response levels. You can see that now. I can easily see difference between some products on the means, but also on the distribution of consumer responses from initial to the end of the use.

Once I understand my data and draw my conclusions, I can further clean my table formatting the elements of the data I want to share. For example, I can use single decimal points for the distributions and the means report. When I am satisfied and want to share to others who might not use JMP. I can easily save my output table in Excel and share with others to make recommendations for the business based on data. As in the case of non-structured data, I do want to understand the relationship between not only my key metrics, but each element of performance we built into our products. This is important for us to understand as we make decisions for the business.

A good way to really understand what are the relationships between different attributes of a product. I rely on the multivariate analysis and JMP. The multivariate analysis allows you to see the correlation between various elements of performance and your key metrics. To do analysis, select multivariate methods from the multivariate analysis options the analyze platform. From there you can start selecting the variables you are interested to understand. In this example, I am interested in the relationship between some of my key performance indicators and all the attributes of the product across different elements. I do like to understand this relationship in a pairwise method, so the relationship between each pair is considered. I always use the default square matrix format.

The result is a very nice color-coded table showing you the strength of relationship between variables. Because human beings are not machines, I do look for record relations that are about 0.6. Obviously, the higher the relation to one, the better the relationship. Sometimes, depending on the stage of the blue moment, I consider different levels of confidence by changing my alpha level directly on the correlation table. For those on your team who are more visually learners, you can create a color map based on the data of your table. Just like in the text analysis, JMP is a very flexible tool that allows you to customize your data. I prefer to show a different color than red because in my organization red typically means a negative response. I can also change the lightness strength to increase the contrast of the strength of the correlations in my data. The darker the color, the higher the correlation. I can already see on this example that one of the attributes of performance doesn't correlate to other attributes, but it has a higher correlation to my key performance indicators.

At the same time, the color map shows that there are some attributes that have a high correlation to my KPI's. This step is very useful to gray a hypothesis, so you can start simplifying your data for more advanced analysis. Sometimes in the life of a project you need to focus on the most important factors that impact the overall score of your product or service, and you also want to understand if those factors are the same or not depending on the type of the product used or service.

Let me show you a quick way to start understanding what are the things that are predicting your response. From the analyze menu, go to the screener and choose predictor screen. Then choose the response you want to predict. In this case, I want to understand predictors of purchasing, so I put that in my quip response. Then I choose the variables of interest as predictors and put them on the quarks or predictables. In this case, I want to see all the elements that might be influencing the purchase intent. The output window shows all my predictors organized in order so that I know what is the biggest driver on my purchase intent. Not surprisingly, it looks like Intent is the biggest contributor, but the other factors have about the same impact and account for about 50% to drive purchase.

I wanted to understand further how these attributes might contribute across different products that I analyze. You can do a quick check easily in jump, but adding a local filter and toggle between different groups to see dynamically any difference on the contributions of each factor to your KPI. As you can see, the predictor variable changes between different products. It looks like for some products, being relevant might be a better predictor to purchase intent than value. This screening tool is very useful to help you start developing hypotheses and set priorities as you continue deep diving on your data analysis and do more sophisticated drivers which I will cover in the next slide.

So far, I have shown you a few techniques in JMP to analyze different types of data and highlighted how flexible JMP tools are to analyze both structured data and unstructured data. I'd like to show you more advanced tools JMP offers for more in-depth analysis. Please join me on October 23, from 4:45 PM to 5:15 PM at the Americas Summit, so I can show you how I leverage the fit model, so I can help guide the product development for the skin and personal care products.

Let me just finalize by telling you that at Procter and Gamble we care to improve all people's lives and understanding how our products delight all kinds of consumers, so we can customize our offerings depending on their habits and their needs. It's important. The partition modeling tool is an advanced tool in JMP, one of the many screening tools I use to start understanding consumer segments and who is delighted or not by your products and services. Here is a demonstration in how I use this tool.

From the start menu, select from the analyze predictive modeling and then partition. The window will ask you to add column variables in the response level and the predictor level. In this example, I want to understand who really enjoy the products in a consumer study, so I am going to select the overall rating as my response level. Then I want to use the habits that my users have and experience with any sweat or other breaks while using the product as my predictor. Because I want to validate my model, I want to call the portion for validation.

This initial window shows the distribution of responses across all 812 consumers in this story. The box on the left corner has the mean of the overall rating for the product and the standard deviation. I like sorting the predictor candidates because I can see what is the most likely prep behavior or experience users had on liking or not the product. In this case, it looks like those that are bothered by stains or body odor and who my exercise might have experienced the products differently. To do in-depth exploration, I can start partitioning the data or creating a key.

The first code shows that in fact, those who are typically bothered by stents rated the products better than those who are not. If explore what other characteristics these 86 people have, I see that in fact, eleven users have applied deodorant to other body parts of their body, and they really love the experience with the products tested. On the other hand, I want to understand what other characteristics might define users that like or didn't like the products that used and that are not water based. Doing a split among those 726 people shows that those who do not exercise did not care very much about the product, but those who do like them split imported. I learned that people who exercised often didn't enjoy the product.

Furthermore, using a product as syndicated as often is better than when consumers that exercise did not use a product very often. This simple yet in-depth analysis is a powerful tool to identify who benefits from our products and who might need some nudging to do different habits. In my role at Procter and Gamble, having JMP platform and using it to tell data stories with structured and non-structured information allows to inform my business and development partners how best we can serve our consumers.

At P&G, innovation is what we do best. We love find solutions to problems leveraging the power of JMP analysis tools, we identify problems to solve. We identify how solutions are doing, so we can then develop products of superior quality. Finding insights from consumers feedback using data analysis is one of the most rewarding jobs in research and development. Thank you so much for your help and attention. I hope to see you at the summit.