Introduction
Unstructured data related marketing problems
Consumer data
Competitor data
Firm strategy data
Data | Objectives | Marketers’ Challenges | How Beyond Text Helps |
---|---|---|---|
Consumer | Understand consumer trends Understand consumer decision processes Consumer profiling and targeting Relationship management Predict consumer behavior | Access to consumer data Consumer thoughts cannot be directly observed High cost and limited ability of useful data Gaining access to and monitoring consumer data from indirect sources | Provides access to unfiltered consumer commentary Data sources often free to access Data typically publicly available Consumers often share freely |
Competitor | Monitor competitor strategy Allow for real-time adjustment of marketing strategies | Competitor strategies are not directly observable | Can observe how consumers react to competitors Data is real-time Can analyze competitor advertising across formats |
Firm | Assess and optimize marketing effectiveness and efficiency | Hard to disentangle the impact of a marketer’s choices Hard to attribute success of marketing activity Integration of divergent data sources | Can test reactions from large numbers of consumers in real-time Link unstructured data to observable outcomes Can convert unstructured data to structured and compare structured data |
Cloud platforms
AWS | GCP | Azure | |
---|---|---|---|
Image And Video Classification | ✓ | ✓ | ✓ |
Image Detection | ✓ | ✓ | ✓ |
Object Detection (Video) | ✓ | ✓ | ✓ |
Speech To Text | ✓ | ✓ | ✓ |
Custom Model (Image and Audio) | ✓ | ✓ | ✓ |
Custom Model (Video) | x | ✓ | x |
Pre-Trained Model (Image, Video, Audio) | ✓ | ✓ | ✓ |
Image analysis
Understanding image analysis
-
Image classification categorizes similar images together. Automated analysis can even identify patterns that humans cannot detect and group based on these patterns.
-
Object detection involves spotting an image within a larger image. It incorporates localization; building a box around each object within an image, before classifying.
-
Supervised approaches use labeled pictures, e.g., cats/not cats. Pixel groupings recurring in labeled images, e.g., cats, are noted, and new images with similar features are given a label.
-
Unsupervised approaches group by common features, but algorithms do not look to group on any prior features, e.g., the computer ‘sees’ the images as similar but does not see them as cats. After the algorithm has created the groups, the researcher can post-hoc label the groups.
Using images for marketing tasks
-
How has our advertising changed over time?
-
Does our image use fit with what we said our target market was supposed to be?
-
Do our products look trendy or classic? Does this fit with what we are supposed to be aiming for?
Using the cloud platforms for image analysis
Cloud platform performance for images
Dataset
Dataset | # of Tags | Tag Description | # of Images | Source |
---|---|---|---|---|
Vehicle | 2 | Vehicle and non-vehicle | 17,760 | |
Apparel | 24 | Combinations of color* & clothing** | 11,385 | |
Real vs Fake Faces | 2 | Real and Fake | 2,041 |
Performance metrics
Image analysis results
Overall Precision | Overall Recall | |||||
---|---|---|---|---|---|---|
Dataset | AWS | GCP | Azure | AWS | GCP | Azure |
Vehicle | 100% | 99.80% | 99.90% | 100% | 99.80% | 99.90% |
Apparel | 97.90% | 97.6% | 97.60% | 98.50% | 97.5% | 97.40% |
(Real vs Fake) Faces | 66% | 97% | 66.40% | 86.50% | 98.50% | 66.40% |
What platform to use?
Platform | Advantages | Disadvantages |
---|---|---|
AWS | Object detection; Comprehensive free tier offering | Facial recognition |
GCP | Facial recognition | Limited free tier offering |
Azure | Single large project possible | Facial recognition |
Video analysis
Understanding video analysis
Using video for marketing tasks
Using the cloud platforms for video analysis
Cloud platform performance for video
Dataset
Dataset | # Tags | Tag Description | Videos | Source |
---|---|---|---|---|
Fight Scenes | 2 | Fight or Not fight | 200 | |
Advertisement 1 | 2 | Exciting or non-exciting | 300 | https://people.cs.pitt.edu/~kovashka/ads/ (Hussain et al., 2017 data) |
Advertisement 2 | 2 | Funny or not-funny | 284 | https://people.cs.pitt.edu/~kovashka/ads (Hussain et al., 2017 data) |
Video analysis results
Dataset | Platform | Overall Precision | Overall Recall | Tag Name | Confidence Threshold | Precision | Recall |
---|---|---|---|---|---|---|---|
Fight Scenes | GCP | 97.6% | 100% | Fight | 8% | 100% | 100% |
No Fight | 8% | 96.6% | 100% | ||||
Exciting | GCP | 68.9% | 70% | Exciting | 50% | 62.7% | 84% |
Not Exciting | 50% | 75.8% | 50% | ||||
Funny | GCP | 72.4% | 73.7% | Funny | 50% | 73.7% | 73% |
Not-Funny | 50% | 73.7% | 74% |
What platform to use?
Platform | Advantages | Disadvantages |
---|---|---|
AWS | More accessible for face-related features Free tier offering good for continuous use | Free tier offering not as good for limited intense tasks |
GCP | Custom model availability Classification availability Free tier offering good for limited intense tasks | Free tier offering not as good for continuous use |
Azure | Free tier offering good for limited intense tasks | Free tier offering not as good for continuous use |
Audio analysis
Understanding audio analysis
Using audio data for marketing tasks
Using the cloud platforms for audio transcription
Cloud platform performance for audio transcription
Dataset
Audio transcription analysis results
AWS | GCP, Sample Rate 48000 Hz | Azure | |
---|---|---|---|
Podcast | Confidence on word | Confidence In ~ 1 min Segment | Overall Confidence |
1 | High: 1.0, Low:0.255 | 0.96, 0.97, 0.97, 0.78, 0.93 | 0.83 |
2 | High: 1.0, Low:0.3453 | 0.96, 0.93, 0.97, 0.87 | 0.87 |
3 | High: 1.0, Low: 0.3383 | 0.97, 0.9, 0.97, 0.97, 0.96 | 0.88 |
4 | High: 1.0, Low: 0.3514 | 0.98, 0.95, 0.97, 0.97 | 0.91 |
5 | High: 1.0, Low: 0.1937 | 0.97, 0.95, 0.97, 0.87 | 0.91 |
What platform to use?
Platform | Advantages | Disadvantages |
---|---|---|
AWS | Using in specialized commercial settings | No textual analysis capacity |
GCP | Medical applications are strong | Free tier offering limited No textual analysis capacity |
Azure | Free tier offering most extensive Multiple audio uploading Speaker indexing Caption formatting Sentiment analysis | Limited domain knowledge |
Discussion and research agenda
Data Type | Questions That Could Create Substantive Contribution |
---|---|
Image | Can the images consumers post on social media predict sales? Can images be used as a leading indicator by practicing marketers? What else besides sales do the images predict, such as engagement? Do some images predict other metrics but not sales? Why? What does this tell us about consumer behavior? Can image mining be used to construct competitive maps? What technical barriers would be faced? How could managers be persuaded of the value of these? Would such maps be of aid in market definition during anti-trust investigations? Do these maps created using consumer posted images predict real-world switching behavior, e.g., closer competitors on the image map are more likely to experience switching between them? Do some types of images work better than others? How can marketers stimulate the posting of specific image types? |
Video | What methods can academics devise to measure the addictiveness of video media, a key consumer behavior question? What factors in video can explain the addictiveness of such media? What are marketers doing to abet, or reduce, addiction? How can firms use these findings ethically? How can young consumers be protected in an effective yet cost-feasible manner? How can consumers understand and manage their own reactions to addictive video? How can parents help their children? Precisely what sort of video is most likely to lead to positive business results, e.g., sales? As a practical matter, how can such insights be integrated with a creative process? Can academics use these insights to understand consumer behavior and provide theoretical explanations for any finding? |
Audio | Can consumer’s vocal tones in call center recordings predict defection? If so, how can vocal tones be used by managers as an early warning system? As a practical matter, what actions can managers take given such early warnings? What tasks would most benefit from transcription services? How will the presence of the transcription services impact academic fields and market research? Will the datasets from transcriptions be made available to fellow academics for replication? If scholars extend other academic’s findings how will this be viewed by journals? What predictive power is lost when audio is transcribed? I.e., what can be predicted from vocal tones that is not predictable from the transcription of the words used? How can we measure loss of meaning through transcription? What types of audio, and from what fields, e.g., sales calls versus in-person service, lose the most meaning when words are conveyed as transcriptions versus listened to? What impact will transcription and translation services have for international business scholarship? How will their quality be assessed and reported? |
Comparison | What is the relative power of the various beyond text areas? Does, for example, product placement in a video generate significantly more impact than placement in a static image? What theory can explain any difference? Given differential costs between audio, image, and video marketing what are the most fruitful avenues for those with limited budgets? How does the effectiveness of beyond text media interact with product and industry type? For what situations do the precise words used matter more than the accompanying video and vice versa? Does TV advertising rely more on soundtrack than an online video? What works in a static image but not in a video and vice versa? Why? |
Integrated Data Sources | How can the training of sales forces and other client facing staff be aided through use of video and audio analysis? How can such training recognize customer heterogeneity and avoid a one-size-fits-all approach? How would the use of the platforms in B2B differ from traditional consumer facing operations? How could a manufacturer use beyond text insights to persuade a retailer to carry their products? What constraints should be put on the marketer’s use of consumer data to protect the privacy of the consumer? How do, or even should, privacy concerns differ between text and beyond text data and within the various beyond text categories? How can marketers help protect the privacy of third parties (background characters) in beyond text data? What actions should regulators take? What is the value of creative? If structure can be created to describe creative, how can we use that to predict market outcomes? What is the difference between the outcomes we observe using various types of creative? Can we create a method to measure the broader idea of creativity in marketing using beyond text techniques? Can we create measures of creativity? Do these relate to meaningful commercial outcomes? The Marketing Accountability Standards Board (MASB) have launched a Measuring Creativity Initiative (MASB, 2023). How can beyond text data help power such initiatives? |