preference mapping for automated recommendation of product attributes for designing marketing...

21
© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential. Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content Moumita Sinha, Rishiraj Saha Roy Adobe Research Labs

Upload: rishiraj-saha-roy

Post on 17-Jul-2015

265 views

Category:

Software


2 download

TRANSCRIPT

Page 1: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing ContentMoumita Sinha, Rishiraj Saha Roy

Adobe Research Labs

Page 2: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Overview

Introduction

Method

Dataset

Results

Conclusions

2

Page 3: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Introduction

Page 4: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Introduction

Marketers need to highlight attributes of their products in campaigns

Preference Mapping is an approach to identify customer preferences based on surveys of product attributes

Using these preferences, recommendations of product attributes can be provided to marketers to design their campaigns

4

Page 5: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Workflow

5

Presenter
Presentation Notes
Boxes in Green are: Part of Workflow; Boxes in Blue: Proposed Algorithm
Page 6: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Method

Page 7: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Method

Consider 𝑘𝑘 products each with 𝑝𝑝 attributes

Example 𝑘𝑘 = 4 products with each product having 𝑝𝑝 = 13 attributes

One of these is the product for which campaign is being designed

Customers go to the product or retailer website and write textual reviews

These reviews are accompanied with positive, neutral or negative sentiments about the various attributes of the products

A preference mapping is then performed with the customer averaged scores of each of the various attributes for the different products

7

Presenter
Presentation Notes
P is the number of attributes and k is the number of products. The k and p have been interchanged in the paper : typo Sentiment score of each sentence in each review has been assigned with Alchemy API. We have chosen review sentences where only one attribute is discussed. Thus sentiment score of the sentence is assumed to be the sentiment score of the attribute. Here a reviewer is simply a customer who has written review(s).
Page 8: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Preference Mapping

Matrix of Reviewer Averaged Scores for 𝑝𝑝 attributes:

𝑋𝑋 = 𝑋𝑋1,𝑋𝑋2, …𝑋𝑋𝑝𝑝𝑇𝑇

𝑋𝑋𝑖𝑖 is a vector with elements 𝑋𝑋𝑖𝑖𝑗𝑗 Reviewer averaged sentiment score for attribute 𝑖𝑖 and product 𝑗𝑗

Principal Component transformation of feature vector

𝑌𝑌 = Γ𝑇𝑇 𝑋𝑋 − 𝜇𝜇 such that 𝜇𝜇 = 𝐸𝐸 𝑋𝑋 Γ ΛΓ𝑇𝑇 = 𝑉𝑉𝑉𝑉 𝑟𝑟 𝑋𝑋

This transformation is such that 𝑉𝑉𝑉𝑉 𝑟𝑟 𝑌𝑌 is maximized

𝜆𝜆1 ≥ 𝜆𝜆2 ≥ … .≥ 𝜆𝜆𝑝𝑝 where 𝜆𝜆𝑗𝑗 = 𝑉𝑉𝑉𝑉𝑟𝑟(𝑌𝑌𝑗𝑗)

8

Presenter
Presentation Notes
Lambda is the eigenvalue matrix of Var(X)
Page 9: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Preference Mapping

𝜆𝜆𝑗𝑗 ’s are eigenvalues (Components of Λ)

Corresponding eigenvectors are: 𝛾𝛾1, 𝛾𝛾2 … 𝛾𝛾𝑝𝑝 (Components of Γ)

Thus 𝑖𝑖𝑡𝑡𝑡 Principal Component (PCi) Score for each product is the weighted sum of the score of the attributes for the

product and the weight being 𝑖𝑖𝑡𝑡𝑡 eigenvector: Y𝑗𝑗 = ∑𝑚𝑚=1𝑝𝑝 𝛾𝛾𝑖𝑖𝑚𝑚𝑋𝑋𝑚𝑚𝑗𝑗

A biplot graph from PC1 and PC2 provides easily interpretable visualization

Shows how products compare among each other

Relative proximity of each attribute to their respective products

Marketing contents can be designed based on the recommendation of this multivariate approach and its visualization

10

Page 10: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Dataset

Page 11: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Dataset

Primary Dataset

Products: 4 cameras

Attributes for each Camera: 13 E.g.: Flash, zoom, battery, quality of automatic mode, photo quality

These attributes are mentioned 583 times in the review data collected

Expert Ratings (To compare results)

Collected from http://www.dcresource.com and http://www.imaging-resource.com

For the same thirteen attributes in 4 cameras

Comments with “exceptional”, “excellent” and “good” for an attribute Score 2

Comments with “weak” and “worst” for an attribute Score 1

12

Page 12: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Results

Page 13: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Averaged Sentiment Scores for the Four Cameras

14

Presenter
Presentation Notes
The sentiment scores from various customer reviews are averaged for each attribute of each product. Then they are scaled and plotted in a radial chart. Here we can see for each product the differing importance of the attributes, based on sentiment scores. This does not give an overall comparison (multivariate account) of the products taken into account all the attributes. We can see that the products vary from each other based on the attributes but what is the comparative score of the products, is not available. The sentiment scores are collected from review statements that have only one attribute mentioned and it is assumed that the sentiment of the statement is the sentiment for the attribute.
Page 14: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Preference Mapping

15

Presenter
Presentation Notes
The PC1 and PC2 together cover 85% of the variability, which calculated from the lambda1 and lambda2. The scores have been scaled before using PCA. The arrows are the eigenvector values for each attribute. The products are the PC score: Y_j There are 4 quadrants and closer the attribute is to a product, the higher positive sentiments customers have for that attribute of that product. Also, we can compare the products simultaneously with respect to the attributes, which was not possible from the radial chart. Thus, Canon Powershot and Nikon have high values for zoom, photo quality etc, as compared to to CanonS100. Canon S100 has a higher sentiment for quality of the color, as compared to Nikon and Canon Powershot.
Page 15: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Correlation: Preference Mapping Scores Vs Rank of attributes Based on Average Sentiment Scores

Camera Kendall-Tau P-Value

Canon G3 0.564 0.007

Canon S100 0.615 0.003

Canon Powershot SD 500

0.641 0.002

Nikon Coolpix 4300 0.294 0.172

Preference mapping has high correlation with intuitive understanding of importance of attributes

Provides further refinement : Provides multivariate relation between products with respect to each attribute

No direct relation between expert ratings and preference mapping score was observed

16

Presenter
Presentation Notes
The averaged scores, as used in the Radial chart are the Xs and the Euclidean distance between the attribute and the products (as shown in the biplot) are the Ys as used to calculate the Kendall-Tau correlation
Page 16: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Conclusions

Page 17: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Conclusions

Preference mapping technique recommends valuable attributes of products to marketers for highlighting in a marketing campaign

By focusing on attributes that are known to have received positive sentiments of customers, the risk in the campaign is minimized

Can potentially increase response rates

18

Page 18: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Conclusions

The proposed technology does not require large amounts of customer preference data to be available internally with the advertiser

Reviews can be collected from any publicly available review site

The comparison with the experts' comments suggests:

What customers value more may be different from what experts consider of high quality in a product

19

Page 19: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Future Work

Page 20: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

Future Work

Multiple factor analysis (MFA) instead of PCA if some or all scores are of categorical nature

Cluster products using attribute sentiment scores as features

Observe the correlation of the clustering output to the representation produced by preference mapping

Quality of reviews can be improved by choosing active users only

21

Page 21: Preference Mapping for Automated Recommendation of Product Attributes for Designing Marketing Content

© 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.

For further questions, the authors can be contacted at:

[email protected]

[email protected]

22