Tutorial presented by Muthusamy Chelliah (Flipkart, India) and Sudeshna Sarkar (IIT Kharagpur, India) at ACM RecSys 2017 https://recsys.acm.org/recsys17/tutorials/#content-tab-1-3-tab
E-commerce websites commonly deploy recommender systems that make use of user activity (e.g., ratings, views, and purchases) or content (product descriptions). These recommender systems can benefit enormously by also exploiting the information contained in customer reviews. Reviews capture the experience of multiple customers with diverse preferences, often on the fine-grained level of specific features of products. Reviews can also identify consumers’ preferences for product features and provide helpful explanations. The usefulness of reviews is evidenced by the prevalence of their use by customers to support shopping decisions online. With the appropriate techniques, recommender systems can benefit directly from user reviews.
This tutorial will present a range of techniques that allow recommender systems in e-commerce websites to take full advantage of reviews. Topics covered include text mining methods for feature-specific sentiment analysis of products, topic models and distributed representations that bridge the vocabulary gap between user reviews and product descriptions, and recommender algorithms that use review information to address the cold-start problem.
The tutorial sessions will be interspersed with examples from an online marketplace (i.e., Flipkart) and experience with using data mining and Natural Language Processing techniques (e.g., matrix factorization, LDA, word embeddings) from Web-scale systems.
Movie recommendation system using collaborative filtering system Mauryasuraj98
The document describes a mini project on building a movie recommendation system. It includes an abstract that discusses different recommendation approaches like demographic, content-based, and collaborative filtering. It also outlines the problem statement, proposed solution, workflow, dataset description, algorithm details, GUI design, result analysis, and applications. The system uses a user-based collaborative filtering model to recommend movies to users based on their preferences and ratings of similar users. Evaluation shows it has good prediction performance.
This document describes a movie recommendation system that uses machine learning techniques like cosine similarity and TF-IDF. It discusses collecting movie data, preprocessing it using techniques like TF-IDF to generate feature vectors, and then calculating cosine similarity between movies to find similar movies and make recommendations. The system was developed in Python using libraries like NumPy, Pandas, and Matplotlib. It demonstrates generating recommendations based on both movie genres and titles and achieves good results. Pseudocode is also provided to explain the technical approach.
Recommender system algorithm and architectureLiang Xiang
1) The document discusses recommender system algorithms and architecture. It covers common recommendation techniques like collaborative filtering, content-based filtering, and graph-based recommendations.
2) It also discusses challenges like cold starts for new users and items. For new users, it recommends using demographic data or initial feedback to understand interests. For new items, it suggests using content information or initial user feedback.
3) The document proposes a feature-based recommendation framework that connects users, items, and latent features to address challenges like heterogeneous data and cold starts. This framework provides explanations but does not support user-based methods.
The document provides an overview of recommender systems. It discusses the typical architecture of recommender systems and describes three main types: collaborative filtering systems, content-based systems, and knowledge-based systems. It also covers paradigms like collaborative filtering, content-based, knowledge-based, and hybrid recommender systems. The document then focuses on collaborative filtering techniques like user-based nearest neighbor collaborative filtering and item-based collaborative filtering. It also discusses latent factor models, matrix factorization approaches, and context-based recommender systems.
In this lecture, I will first cover the recent advances in neural recommender systems such as autoencoder-based and MLP-based recommender systems. Then, I will introduce the recent achievement for automatic playlist continuation in music recommendation.
Collaborative filtering is a technique used in recommender systems to predict a user's preferences based on other similar users' preferences. It involves collecting ratings data from users, calculating similarities between users or items, and making recommendations. Common approaches include user-user collaborative filtering, item-item collaborative filtering, and probabilistic matrix factorization. Recommender systems are evaluated both offline using metrics like MAE and RMSE, and through online user testing.
Recommender systems using collaborative filteringD Yogendra Rao
This document summarizes a student project on implementing recommender systems. The project objectives were to design a website using user-based, item-based, and model-based collaborative filtering as well as MapReduce to generate movie recommendations. The system was tested on the MovieLens dataset using MAE and RMSE metrics, with user-based filtering found to have the best performance. The document outlines the technical aspects of the recommendation system including the technologies used, website architecture, and references.
This document discusses recommender engines, which are systems that predict items a user may be interested in based on their preferences and behaviors. It describes several common recommendation techniques, including demographic filtering, content-based filtering, user-based collaborative filtering, and item-based collaborative filtering. Examples of recommender engines used by Amazon and Digg are provided to illustrate how these techniques are implemented on e-commerce and social news sites. The document concludes that recommender engines provide benefits to both businesses and users by enabling personalized recommendations at scale.
Movie Recommender System Using Artificial Intelligence Shrutika Oswal
In recent years, a huge amount of information is available on the internet and it is very difficult for the user to collect the relevant information. While purchasing any product also a lot of choices available and the user is confused about what to choose. This will be a time-consuming process as well. The search engine will solve this problem to some extent by but it will fail in giving a personalized recommendation. In this presentation, I will describe the different types and working of the recommender system how they gather the data, build recommender, generate recommendations from it, evaluate the performance and effectiveness of the recommender system. The further part of the presentation will describe how to build a movie recommender system using python.
This document discusses recommender systems, including:
1. It provides an overview of recommender systems, their history, and common problems like top-N recommendation and rating prediction.
2. It then discusses what makes a good recommender system, including experiment methods like offline, user surveys, and online experiments, as well as evaluation metrics like prediction accuracy, diversity, novelty, and user satisfaction.
3. Key metrics that are important to evaluate recommender systems are discussed, such as user satisfaction, prediction accuracy, coverage, diversity, novelty, serendipity, trust, robustness, and response time. The document emphasizes selecting metrics based on business goals.
Recommender systems are software agents that analyze a user's preferences through transactions and provide personalized recommendations accordingly. There are several recommendation paradigms including non-personalized rules, personalized rules based on user data, and transaction-based collaborative filtering that learns from user interactions. Context-based recommender systems also consider additional information like time, location, or device to provide adaptive recommendations. Common techniques used in recommender systems include content-based filtering that recommends similar items, collaborative filtering that finds users with similar tastes, and demographic-based recommendations.
Recommendation systems provide users with information they may be interested in based on their preferences and interests. They help address the problem of information overload by retrieving desired information for the user based on their preferences or those of similar users. The two main types of recommendation systems are personalized and non-personalized systems. Common techniques used include collaborative filtering, which finds users with similar tastes, and content-based filtering, which recommends items similar to those a user has liked based on item attributes.
This presentation consist of detail description regarding how social media sentiments analysis is performed , what is its scope and benefits in real life scenario.
The document discusses recommender systems and describes several techniques used in collaborative filtering recommender systems including k-nearest neighbors (kNN), singular value decomposition (SVD), and similarity weights optimization (SWO). It provides examples of how these techniques work and compares kNN to SWO. The document aims to explain state-of-the-art recommender system methods.
This document discusses building a recommendation system for e-commerce. It begins by noting the importance of recommendations, with over 30% of online purchases coming from recommendations. It then discusses gathering data, both explicitly via ratings and reviews, and implicitly via user actions. Main approaches covered include content-based filtering, collaborative filtering using user-user and item-item similarities, and matrix factorization. The document also addresses challenges like sparsity, cold starts, scalability and privacy considerations in implementing recommendation systems.
Recommender Systems represent one of the most widespread and impactful applications of predictive machine learning models.
Amazon, YouTube, Netflix, Facebook and many other companies generate an important fraction of their revenues thanks to their ability to model and accurately predict users ratings and preferences.
In this presentation we cover the following points:
→ introduction to recommender systems
→ working with explicit vs implicit feedback
→ content-based vs collaborative filtering approaches
→ user-based and item-item methods
→ machine learning and deep learning models
→ pros & cons of the methods: scalability, accuracy, explainability
The document discusses recommendation algorithms used by Amazon, including traditional collaborative filtering, cluster models, and search-based methods. It focuses on Amazon's item-to-item collaborative filtering algorithm. This algorithm builds a similar-items table offline by finding items customers tend to purchase together. It then scales well to large data sets, provides high quality recommendations even with limited user data, and performs recommendations quickly.
This document proposes a reinforcement learning framework for explainable recommendations. The framework is model-agnostic, reveals the working mechanism of recommendation models to improve explainability, and can control the quality of explanations. It uses coupled agents with personalized attention networks to select interpretable components for explanations. The agents are trained with policy gradients to maximize rewards for model explainability and presentation quality. The framework was evaluated offline on different models and settings, and with human subjects in a restaurant recommendation task.
Recommending the Appropriate Products for target user in E-commerce using SBT...IRJET Journal
The document proposes a new recommendation system for e-commerce that uses structural balance theory to recommend products when target users have no similar friends or similar product preferences. It discusses limitations of existing recommendation approaches. The proposed system first identifies a target user's "dissimilar enemies" and then determines their "possible friends" according to structural balance theory. Products liked by these "possible friends" are recommended. It aims to leverage structural balance information in user-product networks for more accurate recommendations when collaborative filtering cannot find similar users or items.
This document outlines a movie recommendation system project built using collaborative filtering. The project aims to build a recommendation engine that suggests movies to users based on their preferences and watching history. It will use the MovieLens dataset and implement item-based collaborative filtering. The key steps include importing libraries, preprocessing the data, building the recommendation model using collaborative filtering, and evaluating the model's performance. Collaborative filtering works by comparing a user's preferences to other users to find patterns and provide personalized recommendations. The document also discusses some disadvantages of collaborative filtering like the cold-start problem and difficulty including additional metadata.
This document discusses various approaches for designing effective preference elicitation systems for recommendation engines. It covers challenges like the cold start problem and how to ask users questions to understand their preferences. It also examines different types of interfaces, factors that influence user opinions, and strategies for choosing representative examples to elicit preferences efficiently and accurately. The document concludes with discussing evaluation metrics and opportunities for future work.
[UPDATE] Udacity webinar on Recommendation SystemsAxel de Romblay
A 1h webinar on RecSys for the Udacity NanoDegree Program "How to become a Data Scientist" : https://in.udacity.com/course/data-scientist-nanodegree--nd025.
The link to the ipynb : https://www.kaggle.com/axelderomblay/udacity-workshop-on-recommendation-systems
This document provides an overview of recommendation systems including: how they work using content-based, collaborative filtering, and hybrid approaches; challenges like performance, modelization, and biases; common evaluation methods including offline metrics, A/B testing, and the online-offline gap; interactive learning using reinforcement learning; and concludes with best practices like prioritizing criteria, blending approaches, and continuous evaluation and monitoring. A hands-on tutorial is also proposed to implement a neural network-based recommendation system.
This document discusses recommendation systems and how to develop them. It begins by introducing the speaker and an overview of the topics to be covered. It then explains what recommendation systems are and different types including search, content-based, and collaborative filtering. It discusses drawbacks like cold start problems and sparsity and ways to address them. The document concludes with tips for refining recommendation models like normalization, capturing trends, and temporal factors.
Modern Perspectives on Recommender Systems and their Applications in MendeleyKris Jack
This document discusses modern perspectives on recommender systems and their applications at Mendeley. It covers the evolution of recommender problems from rating prediction to context-aware recommendations. It also discusses common recommender algorithms like collaborative filtering, content-based filtering and hybrid approaches. The document concludes by discussing how Mendeley uses recommenders for related research, researchers to follow, and other use cases.
The document outlines the schedule and activities for a two-day paper prototyping workshop. On the first day, participants will be introduced to paper prototyping techniques and evaluate example prototypes. They will do initial prototyping activities and further refine their prototypes. On the second day, prototyping activities continue with opportunities for feedback from others. Evaluation of paper prototypes is discussed, focusing on usefulness, usability, and user experience. Inspection methods like cognitive walkthroughs and heuristic evaluation are described along with testing methods like co-discovery and wizard of oz techniques. The importance of piloting evaluation protocols is also highlighted.
Summarization and opinion detection in product reviewspapanaboinasuman
This document describes a project to build a system that provides structured summaries of product reviews by extracting product features and associated opinions. It outlines the end-to-end architecture of the system, including modules for crawling reviews, preprocessing text, extracting and analyzing features and opinions, and providing a feature-based summary. An evaluation of the system shows a precision of 75% and recall of 90% for correctly identifying features and opinions.
Recommendations are everywhere : music, movies, books, social medias, e-commerce web sites… The Web is leaving the era of search and entering one of discovery. This quick introduction will help you to understand this vast topic and why you should use it.
The document discusses effective Scrum teams and provides information on Scrum roles and practices. It describes the Product Owner, Scrum Master, and Scrum Team roles and their responsibilities. It also covers topics like user stories, estimating work, themes vs epics, and characteristics of good user stories using the INVEST mnemonic. The document aims to help readers understand how to effectively structure work in Scrum.
Design of recommender system based on customer reviewseSAT Journals
Abstract Recommendations play a significant role in every human life. People choose their ideas based on other’s recommendations since they trust the recommendations more. For giving recommendations there emerged a system called Recommender system. Recommender systems play a important role in E-Marketing. Many companies adopt recommender systems to increase in their sales in the market. They can establish their products such that they can attract more customers by giving offers. Many ranking approaches have emerged to rank the top product recommendation to give to user. Ratings calculated can be an explicit or impliocit rating. Popular sites are Amazon.com, Netflix.com, and Movielens.com etc. These sites help the customers to find relevant product to their interest. They play as a place where customers can find all kinds of items. They do so because recommendations given by other customers have been published after they have used the product. Those customers will have experience about the product. From the customers their view of how is the product usage has been collected. This is used in recommendations. In the Proposed system, customer’s views are used for recommendations. While new customer search products, old users views are published for the particular product. On getting the customer views, one user can trust it since common people have more confidence on words-of other people. Based on product, users are given form to fill their views.On getting views, Ratings are calculated from it. These kind of recommender system give useful recommendations since we collect views of people who are familiar with the item or product. Index Terms: Recommender system, E-Commerce, collaborative filtering, Customer reviews
Recommender System _Module 1_Introduction to Recommender System.pptxSatyam Sharma
This document provides an introduction and overview of a module on recommender systems. The module aims to help students understand the importance and basic concepts of recommender systems. The syllabus covers introduction to recommender systems, different types of recommender systems including collaborative filtering, content-based, and knowledge-based systems. It also discusses hybrid systems, application and evaluation techniques, and emerging topics and challenges. The objective is for students to learn the basic concepts, understand different recommender system types, and be able to evaluate recommender systems as a multidisciplinary field.
This document provides an overview of recommender systems for e-commerce. It discusses various recommender approaches including collaborative filtering algorithms like nearest neighbor methods, item-based collaborative filtering, and matrix factorization. It also covers content-based recommendation, classification techniques, addressing challenges like data sparsity and scalability, and hybrid recommendation approaches.
2 hours training on Mobile UX with Farah Nuraini, Interaction Designer at Traveloka, Indonesia
45 min theory: Research, Analysis, Design solutions and Testing
+ 1h15 min of hands-on exercises with the 5 facilitators from Traveloka.
This document discusses various market analysis tools used in new product development (NPD). It describes tools for idea generation, product optimization, marketing mix optimization, and market prediction. Specific tools covered include brainstorming, the Delphi technique, focus groups, user observation, conjoint analysis, concept testing, prototyping, and diffusion models. The document provides details on how each tool is used, its process, advantages, and disadvantages to help product developers select the right tools for analyzing markets and predicting new product success.
A presentation that gives an overview about latest machine learning and deep learning techniques and use-cases that are prevalent in the eCommerce industry
Supercharge Your Corporate Dashboards With UX AnalyticsUserZoom
This document discusses supercharging corporate dashboards with UserZoom. It begins with polls to understand the audience. It then discusses the current state of user research, including common organizational models and goals to grow the practice. It defines three types of dashboards - individual research engagements, product scorecards, and executive dashboards. Product scorecards provide consistent reporting of UX measures to product teams. Executive dashboards do the same for leadership. The document outlines a vision for UserZoom to be a single place for all dashboard data and describes plans to provide this capability starting in 2017.
Similar to Product Recommendations Enhanced with Reviews (20)
This PDF delves into the aspects of information security from a forensic perspective, focusing on privacy leaks. It provides insights into the methods and tools used in forensic investigations to uncover and mitigate privacy breaches in mobile and cloud environments.
It's your unstructured data: How to get your GenAI app to production (and spe...Zilliz
So you've successfully built a GenAI app POC for your company -- now comes the hard part: bringing it to production. Aparavi addresses the challenges of AI projects while addressing data privacy and PII. Our Service for RAG helps AI developers and data scientists to scale their app to 1000s to millions of users using corporate unstructured data. Aparavi’s AI Data Loader cleans, prepares and then loads only the relevant unstructured data for each AI project/app, enabling you to operationalize the creation of GenAI apps easily and accurately while giving you the time to focus on what you really want to do - building a great AI application with useful and relevant context. All within your environment and never having to share private corporate data with anyone - not even Aparavi.
Increase Quality with User Access Policies - July 2024Peter Caitens
⭐️ Increase Quality with User Access Policies ⭐️, presented by Peter Caitens and Adam Best of Salesforce. View the slides from this session to hear all about “User Access Policies” and how they can help you onboard users faster with greater quality.
Discovery Series - Zero to Hero - Task Mining Session 1DianaGray10
This session is focused on providing you with an introduction to task mining. We will go over different types of task mining and provide you with a real-world demo on each type of task mining in detail.
TrustArc Webinar - Innovating with TRUSTe Responsible AI CertificationTrustArc
In a landmark year marked by significant AI advancements, it’s vital to prioritize transparency, accountability, and respect for privacy rights with your AI innovation.
Learn how to navigate the shifting AI landscape with our innovative solution TRUSTe Responsible AI Certification, the first AI certification designed for data protection and privacy. Crafted by a team with 10,000+ privacy certifications issued, this framework integrated industry standards and laws for responsible AI governance.
This webinar will review:
- How compliance can play a role in the development and deployment of AI systems
- How to model trust and transparency across products and services
- How to save time and work smarter in understanding regulatory obligations, including AI
- How to operationalize and deploy AI governance best practices in your organization
DefCamp_2016_Chemerkin_Yury-publish.pdf - Presentation by Yury Chemerkin at DefCamp 2016 discussing mobile app vulnerabilities, data protection issues, and analysis of security levels across different types of mobile applications.
UiPath Community Day Amsterdam: Code, Collaborate, ConnectUiPathCommunity
Welcome to our third live UiPath Community Day Amsterdam! Come join us for a half-day of networking and UiPath Platform deep-dives, for devs and non-devs alike, in the middle of summer ☀.
📕 Agenda:
12:30 Welcome Coffee/Light Lunch ☕
13:00 Event opening speech
Ebert Knol, Managing Partner, Tacstone Technology
Jonathan Smith, UiPath MVP, RPA Lead, Ciphix
Cristina Vidu, Senior Marketing Manager, UiPath Community EMEA
Dion Mes, Principal Sales Engineer, UiPath
13:15 ASML: RPA as Tactical Automation
Tactical robotic process automation for solving short-term challenges, while establishing standard and re-usable interfaces that fit IT's long-term goals and objectives.
Yannic Suurmeijer, System Architect, ASML
13:30 PostNL: an insight into RPA at PostNL
Showcasing the solutions our automations have provided, the challenges we’ve faced, and the best practices we’ve developed to support our logistics operations.
Leonard Renne, RPA Developer, PostNL
13:45 Break (30')
14:15 Breakout Sessions: Round 1
Modern Document Understanding in the cloud platform: AI-driven UiPath Document Understanding
Mike Bos, Senior Automation Developer, Tacstone Technology
Process Orchestration: scale up and have your Robots work in harmony
Jon Smith, UiPath MVP, RPA Lead, Ciphix
UiPath Integration Service: connect applications, leverage prebuilt connectors, and set up customer connectors
Johans Brink, CTO, MvR digital workforce
15:00 Breakout Sessions: Round 2
Automation, and GenAI: practical use cases for value generation
Thomas Janssen, UiPath MVP, Senior Automation Developer, Automation Heroes
Human in the Loop/Action Center
Dion Mes, Principal Sales Engineer @UiPath
Improving development with coded workflows
Idris Janszen, Technical Consultant, Ilionx
15:45 End remarks
16:00 Community fun games, sharing knowledge, drinks, and bites 🍻
Demystifying Neural Networks And Building Cybersecurity ApplicationsPriyanka Aash
In today's rapidly evolving technological landscape, Artificial Neural Networks (ANNs) have emerged as a cornerstone of artificial intelligence, revolutionizing various fields including cybersecurity. Inspired by the intricacies of the human brain, ANNs have a rich history and a complex structure that enables them to learn and make decisions. This blog aims to unravel the mysteries of neural networks, explore their mathematical foundations, and demonstrate their practical applications, particularly in building robust malware detection systems using Convolutional Neural Networks (CNNs).
Finetuning GenAI For Hacking and DefendingPriyanka Aash
Generative AI, particularly through the lens of large language models (LLMs), represents a transformative leap in artificial intelligence. With advancements that have fundamentally altered our approach to AI, understanding and leveraging these technologies is crucial for innovators and practitioners alike. This comprehensive exploration delves into the intricacies of GenAI, from its foundational principles and historical evolution to its practical applications in security and beyond.
1. Product recommendations
enhanced with reviews
Muthusamy Chelliah
Director, Academic Engagement, Flipkart
Sudeshna Sarkar
Professor, CSE, IIT Kharagpur
Email: sudeshna@cse.iitkgp.ernet.in
11th ACM Conference on Recommender Systems, Como, Italy, 27th-31st August 2017
1
2. Reviews
2
Which aspect is more important?
Food or Location?
What value of the aspect is desired?
Location near the lake or
Location near the conference venue?
4. Product Recommendation
• Suggest new products for
current selection
– Substitute
– Complementary
• Help offer
– Up-sell
– Cross-sell
– Bundle
4
5. Rich content in user reviews
About Looks : Very very impressive like a premium phone but back side is not so attractive,
it's look a like Redmi Series Phone's as well as back camera edge style is boring.
I'm not impress with the picture quality coz its not what Samsung provide even in his low
range phones, if you compare with moto series phones i will give 10/7 to samsung and
moto 10/9, also the shake reduction rate on this phone is very bad, After taking the picture
(naturally our hand shake little bit) the picture is blurry or shaked.
Sound : Not for who listening the music on speakers, While ringing it's good but not
excellent
Well I'm shocked in packaging too as soon as i see the samsung box it's look they given me
a 2 year old box
User Modeling Recommendation Generation
5
6. Review Enhanced Recommendation
1. Introduction
2. Review mining
a) Background: Features and Sentiment, Latent features
b) Statistical approaches. LDA
c) Recent topic models
d) Deep Learning (ABSA)
e) Flipkart ABSA/Recommendation
3. Review based Recommendation
a) Handling Data Sparsity
b) Topic Model, Matrix Factorization and Mixture Model based Recommendation
c) Deep learning based Recommendation
4. Explainable Product Recommendations
5. Summary, Future Trend
6
8. The Recommendation Task
• Rating Prediction: Estimate a utility function
to predict how a user will like an item
• Recommendation: Recommend a set of items
to maximize user’s utility
8
9. Recommender Systems
Data
• Content
– Product Information
• Product Taxonomy
• Product Attributes
• Product Description
– Customer Demographics
• User Activity
– Purchase
– Rating, Preference
– Click / Browse
– User generated reviews
• Social data
Methods
• Content based
– Recommend items similar to those
a user has liked in the past
– Issues: Does not use quality
judgments of users
• Collaborative filtering
– Finds users with similar tastes as
target user to make
recommendation
– Use ratings
– Issues: Sparsity, Cold Start
9
10. Content based recommender systems
Product description Product Ontology User Review
10
• Each item defined by a profile
vector obtained from the item’s
textual description, metadata, etc.
• Weighting methods such as TF-IDF
• User profile vector : by aggregating
the profile vectors of the items that
the user liked or purchased in the past.
• Recommend items with high similarity
between user vector and item vector.
11. User Reviews provide additional rich data.
One Review worth much more than one rating.
Collaborative filtering
Predict user preferences on new items by collecting
rating information from many users (collaborative).
• Many items to choose from
• Few data per user
• No data for new user
Difficulties
• Data sparsity
– too few ratings
– 90%+ items with less than 10 reviews
• Cold-start
– New user/item appear continuously
11
12. Collaborative filtering
Memory Based CF:
1. User-based CF
2. Item-based CF
Model-based CF:
• Train a parametric model with the rating matrix
• Applied to predict ratings of unknown items
o Ex: latent factor model
12
13. Latent Factor Model:
Matrix Factorization
• Map both users and items
to a joint latent factor space
• item 𝑖𝑖 : associated vector 𝑞𝑞𝑖𝑖
• user 𝑢𝑢 : vector 𝑝𝑝𝑢𝑢
• 𝑞𝑞𝑖𝑖
𝑇𝑇
𝑝𝑝𝑢𝑢 captures the
interaction between 𝑢𝑢 and 𝑖𝑖
• This approximates 𝑟𝑟𝑢𝑢𝑢𝑢 : user
𝑢𝑢’s rating of item 𝑖𝑖
• Factor rating matrix
using SVD: obtain , ,
• Reduce the matrix to
dimension
• Compute two resultant
matrices:
and
• Predicting task
16. Why use reviews?
• Reviews provide rich information.
• Reviews help to explain users’
ratings.
• Reviews are useful at modeling
new users and products: one
review tells us much more than
one rating.
• Reviews help us to discover the
aspects of users’ opinions
• Modeling these factors
improve recommendation
• Value for customer
– Help explore space of options
– Discover new things
• Value for company
– Increase trust and loyalty
– Increase conversion
– Opportunities for promotion,
persuasion
– Obtain more knowledge about
customers
16
17. The Value of Reviews
• Recommender systems played an
important part in helping users
find products.
• Reviews were used by users but
only relevant ones.
• Any algorithmic recommender
needs to communicate results to
the user in a meaningful manner.
• The volume of reviews affects
product conversion (upto
270% increase)
• High priced items have a
higher value for reviews.
17
Case Amazon: Ratings and Reviews as
Part of Recommendations
Juha Leino and Kari-Jouko Räihä, RecSys
2007
The Value of Online Customer Reviews
Georgios Askalidis, Edward C. Malthouse
RecSys 2016
18. The importance of Explanations
• Explanations improve user
engagement and
satisfaction with a
recommender system.
• Controllability and
transparency.
• Why are these items
recommended?
• Trust
– user’s confidence
• Persuasive
– Convince users to try/buy
• Satisfaction
– user’s enjoyment
• Help compose judgment
– attributes to a rating
18
19. Product recommendations
enhanced with reviews
Part I
Muthusamy Chelliah
Director, Academic Engagement, Flipkart
Sudeshna Sarkar
Professor, CSE, IIT Kharagpur
Email: sudeshna@cse.iitkgp.ernet.in
11th ACM Conference on Recommender Systems, Como, Italy, 27th-31st August 2017
19
23. Definitions
• Contributor: The person or organization who is expressing their
opinion in text.
• Object: An entity which can be a product, service, person, event,
organization, or topic.
• Review: A contributor-generated text that contains their opinions
about some aspects of the object.
• Aspect: The component, attribute or feature of the object that
contributor has commented on.
• Opinion: An opinion on an aspect is a positive, neutral or negative
view, attitude, emotion or appraisal on that aspect from a
contributor.
23
24. Definitions
• Contributor: Contributor is the person or organization who is
expressing his/her/its opinion in written language or text.
• Object: An object is an entity which can be a product, service,
person, event, organization, or topic.
• Review: Review is a contributor-generated text that contains
the opinions of the contributor about some aspects of the
object.
• Aspect: Aspect is the component, attribute or feature of the
object that contributor has commented on.
• Opinion: An opinion on an aspect is a positive, neutral or
negative view, attitude, emotion or appraisal on that aspect
from a contributor.
24
25. Terminologies
• Features / Attributes / Aspect / Properties of a
product or service
• Opinion
– Overall opinion
– Feature opinion
• Review helpfulness:
– the number of “helpful" votes given by readers
– Can be used to determine quality score
25
26. Aspect-based sentiment analysis (ABSA)
• [Liu ‘16]
• screen is clear and great
• Aspect extraction: screen
• Opinion identification: clear,
great
• Polarity classification: clear is +ve
• Opinion separation/generality:
clear (aspect-specific), great
(general)
– Understand consumer taste of
products
26
[Moghaddam ’10]
Prediction
28. Recommendation: User Needs
• Help me narrow down on my choice
–I’ve shortlisted few cameras and need help
deciding one
–I’m almost leaning towards buying this speaker
system but want to get further confidence and
assurance on sound quality.
28
29. Flipkart Ratings-based Recommendations
• Product discovery with
Rating as a signal
– rank products with high
quality ratings at the
top
• Highlight Ratings to
narrow down purchase
– show ratings attributes
that map to user intent
29
Top Rated phones with good Front camera
Because you searched for good Selfie phones
Cameras with good image stabilization >
Camera good for night mode >
30. Long tail (both in users/products):
induces sparsity/cold-start
30
31. Recommendations Leveraging Reviews
• Good coverage
• High quality and detailed reviews
• Parsing of reviews into meaningful clusters for
Discovery
– Picture Quality (4.5), Sound (3.8), Ease of Installation (2.5)
31
32. ABSA - User needs
• Individual preferences
– easy way to shortlist products matching a criteria
– compare two products on certain aspects
• Too many reviews to read
– interested in only key elements of a product
– be aware of product shortcomings before making
the final buying decision
32
33. ABSA - Partner Needs
• Refine available selection per user preference
• Brands can validate future product design
• Correlation between aspect and sales is a
good indicator of market demand
33
34. ABSA - agenda
• Statistics-based text mining (5 minutes)
• Topic models
– Early (5 minutes)
– Recent (10 minutes)
• Deep learning (10 minutes)
34
37. Aspect extraction
• Frequency-based
– Extract frequently occurring
nouns based on seed words
[Moghaddam ‘10]
• Pattern mining
– Rules for identifying templates
of POS-tagged tokens
(e.g., _JJ_ASP)
• Clustering
– Group aspect nouns [Guo ‘09]
37
38. Attribute-aspect normalization
[Bing ‘16]
• Extract product attributes from description page
• Infer aspect popularity in reviews/map to feature
• Hidden CRF bridges vocabulary gap
38
39. References
• Bing, L., Wong, T.L. and Lam, W. Unsupervised extraction of
popular product attributes from e-commerce web sites by
considering customer reviews. ACM TOIT ’16.
• Guo, H., Zhu, H., Guo, Z., Zhang, X., & Su, Z. (2009, November).
Product feature categorization with multilevel latent semantic
association. CIKM 2009.
• Moghaddam, S., & Ester, M. Opinion digger: an unsupervised
opinion miner from unstructured product reviews. CIKM 2010
39
40. ABSA - agenda
• Statistics-based text mining
• Topic models
– Early (5 minutes)
– Recent (10 minutes)
• Deep learning (10 minutes)
40
41. Topic Modeling: Latent Dirichlet Allocation
(LDA)
41
• Given a document set D, assume for each , there is an associated K-
dimensional topic distribution , which encodes the fraction of words
in d that discuss each of the K topics. That is, words in document d
discuss topic k with probability .
• Each topic k has an associated word distribution , which encodes the
probability that a particular word is used for that topic k.
• and are random vectors each has a Dirichlet prior distribution.
Document d 𝜃𝜃𝑑𝑑,1, 𝜃𝜃𝑑𝑑,2. … , 𝜃𝜃𝑑𝑑,𝑘𝑘, … , 𝜃𝜃𝑑𝑑,𝐾𝐾
𝜙𝜙𝑘𝑘,1
𝜙𝜙𝑘𝑘,2
⋮
𝜙𝜙𝑘𝑘,𝑗𝑗
⋮
𝜙𝜙𝑘𝑘,𝑁𝑁
Topic 1 to K
Topic k
Word j
42. Aspect-based sentiment analysis (ABSA)
• [Liu ‘16]
• screen is clear and great
• Aspect extraction: screen
• Opinion identification: clear,
great
• Polarity classification: clear is +ve
• Opinion separation/generality:
clear (aspect-specific), great
(general)
– Understand consumer taste of
products
42
[Moghaddam ’10]
Prediction
43. LDA – review mining
• Each item: finite mixture over
an underlying set of latent
variables
• Generative, probabilistic
model for collection of
discrete items
– Review corpora generated first
by choosing a value for
aspect/rating pair <am, rm> (θ)
– Repeatedly sample M opinion
phrases<tm, sm>conditioned on θ
43
44. Extract aspects/ratings together
• Topic modeling helps only with document clustering
• Independent aspect identification/ rating prediction leads to errors
[Moghaddam ‘11]
– Same sentiment word shows different opinions for various aspects
• Dataset
– 29609 phrases, 2483 reviews, 5 categories from Epinions
44
45. Aspect-based sentiment analysis (ABSA)
• [Liu ‘16]
• screen is clear and great
• Aspect extraction: screen
• Opinion identification: clear,
great
• Polarity classification: clear is +ve
• Opinion separation/generality:
clear (aspect-specific), great
(general)
– Understand consumer taste of
products
45
[Moghaddam ’10]
Prediction
46. Interdependent LDA model
• [Moghaddam ‘11]
• Generate an aspect am
from LDA
• Generate rating rm
conditioned on sample
aspect
• Draw head term tm and
sentiment sm conditioned
on am and rm respectively
46
47. References
Liu, B. Sentiment analysis. AAAI 2011.
Moghaddam S, Ester M. ILDA: interdependent LDA
model for learning latent aspects and their ratings
from online product reviews. SIGIR ’11
Moghaddam, S., & Ester, M. Aspect-based opinion
mining from online reviews. SIGIR 2012.
47
48. Aspect-based sentiment analysis (ABSA)
• Statistics-based text mining
• Topic models
– Early
– Recent (10 minutes)
• Deep learning (10 minutes)
48
49. Structure in Aspect-Sentiment
• Different consumers interested in different aspect granularities
• Polarity of general to aspect-specific sentiment words through
sparse co-occurrence [Kim ‘13]
• Dataset
– 10414 (laptop)/20862 (digital SLR) reviews, 825/260 targets: Amazon
49
50. Hierarchical aspect-sentiment model
• [Kim ‘13]
• Jointly infer aspect-sentiment tree
– with a bayesian non-parametric model
as prior
• Aspect-sentiment node φk itself is a
tree
– aspect topic at root and sentiment-polar
(S) in leaves
• Topics share semantic theme
– generated from dirichlet (beta)
distribution
50
Light..Heavy
Portability
|
Battery
52. Aspect-based sentiment analysis (ABSA)
• [Liu ‘16]
• screen is clear and great
• Aspect extraction: screen
• Opinion identification: clear, great
• Polarity classification: clear is +ve
• Opinion separation/generality: clear
(aspect-specific), great (general)
– Understand consumer taste of
products
52
[Moghaddam ’10]
Prediction
53. Aspect-sentiment topic
• [Wang ‘16]
• Generated topics inconsistent with human judgment
– Wrong identification of opinion words (nice screen)
• Models based on co-occurrence suffer
– Opinions (smooth screen) NOT identified if mixed in same
topic
53
54. Lifelong learning across categories
• [Wang ‘16]
• Extract/accumulate knowledge from past
• Use discovered knowledge for future learning
– Like human experience
• Dataset:
– 50 product domains with 1000 reviews each -
Amazon
54
58. Attribute-aspect normalization
[Bing ‘16]
• Extract product attributes from description page
• Infer aspect popularity in reviews/map to feature
• Hidden CRF bridges vocabulary gap
58
59. Augmented specification: camera
• [Park ‘15]
• Advanced features hard
for novice customers
• Specs can be enhanced
with reviews
– Opinions on feature
value/importance
– Product-specific words
59
60. DuanLDA
[Park ‘15]
• Each document is
concatenated review
• s is feature-value
pair/specification topic
• Switch x decides
background/spec. topic
60
Review aspect
Feature
- value
61. SpecLDA
[Park ‘15]
• Feature-value pairs with same
feature share info
• Organized as hierarchy for
better topic estimation
• Capture user preference to
indicate a pair
– for feature-/value-related words
• Employ value-word prior in
addition to feature-wordReview aspect Feature
value
Feature
61
62. Evaluation: top words for a spec. topic
[Park ‘15]
• Spec. (“Display – type”,“2.5in LCD display”)
• DuanLDA – does not use value word prior
• SpecLDA – feature and value topics form a cluster
62
63. ABSA Summary (Topic models)
LDA: sample opinion
phrases, not all terms
ILDA: Correspondence
betn. aspect/rating
HASM: aspect-sentiment
tree at each with topic root
and S sentiment leaves SpecLDA: feature-,
value- word, aspect
JAST: aspect: (term,
opinion), polarity,
generic opinion
63
64. References
Kim, S., Zhang, J., Chen, Z., Oh, A. H., & Liu, S. A Hierarchical
Aspect-Sentiment Model for Online Reviews. AAAI 2013.
Park, D.H., Zhai, C. and Guo, L. Speclda: Modeling product
reviews and specifications to generate augmented
specifications. SDM 2015.
Wang, S., Chen, Z., & Liu, B. Mining aspect-specific opinion
using a holistic lifelong topic model. WWW 2016.
64
65. Aspect-based sentiment analysis (ABSA)
• Statistics-based text mining
• Topic models
– Early
– Recent
• Deep learning (10 minutes)
65
66. Aspect-based sentiment analysis (ABSA)
• [Liu ‘16]
• screen is clear and great
• Aspect extraction: screen
• Opinion identification: clear,
great
• Polarity classification: clear is +ve
• Opinion separation/generality:
clear (aspect-specific), great
(general)
– Understand consumer taste of
products
66
[Moghaddam ’10]
Prediction
68. Semantic compositionality
• Rating prediction [Socher ‘13]
• Word spaces cannot express meaning of longer phrases in a principled way
• Sentiment detection requires powerful models/supervised resources
68
69. Recursive deep models
• Rating prediction
[Socher ‘13]
• Parent vectors
computed bottom-up
• compositionality
function varies for
different models
• Node vectors used as
features for classifier
69
70. Semantic product feature mining
• Aspect extraction [Xu ‘14]
• Reviews features mined using
syntactic patterns
• Syntax helps leverage only context
– Suffering from sparsity of
discrete information
• Term-product association needs
semantic clues
70
72. Recurrent neural networks
• Aspect extraction [Liu ’15]
• window-based approach for short-term dependencies
• bidirectional links for long-range dependencies
• linguistic features to guide training
72
73. Deep memory network: LSTM
• Aspect extraction [Tang ‘16]
• Fast processor but small
memory
• Need to capture semantic
relations of aspect/context
words
• LSTM incapable of
exhibiting context clues
• Selectively focus instead on
context parts
73
74. Deep memory network: attention
• Aspect extraction [Tang ‘16]
• Multiple hops are stacked
– so abstractive evidence is
selected from external memory
• Attention mechanism
– weight to lower position while
computing upper representation
• linear/attention layers with
shared parameters in each hop
• Context word importance
based on distance to aspect
74
75. ABSA summary (deep learning)
Content/location attention to
emphasize context word for LSTM
Similarity graph for lexical and
CNN for context semantics
Window for short-term and
bidirectionality for long-range
dependencies in RNN Parent vectors
computed bottom-
up in RecNN for
semantic
compositionality
75
76. References
Liu, P., Joty, S. R., & Meng, H. M. Fine-grained Opinion Mining with
Recurrent Neural Networks and Word Embeddings. EMNLP ’15.
Socher, R., Perelygin, A., Wu, J.Y., Chuang, J., Manning, C.D., Ng, A.Y.
and Potts, C., Recursive deep models for semantic compositionality
over a sentiment treebank, EMNLP ’13.
Tang, D., Qin, B. and Liu, T. Aspect level sentiment classification with
deep memory network. EMNLP ’16
Xu, L., Liu, K., Lai, S. and Zhao, J.. Product Feature Mining: Semantic
Clues versus Syntactic Constituents. ACL 2014.
76
78. Trusted advisor
Objective
● Transform Flipkart into a research
destination
● Shorten the purchase cycle for a user
Aspect Reviews
● Identify what users care about while
buying a specific product
● Organize/classify reviews into these
aspects & sub-aspects
Aspect Ratings
● Organize sentiments (polarity & strength)
● Summarize sentiments in the form of a
user rating at an aspect level
78
79. Future directions
• Product comparison
– Model entities assessed
directly in a single review
sentence [Sikchi ‘16]
– Leverage product spec. to
learn attribute importance
• Question/answer
– Summarize reviews with
real-world questions
– Rerank initial candidates
for diversity [Liu ‘16]
79
80. References
• Liu, M., Fang, Y., Park, D.H., Hu, X. and Yu, Z.
Retrieving Non-Redundant Questions to
Summarize a Product Review. SIGIR 2016.
• Sikchi, A., Goyal, P. and Datta, S., Peq: An
explainable, specification-based, aspect-
oriented product comparator for e-
commerce. CIKM 2016.
80
83. Handling Cold-Start
• L1 ranking – leverage content (e-commerce has structured data),
business provided attributes (CTR, conversion rate)
• L2 ranking - many online learning algorithms – AdaGrad, FTRL,
AdPredictor for real-time learning (per-coordinate learning) , can
handle new users/products very soon
• Other sources: Reviews, ratings
– Some issues
• Reviews/ratings are always biased to extremes
• Ratings: Users very satisfied or unsatisfied
• Reviews: more information to type. Biased towards
angst/unsatisfactory experience
83
84. Future directions
84
• Bundle [Pathak ‘17]
– Personalized/not
considered in isolation
– Assess size/compatibility
• Upsell [Lu ‘14]
– Maximize revenue thru
utility, saturation and
competing products
85. References
• Pathak, A., Gupta, K., & McAuley, J.
Generating and Personalizing Bundle
Recommendations on Steam. SIGIR ’17.
• Lu, W., Chen, S., Li, K., & Lakshmanan, L. V.
Show me the money: Dynamic
recommendations for revenue maximization,
VLDB ‘14.
85
87. Product recommendations enhanced
with reviews
Part 2
Muthusamy Chelliah
Director, Academic Engagement, Flipkart
Sudeshna Sarkar
Professor, IIT Kharagpur
Email: sudeshn@cse.iitkgp.ernet.in
11th ACM Conference on Recommender Systems, Como, Italy, 27th-31st August 2017
87
88. Review Enhanced Recommendation
1. Introduction
2. Review mining
a) Background: Features and Sentiment, Latent features
b) Statistical approaches. LDA
c) Recent topic models
d) Deep Learning (ABSA)
e) Flipkart ABSA/Recommendation
3. Review based Recommendation
a) Handling Data Sparsity
b) Topic Model, Matrix Factorization and Mixture Model based Recommendation
c) Deep learning based Recommendation
4. Explainable Product Recommendations
5. Summary, Future Trend
88
90. Issues with Recommender Systems
Collaborative Filtering
• Cold Start
– New items, New Users
• Sparsity
– it is hard to find users
who rated the same
items.
• Popularity Bias
– Cannot recommend
items to users with
unique tastes.
Content Based
• Hard to identify quality judgement.
• Ignores user’s ratings.
90
Review Based
• Text information from reviews address
sparseness of rating matrix.
– Bag of words model
• More sophisticated models required to
capture contextual information in
documents for deeper understanding:
CNN, etc.
• Enable Explanations
91. Review Aware Recommender Systems
Traditional
Models
Weight on quality of
ratings, Recsys’12
Use reviews as
content, Recsys’13
Aspect Weighting,
WI’15
TriRank, CIKM’15
Topic Models
and Latent
Factors
HFT, Recsys ‘13
RMR, Recsys‘14
CMR, CIKM‘14
TopicMF, AAAI’14
FLAME, WSDM’15
Joint Models of
Aspects and
Ratings
JMARS,KDD’14
EFM, SIGIR’14
TPCF, IJCAI’13
SULM, KDD’17
Deep Learning
Models
ConvMF, Recsys’16
DeepCONN,
WSDM’17
TransNets, Recsys’17
D-Attn, Recsys’17
91
92. Review Quality Aware Collaborative
Filtering, by Sindhu Raghavan, Suriya
Gunasekar, Joydeep Ghosh, Recsys’12
Using weights based on reviews
• Propose a unified framework
for performing the three tasks:
– aspect identification
– aspect-based rating inference
– weight estimation.
• Attach weights or quality scores to
the ratings.
• The quality scores are determined
from the corresponding review.
• Incorporate quality scores as
weights for the ratings in the basic
PMF framework
92
Review Mining for Estimating
Users’ Ratings and Weights for
Product Aspects, by Feng Wang
and Li Chen, WI’15
93. Sentimental Product Recommendation
1. Convert unstructured reviews into rich product descriptions
a) Extract Review Features using shallow NLP
b) Evaluate Feature Sentiment by opinion pattern mining
2. Content-based recommendation approach based on feature
similarity to a query product
a) Similarity-Based Recommendation
b) Sentiment-Enhanced Recommendation: seek products with better
sentiment than the query product.
93
Sentimental Product Recommendation, by Ruihai Dong, Michael P. O’Mahony,
Markus Schaal, Kevin McCarthy, Barry Smyth, Recsys 2013
95. HFT: Linking Reviews
1. Latent Dirichlet Allocation (LDA) finds low dimensional structure in review
text (topic representation)
2. SVD learns latent factors for each item.
• Statistical models combine latent dimensions in rating data with topics in
review text.
• Use a transform that aligns latent rating and review terms.
• Link them using an objective that combines the accuracy of rating
prediction (in terms of MSE) with the likelihood of the review corpus (using
a topic model).
95
Hidden Factors and Hidden Topics: Understanding Rating Dimensions with Review
Text: McAuley and Leskovec, Recsys 2013
96. HFT: Optimization
• A standard recommender system : 𝑟𝑟̂𝑢𝑢,𝑖𝑖 = 𝑝𝑝𝑢𝑢. 𝑞𝑞𝑖𝑖
• Minimize MSE:
min
𝑃𝑃,𝑄𝑄
1
𝒯𝒯
� 𝑟𝑟𝑢𝑢𝑢𝑢 − 𝑞𝑞𝑖𝑖 𝑝𝑝𝑢𝑢
2
𝒯𝒯
+ 𝜆𝜆1 � 𝑝𝑝𝑢𝑢
2
+𝜆𝜆2 � 𝑞𝑞𝑖𝑖
2
𝑥𝑥𝑥𝑥
• Instead use review text as regularizer:
min
𝑃𝑃,𝑄𝑄
1
𝒯𝒯
� 𝑟𝑟𝑢𝑢𝑢𝑢 − 𝑞𝑞𝑖𝑖 𝑝𝑝𝑢𝑢
2
𝒯𝒯
− 𝜆𝜆 𝑙𝑙 𝒯𝒯|Θ, 𝜙𝜙, 𝑧𝑧
Rating error Corpus likelihood
96
97. HFT: Combining ratings and reviews
• Matrix factorization and LDA project users and items
into low-dimensional space
• Align the two
97
98. HFT: Combining ratings and reviews
𝜃𝜃𝑖𝑖,𝑘𝑘 =
exp 𝒦𝒦 𝑞𝑞𝑖𝑖, 𝑘𝑘
∑ 𝑒𝑒𝑒𝑒𝑒𝑒 𝒦𝒦 𝑞𝑞𝑖𝑖. 𝑘𝑘𝑘𝑘′
𝑞𝑞𝑖𝑖 ∈ ℛ 𝐾𝐾
𝜃𝜃𝑖𝑖 ∈ Δ𝐾𝐾
By linking rating and opinion models, find topics in reviews
that inform about opinions
98
99. HFT: Model Fitting
Repeat Step 1 and Step 2 until convergence:
Step 1: Fit a rating model regularized by the topic
(solved via gradient ascent using L-BFGS)
Step 2: Identify topics that “explain” the ratings.
Sample with probability
(solved via Gibbs sampling)
99
100. Ratings Meet Reviews, a Combined
Approach to Recommend
• Use LDA to model reviews
• Use mixture of Gaussians
to model ratings
• Combine ratings and
reviews by sharing the
same topic distribution
• N users 𝑢𝑢𝑖𝑖
• M items 𝑣𝑣𝑗𝑗
• 𝑢𝑢𝑖𝑖, 𝑣𝑣𝑗𝑗 defines the observed
ratings 𝑋𝑋 = 𝑥𝑥𝑖𝑖𝑖𝑖 ,
• Associated review 𝑟𝑟𝑖𝑖𝑖𝑖
100
• Ratings meet reviews, a combined approach to
recommend, Ling et al., Recsys 2014.
102. Use LDA to
model the
reviews
Use mixture of
Gaussians to model
the ratings
Use the same
topic distribution
to connect the rating
part and the review
part
102
104. Results for Cold Start Setting
• The improvement of HFT and RMR over
traditional MF is more significant for datasets
that are sparse.
104
105. CMR
• Predict user’s ratings by considering review texts as well as
hidden user communities and item groups relationship.
• Uses Co-clustering to capture the relationship between
users and items.
• Models as a mixed membership over community and group
respectively.
• Each user or item can belong to multiple communities or
groups with varying degrees.
105
Collaborative Filtering Incorporating Review Text and Co-clusters of Hidden User
Communities and Item Groups, Yinqing Xu Wai Lam Tianyi Lin, CIKM 2014
106. Collapsed Gibbs sampling algorithm to perform approximate inference.
CMR Graphical Model
106
U: user collection
V: item collection
𝑟𝑟𝑢𝑢𝑢𝑢: u’s rating of item v
𝑑𝑑𝑢𝑢𝑢𝑢: u’s review of item v
𝑤𝑤𝑢𝑢𝑢𝑢𝑛𝑛: nth word of 𝑑𝑑𝑢𝑢𝑢𝑢
𝑧𝑧𝑢𝑢𝑢𝑢𝑢𝑢: topic of 𝑤𝑤𝑢𝑢𝑢𝑢𝑛𝑛
𝜋𝜋
𝑐𝑐
𝑢𝑢
: community membership of u
𝜋𝜋
𝑔𝑔
𝑣𝑣
: group membership of v
𝑐𝑐𝑢𝑢𝑢𝑢: user community of 𝑑𝑑𝑢𝑢𝑢𝑢 and 𝑟𝑟𝑢𝑢𝑢𝑢
𝑔𝑔𝑢𝑢𝑢𝑢: item group of 𝑑𝑑𝑢𝑢𝑢𝑢 and 𝑟𝑟𝑢𝑢𝑢𝑢
𝜃𝜃𝑐𝑐𝑐𝑐: topic distribution of co-cluster (c,g)
𝜓𝜓𝑐𝑐𝑐𝑐: rating distribution of co-cluster (c,g)
𝜙𝜙𝑘𝑘: topic distribution of topic k
108. TopicMF
• Matrix Factorization (MF) factorizes
user-item rating matrix into latent user
and item factors.
• Simultaneously topic modeling
technique with Nonnegative Matrix
Factorization (NMF) models the latent
topics in review texts.
Align these two tasks by using a transform from item and user latent vectors to topic
distribution parameters so as to combine latent factors in rating data with topics in
user-review text.
108
TopicMF: Simultaneously Exploiting Ratings and Reviews for
Recommendation, by Yang Bao, Hui Fang, Jie Zhang, AAAI,2014
109. aspects
JMARS Model
Aspect Modeling
(Interest of the user/ Property of the item)
𝜃𝜃𝑢𝑢~ 𝒩𝒩 0, 𝜎𝜎2
user−aspect𝐼𝐼
𝜃𝜃𝑚𝑚~ 𝒩𝒩 0, 𝜎𝜎2
item−aspect𝐼𝐼
𝜃𝜃𝑢𝑢𝑢𝑢~𝑒𝑒𝑒𝑒𝑒𝑒 𝜃𝜃𝑢𝑢 + 𝜃𝜃𝑚𝑚
Rating Modeling
Aspect-specific rating
𝑣𝑣𝑢𝑢
𝑇𝑇
𝑀𝑀𝑎𝑎 𝑣𝑣𝑚𝑚 + 𝑏𝑏𝑢𝑢 + 𝑏𝑏𝑚𝑚
Overall Rating
𝒓𝒓� 𝒖𝒖𝒖𝒖 = � 𝜽𝜽𝒖𝒖𝒖𝒖𝒖𝒖 𝒓𝒓𝒖𝒖𝒖𝒖𝒖𝒖
𝒂𝒂
𝑀𝑀𝑎𝑎 emphasizes the aspect specific properties.
Rating and Review Model
user item
109
Jointly Modeling Aspects, Ratings and Sentiments for Movie Recommendation (JMARS), by
Qiming Diao, Minghui Qiu, Chao-Yuan Wu, KDD 2014
110. JMARS Language Model
110
We assume that the review language model is given by
a convex combination of five components.
1. Background 𝝓𝝓𝟎𝟎: words uniformly distributed in
every review.
2. Sentiment 𝝓𝝓𝒔𝒔: not aspect specific content such as
great, good, bad.
3. Item-specific 𝝓𝝓 𝒎𝒎: any term that appears only in the
item.
4. Aspect-specific 𝝓𝝓𝒂𝒂: words associated with specific
aspects. Music, sound, singing
5. Aspect-specific sentiment words 𝝓𝝓𝒂𝒂𝒂𝒂: to express
positive or negative sentiments.
Each of the language models is a multinomial
distribution.
111. Inference and Learning
( hybrid sampling/optimization)
• The goal is to learn the hidden factor vectors, aspects, and sentiments of
the textual content to accurately model user ratings and maximize the
probability of generating the textual content.
• Use Gibbs-EM, which alternates between collapsed Gibbs sampling and
gradient descent, to estimate parameters in the model.
• The Objective function consist of two parts :
1. The prediction error on user ratings.
2. The probability of observing the text conditioned on priors.
ℒ = � 𝜖𝜖−2 𝑟𝑟𝑢𝑢𝑢𝑢 − 𝑟𝑟̂𝑢𝑢𝑢𝑢
2
− log 𝑝𝑝 𝑤𝑤𝑢𝑢𝑢𝑢|Υ, Ω
𝑟𝑟𝑢𝑢 𝑢𝑢∈ℛ
111
112. Qualitative Evaluation
Aspect rating, Sentiment and Movie-Specific Words
To evaluate if the model is capable of interpreting the reviews correctly, the
learned aspect ratings are examined.
112
113. TPCF
• Topic creation and manual verification
and grouping.
• Faceted opinion extraction
• Based on the topics a user writes
about, create an individual interest
topic profile.
• Personalize the product rankings for
each user, based on the reviews that
are most relevant to her profile.
113
Recommendation Using Textual Opinions
Claudiu-Cristian Musat, Yizhong Liang, Boi Faltings, IJCAI 2013
114. Sentiment Utility Logistic Model
• ŒSULM builds user and item profiles
– Predicts overall rating for a
review
– Estimating sentiment utilities
of each aspect.
• SULM identifies the most
valuable aspects of future user
experiences.
• Uses Opinion Parser for aspect
extraction and aspect sentiment
classification.
• SULM estimates the sentiment
utility value for each of the aspects
k using the matrix factorization
approach.
• Tested on actual reviews across
three real-life applications.
114
Aspect Based Recommendations: Recommending Items with the Most Valuable
Aspects Based on User Reviews, by Konstantin Bauman, Bing Liu, Alexander
Tuzhilin, KDD’17
117. ConvMF
• Integrate CNN into PMF for the recommendation task.
• Item variable plays a role of the connection between PMF and CNN in
order to exploit ratings and description documents.
117
CNN
118. Optimization
• Use Maximum a posteriori to solve U, V, W
• When U, V are temporarily fixed, to optimzie W, use
backprop with given target variable 𝑣𝑣𝑗𝑗
• ConvMF+ used pretrained word embedding.
118
119. Deep Cooperative Neural Network
• (DeepCoNN) models users and items jointly.
• Two parallel convolutional neural networks to model user behaviors and
item properties from review texts.
• Utilizes a word embedding technique to map the review texts into a lower-
dimensional semantic space as well as keep the words sequence
information. Œ
• The outputs of the two networks are concatenated.
• The factorization machine captures their interactions for rating prediction.
Joint deep modelling of Users and Items using Review for
Recommendation, Lei Zheng, Vahid Noroozi, Philip S. Yu, WSDM 2017
119
121. User and Item Network
User Network
Lookup Layer:
• Reviews are represented as a matrix of
word embeddings.
• All reviews written by user 𝑢𝑢 are
merged into a single document
𝑉𝑉
𝑢𝑢
1: 𝑛𝑛
= 𝜙𝜙 𝑑𝑑
𝑢𝑢
1
⨁𝜙𝜙 𝑑𝑑
𝑢𝑢
2
⨁𝜙𝜙 𝑑𝑑
𝑢𝑢
3
⨁ ⋯ ⨁ 𝜙𝜙 𝑑𝑑
𝑢𝑢
𝑛𝑛
Item Network
Lookup Layer:
• All reviews written on item 𝑖𝑖 are
merged into a single document
CNN Layers:
Each neuron j in the convolutional layer
uses filter kj on a window of words size t.
Max-pooling layer
𝑜𝑜𝑗𝑗 = max 𝑧𝑧1, 𝑧𝑧2, ⋯ , 𝑧𝑧𝑛𝑛−𝑡𝑡+1
Output from multiple filters:
𝑂𝑂 = 𝑜𝑜1, 𝑜𝑜2, ⋯ , 𝑜𝑜𝑛𝑛𝑛
Fully connected layer: weight matrix W.
The output of the fully connected layer is
considered as features of U.
125. TransNets
• DeepCONN performance is lower when the
corresponding review is not in the training set.
• TransNets extends the DeepCoNN model by introducing
an additional latent layer representing an approximation
of the review corresponding to the target user-target
item pair.
• Regularize this layer at training time to be similar to
another latent representation of the target user’s review
of the target item.
125
128. Training TransNets
• During training, force the Source Network’s representation 𝑧𝑧𝐿𝐿 to be
similar to the encoding of rev𝑢𝑢𝑢𝑢 produced by the Target Network.
Repeat for each review:
• Step 1: Train Target Network on the actual review. Minimize L1 loss:
𝑟𝑟𝑢𝑢𝑢𝑢 − 𝑟𝑟̂𝑇𝑇
• Step 2: Learn to Transform: minimize a L2 loss computed between 𝑧𝑧𝐿𝐿 and
the representation 𝑥𝑥𝑇𝑇 of the actual review.
• Step 3: Train a predictor on the transformed input: Paramaters of 𝐹𝐹𝐹𝐹𝑠𝑠 are
updated to minimize a L1 loss : 𝑟𝑟𝑢𝑢𝑢𝑢 − 𝑟𝑟̂𝑆𝑆
At test time, TransNet uses only the Source Network.
128
129. Extended TransNet
• TransNet does not use user, item identity.
• Ext-TN learn a latent representation of the users and items,
similar to Matrix Factorization methods.
• Input to 𝐹𝐹𝐹𝐹𝑆𝑆: latent user and item representation concatenated
with the output of Transform layer.
129
132. D-Attn
• The local attention model (L-Attn)
selects informative keywords
from a local window.
• The global attention model (G-
Attn) ignores irrelevant words
from long review sequences.
• The outputs of L-Attn and G-Attn
are concatenated and followed by
a FC layer.
• Dot product of the vectors from
both networks estimate ratings.
132
135. Reviews Explain Ratings
• Reviews justify a user’s rating:
– by discussing the specific
properties of items (aspects)
– by revealing which aspects the
user is most interested in.
135
136. Hidden factor/topics
• Interpretable textual
labels for rating
justification.
• Topics obtained
explain variation in
rating/review data
Hidden factors and hidden topics: understanding rating dimensions with review text,
Mcauley et al, RecSys 2013.
136
137. EFM
137
Explicit Factor Models for Explainable Recommendation based on
Phrase-level Sentiment Analysis
Yongfeng Zhang, Guokun Lai, Min Zhang, Yi Zhang, Yiqun
Liu,Shaoping Ma, SIGIR 2014
138. User-Feature Attention Matrix X
• Suppose feature 𝐹𝐹𝑗𝑗is mentioned by
user 𝑢𝑢𝑖𝑖 for 𝑡𝑡𝑖𝑖𝑖𝑖 times.
𝑋𝑋𝑖𝑖𝑖𝑖 = �
0, 𝑖𝑖𝑖𝑖 𝑢𝑢𝑖𝑖did not mention 𝐹𝐹𝑗𝑗
1 + 𝑁𝑁 − 1
2
1 + 𝑒𝑒−𝑡𝑡𝑖𝑖𝑖𝑖
− 1 otherwise
• Rescale 𝑡𝑡𝑖𝑖𝑖𝑖 into the same range [1,N]
as the rating matrix A by
reformulating the sigmoid function.
Item-Feature Quality Matrix Y
138
For each item 𝑝𝑝𝑖𝑖,
extract the corresponding (𝐹𝐹, 𝑆𝑆′
) pairs form all its
reviews.
If feature 𝐹𝐹𝑗𝑗 is mentioned for 𝑘𝑘 times on item 𝑝𝑝𝑖𝑖, and
the average of sentiment of feature 𝐹𝐹𝑗𝑗 in those
mentions are 𝑠𝑠𝑖𝑖𝑖𝑖.
𝑌𝑌𝑖𝑖𝑖𝑖 = �
0, if item 𝑝𝑝𝑖𝑖 is not reviewed on 𝐹𝐹𝑗𝑗
1 + 𝑁𝑁 − 1
2
1 + 𝑒𝑒−𝑘𝑘.𝑠𝑠𝑖𝑖𝑖𝑖
− 1 otherwise
Rescaled to [1,N]
EFM
139. Factorization Model: Integrating Explicit and
Implicit Features
• Build a factorization model over user-
feature attention matrix X and item-
feature quality matrix Y.
minimize
𝑈𝑈1, 𝑈𝑈2, 𝑉𝑉
𝜆𝜆𝑥𝑥 𝑈𝑈1. 𝑉𝑉 − 𝑋𝑋
2
𝐹𝐹
+ 𝜆𝜆𝑦𝑦 𝑈𝑈2. 𝑉𝑉 − 𝑌𝑌
2
𝐹𝐹
• Introduce 𝑟𝑟𝑟 latent factors: 𝐻𝐻1 and 𝐻𝐻2.
• We use 𝑃𝑃 = 𝑈𝑈1 𝐻𝐻1 and 𝑄𝑄 = 𝑈𝑈2 𝐻𝐻2 to
model the overall rating matrix A.
• U1: explicit factors
• H1: Hidden factors
minimize
𝑃𝑃, 𝑄𝑄
𝑃𝑃. 𝑄𝑄 − 𝐴𝐴
2
𝐹𝐹
139
minimize
𝑈𝑈1, 𝑈𝑈2, 𝑉𝑉, 𝐻𝐻1, 𝐻𝐻2
� 𝑃𝑃. 𝑄𝑄 − 𝐴𝐴
2
𝐹𝐹
+ 𝜆𝜆𝑥𝑥‖ 𝑈𝑈1. 𝑉𝑉
− 𝑋𝑋‖2
𝐹𝐹
+ 𝜆𝜆𝑦𝑦 𝑈𝑈2. 𝑉𝑉 − 𝑌𝑌
2
𝐹𝐹
+ ⋯ �
140. Explanation Generation
• Template based explanation
and an intuitional word cloud
based explanation.
• Feature-level explanation for a
recommended item.
• Provide disrecommendations
by telling the user why the
current browsing item is
disrecommended.
140
You might be interested in <feature>
on which this product performs well
You might be interested in <feature>
on which this product performs poorly.
141. Trirank Approach to Aspect based
Recommendation
Offline Training
• Extract aspects from user reviews
• Build tripartite graph of User, Product, Aspect
• Graph Propagation: Label propagation from each vertex
• Machine learning for graph propagation
TriRank: Review-aware Explainable Recommendation by Modeling Aspects: Xiangnan
He, Tao Chen, Min-Yen Kan, Xiao Chen
141
143. Trirank – Review Aware Recommendation
• Uses User-Item-Aspect ternary relation
– Represented by heterogeneous tripartite graph
• Model reviews in the level of aspects
• Graph-based method for vertex ranking on tripartite graph accounting for
– the structural smoothness (encoding collaborative filtering and aspect
filtering effects)
Smoothness implies local consistency: that nearby vertices should not vary
too much in their scores.
– fitting constraints (encoding personalized preferences, prior beliefs).
• Offline training + online learning.
– Provide instant personalization without retraining.
143
144. Basic Idea: Graph Propagation
144
Label Propagation from the target user’s historical item nodes captures
Collaborative Filtering
145. Machine Learning for Graph Propagation
[He et al, SIGIR 2014]
145
Smoothness Regularizer: Nearby vertices should
not vary too much
� � 𝑦𝑦𝑖𝑖𝑖𝑖
𝑢𝑢𝑖𝑖
𝑑𝑑𝑖𝑖
−
𝑝𝑝𝑗𝑗
𝑑𝑑𝑗𝑗
2
𝑗𝑗∈𝑃𝑃𝑖𝑖∈𝑈𝑈
Fitting Constraints (initial labels):
Ranking scores should adhere to the initial labels.
� 𝑝𝑝𝑗𝑗 − 𝑝𝑝𝑗𝑗
0 2
𝑗𝑗∈𝑃𝑃
Optimization (Gradient Descent):
𝑝𝑝 = 𝑆𝑆𝑌𝑌 𝑢𝑢 + 𝑝𝑝0
𝑢𝑢 = 𝑆𝑆𝑌𝑌
𝑇𝑇
𝑝𝑝, where 𝑆𝑆𝑌𝑌 =
𝑦𝑦𝑢𝑢𝑢𝑢
𝑑𝑑𝑢𝑢 𝑑𝑑𝑖𝑖
• Input:
– Graph Structure (Matrix 𝑌𝑌)
– Initial labels to propagate (vectors
𝑝𝑝0
)
• Output:
– Scores of each vertex (vectors 𝑢𝑢, 𝑝𝑝)
147. Trirank Approach
• Graph propagation in the tripartite graph:
Initial labels should encode:
• Target user’s preference on aspects/ items/ users:
a0: reviewed aspects.
p0: ratings on items.
u0: similarity with other users (friendship).
147
148. Online Learning
Offline Training (for all users):
1. Extract aspects from user reviews
2. Build the tripartite graph model (edge weights)
3. Label propagation from each vertex and save the scores.
• i.e. store a |V|×|V| matrix 𝑓𝑓 𝑣𝑣𝑖𝑖, 𝑣𝑣𝑗𝑗 .
Online Recommendation (for target user 𝑢𝑢𝑖𝑖):
1. Build user profile (i.e., 𝐿𝐿𝑢𝑢 vertices to propagate from).
2. Average the scores of the 𝐿𝐿𝑢𝑢 vertices:
148
149. Explainability
• Collaborative filtering + Aspect filtering
– Item Ranking
– Aspect Ranking
– User Ranking
(Similar users also
choose the item)
(Reviewed aspects
match with the item)
149
150. FLAME
• FLAME learns users’ personalized preferences on different aspects from their
past reviews, and predicts users’ aspect ratings on new items by collective
intelligence.
• Propose a new model combining aspect-based opinion mining and
collaborative filtering to collectively learn users’ preferences on different
aspects.
• Preference diversity: For example, the food of the same restaurant might be
delicious for some users but terrible for others.
• To address the problem of Personalized Latent Aspect Rating Analysis, we
propose a unified probabilistic model called Factorized Latent Aspect ModEl
(FLAME), which combines the advantages of both collaborative filtering and
aspect-based opinion mining
150
FLAME: Probabilistic Model Combining Aspect Based Opinion Mining and
Collaborative Filtering, Yao Wu and Martin Ester, WSDM’15
151. Aspect Weights, Aspect Values
Produce more persuasive recommendation
explanations by the predicted aspect ratings and
some selected reviews written by similar users.
• user-1 likes to comment on the
Location and Sleep
• user-2 cares more about the Service,
Clean and Location.
151
152. Senti-ORG: Explanation Interface
• An explanation interface that emphasizes explaining the tradeoff
properties within a set of recommendations in terms of both their static
specifications and feature sentiments extracted from product reviews.
• Assist users in more effectively exploring and understanding product
space.
152
Explaining Recommendations Based on Feature Sentiments in Product Reviews
Li Chen, F Wang, IUI 2017
154. Future Trends
• Explanation Interfaces
• User control on
recommendations.
• Use with other knowledge
sources: context, ontology,
…
Use of reviews beyond
recommendations.
• Addressing Complex and
Subjective Product-Related
Queries, (McAuley WWW’16)
154