Deep Learning-based Search and Recommendation systems using - PowerPoint PPT Presentation

Deep Learning-based Search and Recommendation systems using TensorFlow Abhishek Kumar Dr. Vijay Agneeswaran MARCH 06, 2018 Strata Conference – San Jose ( 2018 ) S A P I E N T R A Z O R F I S H 2 2 S A P I E N T R A Z O R F I S H

Session Logistics 1. Access the work environment using the following link [ ] 2. Code and presentation available at link : [ ht 8 ] http://bi bit.ly/s /strata-dl dl-ca ca-20 2018 3. Connecting to the speakers [ Please send introductory note in LinkedIn invite ] 1. Abhishek Kumar ( ht http://bit.ly/kum kumara rabhi hishe hek @me meabhishekkuma mar ) 2. Dr. Vijay Agneeswaran ( ht http:/ ://bi bit.ly/v /vijaysa @a_ a_vi vijay aysrinivas vas ) 4. Don’t forget to tweet #str strata tadata ta S A P I E N T R A Z O R F I S H 3

About the Speaker A B H I S H E K K U M A R Senior Data Scientist , Sa SapientRa tRazorfish Masters from Un Univers ersity of of Califor ornia, Be Berkeley Pluralsight Pl t Aut Autho hor • Doing Data Science with Python • R Programming Fundamentals • Machine Learning with ENCOG • Currently authoring : “ Deploying Machine Learning Models with Tensorflow Serving ” S A P I E N T R A Z O R F I S H 4

About the Speaker D R . V I J A Y A G N E E S W A R A N Senior Director and Head of Data Science, Sa SapientRa tRazorfish MS (Research ) & PhD , IIT IIT Madras Post doctoral research fellowship, LS LSIR La Labs bs Professional member : AC ACM, IEEE (Seni nior) 4 Full US Patents and multiple publications 4 (including IE IEEE journals) Regular Speaker @ O’ O’Reilly Str Strata ta co confer eren ence ce S A P I E N T R A Z O R F I S H 5

Audience Profiling 1. Machine Learning? 2. Deep Learning? 3. Search and Recommendation Systems? 4.Tensorflow? S A P I E N T R A Z O R F I S H 6

Session Agenda 4 Levels of Learning FOUNDATION [ 30 MINS ] Search [ 1 HR ] Recommendation and LTR [ 1 HR ] PRODUCTION [ 30 Mins ] 1. High Level Overview For 1. Embedding 1. Embedding in RecSys 1. TF in Production : Training Problem Space [ Search, & Inference 2. Demo : Embedding in TF 2. Demo : Build DL based Recommendations, RecSys with Explicit 2. Tensorflow Serving Learning to Rank ] 3. Image Search using Feedback using TF CovNet 3. RecSys Architecture 2. Deep Learning Primer 3. Demo : Build hybrid 4. Demo : Image Search in 3. Why Tensorflow for Deep RecSys using TF TF Learning ? 4. Learning on Rank 5. Demo : Build DL based RecSys with Implicit Feedback using TF BREAK S A P I E N T R A Z O R F I S H 7

By the end of this session… 1. You will have basic foundation of Deep Learning. 2. You will have good understanding of Recommendation Systems, Search and Ranking Systems 3. You will be able to transform the concepts and build DL models using Tensorflow 1. Deep learning based Image retrieval system 2. Deep learning based hybrid RecSys on explicit feedback 3. Deep learning based RecSys and Learning to Rank model on implicit feedback 4. You will have high level idea to take the lab scale solution to a production ready system S A P I E N T R A Z O R F I S H 8

Quick Chat with Your Neighbor 1. Introduce yourself to your neighbor 2. What they are looking to learn from the tutorial S A P I E N T R A Z O R F I S H 9

Probl Problem em Spac pace e : Search and Recommendation S A P I E N T R A Z O R F I S H 10

We now live in the connected age 1900 1960 1990 2008+ Manufacturing Distribution Information Connected Economy Economy Economy Economy Mass manufacturing Global connections and Connected PCs and iPhone and Facebook makes industrial transportation systems supply chains mean those launch in 2007/8 powerhouses successful make distribution key that control information heralding a new era in flow dominate transparency, empowerment, and experimentation. “A customer can have “Strategy “The great challenge… “The customer is the a car painted any color he is…globalization, taking is to make productive center of your universe.” wants as long as it's black” your products around the tremendous new the world; be the low- resource, the knowledge cost producer” worker. ” *Adapted from Forrester Research “Age of the Customer” graphic S A P I E N T R A Z O R F I S H 11

The Connected Consumer is in charge Empowered and demand transparency Value experiences over most things Embrace and seek new companies to engage with Demand personalization and real-time relevancy S A P I E N T R A Z O R F I S H 12

Research proves that consumer experience does matter 81% 95% 86% 75% CU STO ME R S D ATA CU STO ME R S SH O PPING TIME IS SPE NT O N PR O D U CT W ITH IN OR G ANIZATION D E MAND IMPR OVE D SAID TH AT D ISCOVE R Y & R E SE AR CH R E SPO NSE TIME R E MAINS U NTAPPE D PE R SO NALIZAT IO N H AS H AD O NLINE B Y 5 0% CU STO ME R S SO ME IMPACT O N PU R CH ASING D E CISIO N Source : http://www.nextopia.com/wp-content/uploads/2015/01/personalization-ecommerce-infographic.png https://blog.hubspot.com/blog/tabid/6307/bid/23996/Half-of-Shoppers-Spend-75-of-Time-Conducting-Online-Research-Data.aspx http://possible.mindtree.com/rs/574-LHH-431/images/Mindtree%20Shopper%20Survey%20Report.pdf http://www.getelastic.com/using-big-data-for-big-personalization-infographic/ S A P I E N T R A Z O R F I S H 13

Problem Space : Search Search Engines Challenges - How to represent text, images, Search term.. audios - TF-IDFs? - Metadata for binary? User + Interactions - Search in other languages? Personalized Search - Search Quality - Well-ranked results - By providing better search results, Netflix estimates that Engine it is avoiding canceled Filtered Results Re-Ranked Results subscriptions that would Indexing Recency reduce its revenue by $1B Similarity Calculation annually. [ Link ] Impression Discounting S A P I E N T R A Z O R F I S H 14

Problem Space : Recommendation Challenges Recommendation Engines - How to represent users and items? - How to build hybrid systems with User both interactions ( collaborative ) Item and user/item metadata ? - How to use dynamic user behaviors? - How to use implicit ( view, share ) feedback ? Engine Recommended Re-Ranked Results Results Interactions Results Diversity Recency Impression Discounting Other Users Interactions S A P I E N T R A Z O R F I S H 15

Why Deep Learning for Search and Recommender System? - Direct content Feature extraction instead of metadata - Text, Image, Audio - Better representation of users and items for Recsys - Hybrid algorithms and heterogeneous data can be used - Better suited to model dynamic behavioral patterns and complex feature interactions S A P I E N T R A Z O R F I S H 16

Deep Learning Primer S A P I E N T R A Z O R F I S H 17

What is Deep Learning ? Artificial Class of machine learning algorithms Intelligence - That uses hierarchy of non-linear processing layers and Machine complex model structures Learning - Layers learn to represent different representation of data Deep - Higher level features are constructed from lower level Learning abstract features - Trendy name for “Neural Networks with deep layers” Enabling Big Data Tech S A P I E N T R A Z O R F I S H 18

Simple Neural Network With 2 Layers OUTPUT LAYER INPUT LAYER Limitation : Can learn only linear relationship S A P I E N T R A Z O R F I S H 19

Simple Neural Network with At Least One Hidden Layer Universal Approximator OUTPUT LAYER INPUT LAYER HIDDEN LAYER S A P I E N T R A Z O R F I S H 20

Neural Network Training : Backpropagation Output calculation Error Calculation Error Propagation and Weights Update ( using Gradients) OUTPUT LAYER INPUT LAYER HIDDEN LAYER S A P I E N T R A Z O R F I S H 21

What Changed Now ? - Mor More data - More complex models need more data to avoid overfitting - Deep learning models have higher VC dimension - Co Computin ing Power - Computing power have increased significantly - Specialized hardware such as GPUs and TPUs - Re Research Breakth through - Hinton’s work on layerwise training led a new paradigm to train deep networks - Non-saturating activation functions ( variation of ReLUs ) - Dropouts helped to achieve regularization easily - Adaptive learning rate helped to avoid problems of local minima and led to better convergence S A P I E N T R A Z O R F I S H 22

Popular Neural Network Architectures : Deep Feed forward MULTIPLE HIDDEN OUTPUT LAYER INPUT LAYER LAYERS S A P I E N T R A Z O R F I S H 23

Popular Neural Network Architectures : Convolution Neural Network ( CovNet) S A P I E N T R A Z O R F I S H 24 Image Credit : https://towardsdatascience.com/applied-deep-learning-part-4-convolutional-neural-networks-584bc134c1e2

Convolution Neural Network ( CovNet) : Components CONVO VOLUTION Mathematical Operation on two sets of information NON-LI NO LINEARITY Feature Filter / Map Input PO POOLING Kernel S A P I E N T R A Z O R F I S H 25 Image Credit : https://towardsdatascience.com/applied-deep-learning-part-4-convolutional-neural-networks-584bc134c1e2