This document summarizes 5 papers being presented at the WWW2019 research track on Mobile and Ubiquitous Computing. It also discusses trends in deep learning apps on Android smartphones based on an analysis of over 16,500 apps from Google Play. Key findings include an increase in deep learning apps from 166 in June 2018 to 211 in September 2018, with CNNs being the most common neural network used.
This document summarizes 5 papers being presented at the WWW2019 research track on Mobile and Ubiquitous Computing. It also discusses trends in deep learning apps on Android smartphones based on an analysis of over 16,500 apps from Google Play. Key findings include an increase in deep learning apps from 166 in June 2018 to 211 in September 2018, with CNNs being the most common neural network used.
This document provides an overview and summary of the RecSys2018 conference that took place from October 2-7, 2018. It includes information about the keynote speakers, presented papers, and topics discussed. Some of the major topics covered include explainable recommendations, algorithmic confounding and homogeneity, offline evaluation of implicit feedback, calibrated recommendations, and using contextual bandits for artwork personalization at Netflix. The document also lists the dates and topics for each day of paper sessions and industry talks.
The document discusses Amazon Web Services' (AWS) machine learning and artificial intelligence services that were announced or highlighted at the 2018 re:Invent conference. It provides an overview of 1) AI services like Amazon Personalize and Amazon Forecast, 2) machine learning frameworks and infrastructure like Amazon SageMaker, Elastic Inference, and EC2 instances, and 3) machine learning services like Amazon Rekognition and Amazon Comprehend. It also summarizes new services announced at re:Invent like Amazon SageMaker Ground Truth, AWS Marketplace for ML/AI, and Amazon SageMaker Neo.
Orion an integrated multimedia content moderation system for web servicescyberagent
This document describes Orion, an integrated content moderation system developed by CyberAgent to moderate user generated content on their various social networking services and apps. The system combines automatic filtering using over 300 filters with manual review by human operators. It processes millions of posts daily. Since deploying Orion, the proportion of content requiring manual review has decreased by up to 5 times, and criminal activity on the company's services has sharply declined. The system provides reporting and monitoring to ensure a high quality of moderation.
Orion an integrated multimedia content moderation system for web servicescyberagent
This document describes Orion, an integrated content moderation system developed by CyberAgent to moderate user generated content on their various social networking services and apps. The system combines automatic filtering using over 300 filters with manual review by human operators. It processes millions of posts daily. Since deploying Orion, the percentage of content requiring manual review has decreased by up to 5 times, and criminal activity on the company's services has sharply declined. The system provides reporting to monitor operator performance and ensure high quality moderation.
5. イントロダクション
Search 分の分野概観
5
小分野 論文リスト
検索一般
● Leveraging Fine-Grained Wikipedia Categories for Entity Search
● Subgraph-augmented Path Embedding for Semantic User Search on
Heterogeneous Social Network
● Ad Hoc Table Retrieval using Semantic Similarity
対話検索・
クエリ提案
● Query Suggestion with Feedback Memory Network
● Conversational Query Understanding Using Sequence to Sequence Modeling
Hashing ● Scalable Supervised Discrete Hashing for Large-Scale Search
プライバシー
● Privacy and Efficiency Tradeoffs for Multiword Top K Search with Linear Additive
Rank Scoring
データ整備
● StaQC: A Systematically Mined Question-Code Dataset from Stack Overflow
● Strategies for Geographical Scoping and Improving a Gazetteer
検索行動
分析
● “Satisfaction with Failure” or “Unsatisfied Success”: Investigating the Relationship
between Search Success and User Satisfaction
● Search Process as Transitions Between Neural States
※Search / Mining の分類や
小分野は発表者の主観による
分類です
6. イントロダクション
Mining 分の分野概観
6
小分野 論文リスト
機械学習・
アルゴリズム
● Parabel: Partitioned Label Trees for Extreme Classification with Application to
Dynamic Search Advertising
● Learning from Multi-View Multi-Way Data via Structural Factorization Machines
● Online Compact Convexified Factorization Machine
● Learning on Partial-Order Hypergraphs
● Manifold Learning for Rank Aggregation
レビュー
分析
● A Sparse Topic Model for Extracting Aspect-Specific Summaries from Online
Reviews
● Neural Attentional Rating Regression with Review-level Explanations
行動分析
● Detecting Crowdturfing “Add to Favorites” Activities in Online Shopping
● Understanding and Predicting Delay in Reciprocal Relations
その他
● Finding Subcube Heavy Hitters in Analytics Data Streams
● Joint User- and Event- Driven Stable Social Event Organization
● TEM: Tree-enhanced Embedding Model for Explainable Recommendation
● Hierarchical Variational Memory Network for Dialogue Generation
※Search / Mining の分類や
小分野は発表者の主観による
分類です
7. イントロダクション
ざっくり概要 – Search 分 (1/2)
● Leveraging Fine-Grained Wikipedia Categories for Entity Search
○ クエリのメイン語 headword とそれ以外 modifier に注目した category matching で精度 ↑
● Subgraph-augmented Path Embedding for Semantic User Search on Heterogeneous Social
Network
○ 色々なタイプの関係 (e.g. schoolmates 等) があるネットワーク (heterogeneous social
network) における「特定ユーザ」と「関係」を入力としたユーザ検索を実現
● Ad Hoc Table Retrieval using Semantic Similarity
○ クエリから表を検索。クエリと表を同じ embedding space に置いてマッチする
● Query Suggestion with Feedback Memory Network
○ 検索結果ページでのクリック履歴から、次にクエリされそうなフレーズを予測 w/ seq2seq (を
改変したモデル)
● Conversational Query Understanding Using Sequence to Sequence Modeling
○ 文脈を考慮できる stateful な対話検索が目的。context も利用した seq2seq で発話生成
● Scalable Supervised Discrete Hashing for Large-Scale Search
○ 教師あり hashing。大規模データ対応・計算過程で discrete constraints に違反しないと言う
好特性
7
8. イントロダクション
ざっくり概要 – Search 分 (2/2)
● StaQC: A Systematically Mined Question-Code Dataset from Stack Overflow
○ accepted answer の複数のコード片から「それ単体で解決できるコード」を systematic に判
定
● Strategies for Geographical Scoping and Improving a Gazetteer
○ 複数の地理情報 DB (gazetteer) を統合。各 DB が異なるデータタイプ(点、範囲)だったり不
正確なデータでも、うまく統合できる確率的なモデルを提案
● Privacy and Efficiency Tradeoffs for Multiword Top K Search with Linear Additive Rank
Scoring
○ searchable encryption。従来研究でまだだった ranking (i.e. top-k search) を実現
● “Satisfaction with Failure” or “Unsatisfied Success”: Investigating the Relationship between
Search Success and User Satisfaction
○ 「ユーザが検索に満足してても、実際には誤った情報で満足している」など、ユーザ満足度と
検索の成功の間にあるギャップについて詳しく調査
● Search Process as Transitions Between Neural States
○ 検索行動が4つの過程からなるとし、各過程で脳活動がどのように異なるか・共通しているか
を fMRI で調査
8
9. イントロダクション
ざっくり概要 – Mining 分 (1/2)
● Parabel: Partitioned Label Trees for Extreme Classification with Application to Dynamic
Search Advertising
○ ラベル数が非常に多い分類問題(extreme classification)を同精度で 600-900 倍早く学習で
きる手法を提案。似たようなラベルをまとめて (label trees) 1-vs-All 爆発しないように工夫
● Learning from Multi-View Multi-Way Data via Structural Factorization Machines
○ 色々な種類の素性をそのまま使うとベクトル大き過ぎとか問題 → 潜在空間にうまく落とす手
法を提案
● Online Compact Convexified Factorization Machine
○ FM を頑張ってオンライン凸最適化問題にしてオンライン化。分類・回帰とも精度向上
● Learning on Partial-Order Hypergraphs
○ グラフベース学習手法を POH (hypergraph を拡張したデータ構造) に適用できるように
● A Sparse Topic Model for Extracting Aspect-Specific Summaries from Online Reviews
○ レビュー文書群から商品の評価視点がうまく取り出せるような topic model を提案
● ☆ Neural Attentional Rating Regression with Review-level Explanations
○ レビュー点数を、レビュー有用度を考慮して推定。レビュー点数推定精度の向上に加え、有用
度予測では「有用とした人数」よりも高い精度を実現
9
10. イントロダクション
ざっくり概要 – Mining 分 (2/2)
● ☆ Detecting Crowdturfing “Add to Favorites” Activities in Online Shopping
○ 「お気に入りに追加」しまくって順位を上げるタイプのスパムを分析・検出
● Understanding and Predicting Delay in Reciprocal Relations
○ Tumblr で「フォロー返し」するまでの時間を分析 + 分析に基づいた予測手法を提案
● Finding Subcube Heavy Hitters in Analytics Data Streams
○ 高次元・ストリーミングデータに対応可能な heavy hitters 抽出手法を提案
● Manifold Learning for Rank Aggregation
○ 従来の rank aggregation では文書間の独立性が前提であったが、manifold で非独立性を考
慮
● Joint User- and Event- Driven Stable Social Event Organization
○ ユーザ–イベント選好とユーザ間選好を考慮したヒューリスティックにより効率的に Social
Event Organization 問題を解く
● TEM: Tree-enhanced Embedding Model for Explainable Recommendation
○ 理由が説明可能な推薦。GBDT で素性 (cross feature) 抽出 → embed
● Hierarchical Variational Memory Network for Dialogue Generation
○ 階層構造と variational memory network を seq2seq モデルに導入。長文での返答が可能
に
10