まとめページへのリンク

Research Track Session RT17: Interpretability

Learning Interpretable Metric between Graphs: Convex Formulation and Computation with Graph Mining

Tomoki Yoshida (Nagoya Institute of Technology); Ichiro Takeuchi (Nagoya Institute of Technology & National Institute for Material Science & RIKEN Center for Advanced Intelligence Project);Masayuki Karasuyama (Nagoya Institute of Technology & National Institute for Material Science & Japan Science and Technology Agency)

Intro

Metric on Graph

Machine learning classifiers implicity/explicity require distance function
However neighbor classifier, kernel method, ...

Existing approach

graph kernel, deep-net

propose: Interpretable graph metric learning(IGML)

Subgraph Representation
Graph can be seen as a set of subgraphs

disadvantage

computationaly intractable
not optimal for classification

proposed: supervised learning of metric on graphs with sparse penalty
Optimizing Distance function:

parameterized subgraph-based distance function
pairwise loss funxction
pairwise loss + sparse penalty
convex problem
identify discriminative subgraph
a huge number of variables needed to be optimaized
subest-based optimization
- select small number of subgraph
- optimizing with respect to MH only for H<F

pruning on graph Mining Tree
Tree Traverse with Pruning

On Identified Subgraph Space

observing estimated subgraph weights
Applying smoe ML algorothm such as decision tree

conclusion

IGML: learning metric through labeled data
discriminative
convex
sparse
identify discriminative subgraphs

Axiomatic Interpretability for Multiclass Additive Models

Xuezhou Zhang (University of Wisconsin-Madison); Sarah Tan (Cornell University); Paul Koch (Microsoft Research); Yin Lou (Ant Financial); Urszula Chajewska (Microsoft); Rich Caruana (Microsoft Research)

is binary logistic regression interpretable?
YES, given that the features themselves are interpretable.

Now, what happens if 3 class.
Diabetes, Type 1, Type 2, normal
remedy the inconsistency between

how a model is interpretable?

model: Generalized additive model(GAM)
three classes
one feature
the generalized linear model(GLM)
generalized additive model(GAM)
Different representations of the same model

the functions are informative?
the axziom of monotonicity
problem always have unique global optimal solution
the process of construction takes almost zero computation cost

Interbility in Action on Infant Mortality
"The risk of all caused of death increase as birthweight decrease" wrong

InterpretML Demo: Global interpretation
InterpretML Demo: local interpretation

Incorporating Interpretability into Latent Factor Models via Fast Influence Analysis

Weiyu Cheng (Shanghai Jiao Tong University); Yanyan Shen (Shanghai Jiao Tong University); Linpeng Huang (Shanghai Jiao Tong University); Yanmin Zhu (Shanghai Jiao Tong University)

background

LFM is SOTA for CF.
uncertainty of the model
Influence functions

goal of influence function is to estimate the influence of training points on a model's predictions
Interpreting LFMs with influence functions

influence analysis for MF
fast influence analysis for MF
data: yelp and movielens

experiments - effectiveness
running time of IA end IA for MF

conclusion

explanation method based on influence function
...

Improving the Quality of Explanations with Local Embedding Perturbations

Yunzhe Jia (University of Melbourne); James Bailey (University of Melbourne); Kotagiri Ramamohanarao (University of Melbourne); Christopher Leckie (University
of Melbourne); Michael E. Houle (National Institute of Informatics)

importance of the generation of the appropriate neighborhood for local explanation methods

logical reasoning behind the model of a prediction
Global explanation
local explanation for the prediction

Current framework(LIME)
step1: generate synthetic neighborhood of x
step2: label the synthetic data
step3: generate the local explanation from synthetic neighborhood

"good" neighbors vs "bad" neighbors

LID a possible measurement of the quality neighbors
LID access the intrinsic dimensionality of the local data submanifold around given instance
Local intrinsic dimensionality(LID) - Intuition

neighbors local uniform 1D
neighbors local uniform 2D
chaotic and far neighbors distance are .. locally ~7D

what is the local intrinsic dimensional (LID) at a point X?
Estimation of Local intrinsic dimensionality
Locally intrinsic dimensionality (LID) of different synthetic sets.
local explanation extraction framework
LEAP: local embedding aided perturbation

experiment

explanation quality
explanation fidelity
methodology

Comparison of the quality of explanations generated by different perturbation
locality analysis

Summary:

synthetic neighborhood
generation is crucial process for local explanation extraction
local intrinsic dimensionality can be used as a basis for generating better quality neighboorhoods
our proposed method can be easily integrated with existing frameworks for explanation such as LIME

Log2Intent: Towards Interpretable User Modeling via Recurrent Semantics Memory Unit

Zhiqiang Tao (Northeastern University); Sheng Li (University of Georgia); Zhaowen Wang (Adobe Research); Chen Fang (ByteDance AI Lab); Longqi Yang (Cornell University); Handong Zhao (Adobe Research); Yun Fu (Northeastern University)

baskground

User moding learnins generic user representations for downstream tasks
Log-trace data

3 goal

understand user behaivior
learn user represantation
provide personal asistance

challenge

how to mine user's temporal characteristic(usage habit) from long log history
how to regularize and further interpret the unstructured user log data.

tutorial dataset

collect 2034 tutorials with various

Insight in Tutorial

structured inside tutorial
key: the same software annotation action always corresponds to different sentences annotations due to different temporal context

session-to-session modeling

temporal encoder
semantics encoder + memory network
user action decoder

structured and unstructured User data

recurrent semantics memory unit(RMSU)

each action <> one memory unit
memory parameterization
recurrent memory addressing & reading

semantic attention mechanism

dataset:

userlogdataset
from photoshop
tutrial dataset
behance dataset
validation criteria

our approach effectively annotate tutorial log sequence via attention mechanism

log interpretation
user interest detection
task: User action Prediction

summarization

user log-trace data enrich user representation
log-term temporal dependency captures user habit
user log modeling enables personalization
recurrent semantic attentions on tutorial memory provide interpretability

Interpretable and Steerable Sequence Learning via Prototypes

Yao Ming (Hong Kong University of Science and Technology); Panpan Xu (Bosch Research North America); Huamin Qu (Hong Kong University of Science and Technology); Liu Ren (Bosch Research North America)

motivation

interpretability
steerablity

case-based reasoning
predicts a new case by consulting similar cases in the past
prototype learning
ProtoSequence Network(ProSeNet)

Prosenet - explanable prediction

learning readable prototype

jointly training a sequence decoder
prototyping projection
project each prototype to its closest embedding of the training sequences
incorporating expert knowledge.

inductive biases for interpretability.

diversity
sparsity
simplisity

challenges:

duplicate prototypes
diversity & sparsity
verbosity -> simplicity

beam-search

experiments

ECG heartbeat classification (time series)
predictive diagnostic of vehicle faluts
Protein Sequence Classification

user experiments

conclusions

sequence learning techniques that produce interpretable and prototypes
experiments and case studies on various real-world sequence classification tasks
user experiment demonstrating the benefits of human-involvements in interpretability