Social and ecological factors influencing movement and organizational patterns in sheep Habiba, Caitlin Barale, Ipek Kulahci, Rajmonda Sulo and Khairi.

Social and ecological factors influencing movement and organizational patterns in

sheep

Habiba, Caitlin Barale, Ipek Kulahci,

Rajmonda Sulo and Khairi Reda

complementary approaches from ecology and computer science

How do we identify the key individuals in a group?

•Personality 'types’•Group social dynamics•Group movement patterns

Approaches

Ecology• Direct observation

• Information on:• Proximity

• Individual movement patterns

• Behavior

• Interactions

• Individual state

Computer Science• Remote sensing

• Information on:• Proximity

• Group-level movement patterns

• Clustering

• Velocity

• Turning angle

How can we use concepts from computer science to study animal behavior?

•Association visualizations

•Static networks

•GPS clustering algorithms

•Dynamic networks

•Social network analysis

•Agent-based modeling

GPS Data

Data Collection

General Statistics Map with GPS readings overlaid (Khairi)

Personality Analysis

• PCA to generate personality ‘types’

• Static social networks

• Association visualizations

Push/pullPush/pushedFollowed pullsPull/followerPush/follower

Lactation state

Join time in the morningsGroup joiningAlone grazingLocation- frontLocation- backLocation- edgeLocation- front during movement

Personality components

Push/pullPush/pushedFollowed pullsPull/followerPush/follower

Lactation state

Join time in the morningsGroup joiningAlone grazingLocation- frontLocation- backLocation- edgeLocation- front during movement

Personality components

PCA component 1 - spatial

PCA component 2 - behavioral

PCA 2Push/pull

Push/pushedFollowed pulls

PCA1Location- backLocation- edge

Location- front during movement

Personality scores

Group movement can be initiated either by pulling or pushing

Pull-follow and displacement networks

What does the network look like from the perspective of individuals?

Social Network metrics

DegreeBetweennessClosenessCluster coefficientEigenvector

Key individuals – pull network

Size: # of times an individual pulls

Color: PCA category

Lactating

Not-lactating

Key individuals – push network

Size: # of times an individual is pushed

Color: PCA category

Lactating

Not-lactating

Key individuals – nearest neighbor network

Size: # of times two individuals are

neighbors while aligned-grazing

Color: PCA category

Lactating

Not-lactating

Moving from static networks to dynamic ones

Static Vs Dynamic Networks

T = i

…

T = i+1 T = i+2 T = i+3 T = i+4

…

GPS clustering

What is an edge ?

Use k-means algorithm to spatially cluster gps readings. Any two sheep in the same cluster have an edge between them.

Use the Simple Structure Index (SSI) to determine the number of optimal clusters in each time interval

Using different graph theoretic measures, the goal is to identify sampling rates at which the measures contain minimal noise while still being informative of the underlying dynamics in the network.

What is the right temporal resolution?

Apply information theoretic concepts to quantify noise and information and compute the best trade-off across different sampling rates.

Algorithm TWIN

For w < wmax

* Compute the interaction network.

* Compute the different structure measures such

as density, and centrality, etc.Quantify the amount of noise and information inherent

in the network measure.Compute a measure of goodness of fit based on a

trade-off between the two.

TWIN Results

Optimal interval = 3min Optimal interval = 5min

GPS data - issues

Data Quality

Discrepancy of sensor readings on the same sheep Missing Values

Static Vs Dynamic Networks

T = i

…

T = i+1 T = i+2 T = i+3 T = i+4

…

Diffusion Process

• Independent Cascade Diffusion Model

pr

prprpr

pr

Diffusion Process

• Linear Threshold Diffusion Model

өi өj

өk

өl

bi,j

bj,lbi,l

bk,l

bi,k

Independent cascade

Independent cascade

Independent cascade

Text

Linear threshold model

Community Identification

•A dynamic community is a subset of individuals that stick together over time.

•NOTE: Communities ≠ Groups

5 4 32 1

5

4

5

4

1

4

12 3 4

5 2

2 3

5 2 3 1

t=1

t=2

t=3

t=4

t=5

Approach: Assumptions

• Individuals and groups represent exactly one community at a time.

• Concurrent groups represent distinct communities.

Desired

Required

•Conservatism: community affiliation changes are rare.•Group Loyalty: individuals observed in a group belong to the same community.•Parsimony: few affiliations overall for each individual.

Approach: Color = Community

Valid coloring: distinct color of groups in each time step

Dynamic structure measures

Dynamic structure measures

• Interesting behavioral aspects to model– Social foraging• Alignment• Sub-group formation from local interactions• “Exploratory” random walks

– Foraging strategy• How do sheep find good food in patchy landscape?• How does information about food resources spread in the

herd?• Competition

– Pull and push

Agent-based model

• Interesting behavioral aspects to model– Social foraging• Alignment• Sub-group formation from local interactions• “Exploratory” random walks

– Foraging strategy• How do sheep find good food in patchy landscape?• How does information about food resources spread in the

herd?• Competition

– Pull and push

Agent-based model

– Sheep watch closest N neighbors and react to their behavior

– No personality variance among agents– Agents assumed to be “hungry”. They graze

continuously without rest– Agents have 3 states: GRAZE, WALK, PANIC– Low vegetation quality and patchy landscape;

agents need to move frequently to maximize gain

Assumptions

• Agents switch back and forth between grazing and social walking

• State transition governed by a Gaussian timer– Grazing bouts: mean=6 sec, SD=2 sec– Social walking: mean=2-8 sec (depending on

isolation), SD=2.5 sec

• Certain events cause agents to switch state– Agents minimize grazing and take longer walks if

they are isolated from subgroup

Timing

Local rules

Collision avoidance Cohesion Alignment

Random biased-walksP = 0.05 (when walking)

Pulling

Agents that find themselves drifting away turn back and rejoin the subgroup

Subgroup awareness

Demo

Conclusions

Sheep 15 – an influential individual??

• Central in the observational networks– High betweeness and degree– Highest number of successful pulls

• Role in correlational plots– Often an outlier– Makes trends more significant

• GPS – doesn’t stand out• Why?? Missing data…

Sheep 19 – a loner??

• Static networks

• Not well connected– >>>

• Animation• Little interaction with others in herd

• GPS – consistently low spread values

Are social network measures of individuals consistent across the three networks?

Push & pull networks: + correlation (betweenness & closeness)Push & nearest neighbor networks: - correlation (eigenvector &

closeness)

Network comparison

Is there consistency between observational subgroups and GPS clustering data?

Is there consistency between personality analysis obtained from observations and that obtained from GPS data aggregated at the “optimal” time interval?

Bite rate correlates with relative position

p < 0.005, R2 = 0.47

CS future directions

•Community structure evolution.

•Compare the observational data with “unsupervised” data collection and interpretation methods.

•Network comparison of different population to discover their social dynamics independent of their environment and other factors.