result.md

ReimplementationOfTwoPapers.

Overview

• This repo aims to reimplement Deepmove and MobilityUpperBoundPrediction project using another framework or pure Python.

• Weekly summary will be uploaded in weekly_summary folder.

  • week1 summary
  • week2 summary
  • week3 summary
  • week4 summary
  • week5 summary

• You can find the report here.

Results

---

• Deepmove

  I got these results from reimplemented tensorflow models as follows.

  model_mode	L2	attn_type	clip	dropout	hidden_size	learning_rate	loc_size	rnn_type	tim_size	uid_size	original_acc	my_acc
  markov											0.082	0.082
  simple	1.00E-06	dot	5	0.3	500	0.0001	500	LSTM	10	40	0.09587167	0.082337454
  simple_long	1.00E-05	dot	5	0.5	200	0.0007	500	LSTM	10	40	0.117923069	0.082788173
  attn_avg_long_user	1.00E-05	dot	5	0.2	300	0.0007	100	LSTM	10	40	0.133689175	0.135096371
  attn_local_long	1.00E-06	dot	2	0.6	300	0.0001	300	LSTM	20	40	0.145342384	0.150585050

  Besides, I also plot user 4's predicted and real trajactory under four different models using matplotlib.

  
  

  You can find my results in this folder.

---

• MobilityUpperBoundPrediction

  Original results

  

  My results

  

  You can find more details in this folder.

---

• Cross dataset test

  First, I applied DeepMove dataset to MobPrediction algorithm, and got the following table.

  	training set
  number of failed personIDs	355
  possibility using DL method	0.4857
  possibility using RL method	0.3307

  Which means we may achieve better results for trajactory prediction if using another algorithm rather than DeepMove algorithm.

  Then, I applied two datasets from DataGeolife , namely, spatial resolution 40.54km^2 with temporal resolution 0:05:00 (dataset1) and spatial resolution 618m^2 with temporal resolution 1:00:00 (dataset2) to DeepMove algorithm. Then I got the following table.

  	dataset1	dataset2
  possibility using DL method	0.9765	0.6201
  possibility using RL method	0.9688	0.2726
  possibility using DeepMove	0.8750446	0.04101216

  DeepMove key parameters for dataset1 and dataset2:

    --loc_emb_size=100 --uid_emb_size=40 --tim_emb_size=10 --hidden_size=300 --epoch_max=50 --dropout_p=0.6 --learning_rate=0.001 --clip=2  --model_mode=simple --min-lr=1e-5 --use_geolife_data=True
  

  It seems DeepMove has a great precise prediction since predicting trajectory is so hard.