JGibbLDA的输出文件

五个输出文件:

---

• model-final.towords
  每个topic下面的words和words分布,并按分布排序:

1Topic 0th:
	bill 0.005843543826578699
	lai 0.003958529688972668
	seventh 0.0020735155513666352
	immedi 0.0020735155513666352
	anaheim 0.0020735155513666352
	concern 0.0020735155513666352
	month 0.0020735155513666352
	american 0.0020735155513666352
	decision 0.0020735155513666352
	risk 0.0020735155513666352
	demis 0.0020735155513666352
	maintain 0.0020735155513666352
	rose 0.0020735155513666352
Topic 1th:
	photo 0.0040191387559808615
	seri 0.0040191387559808615
	fashion 0.0040191387559808615
	subject 0.0021052631578947372
	left 0.0021052631578947372
	briefli 0.0021052631578947372
	discov 0.0021052631578947372
	hick 0.0021052631578947372
	line 0.0021052631578947372
	lip 0.0021052631578947372
	whom 0.0021052631578947372
	beij 0.0021052631578947372
	man 0.0021052631578947372
	xiaop 0.0021052631578947372
	highli 0.0021052631578947372
	california 0.0021052631578947372
Topic 2th:
	enjoi 0.0039069767441860465
	fan 0.0039069767441860465
	shuttle 0.0039069767441860465
	return 0.002046511627906977
	little 0.002046511627906977
	plai 0.002046511627906977
	songwrit 0.002046511627906977
	sai 0.002046511627906977
	debat 0.002046511627906977
	depart 0.002046511627906977
	disagre 0.002046511627906977
	iii 0.002046511627906977
	jan 0.002046511627906977
	mere 0.002046511627906977
	relat 0.002046511627906977
	stress 0.002046511627906977

• model-final.phi
  topic -words distributions
  是上面towords文件的原始数据文件,未排序,是数组里面直接保存的数据.

• model-final.theta
  每个doc下的topic分布,无排序

• model-final.assign
  每个word所属的topic代号:
  0:35
  0代表word的id,35代表第36个topic

0:35 1:30 1:52 2:11 2:11 2:11 2:11 3:52 4:11 5:26 6:52 7:38 8:98 9:93 10:11 11:11 12:33 12:52 12:11 12:11 13:11 14:52 14:52 15:12 16:11 16:11 
7:11 18:11 18:11 19:13 20:52 20:52 21:66 22:93 23:21 24:98 25:81 26:11 27:11 28:7 29:52 30:11 31:11 31:11 31:11 32:11 32:11 33:11 34:57 34:52 34:52 35:11 36:25 37:11 37:11 37:11 38:32 39:11 40:1 40:52 41:52 42:52 42:52 43:52 43:52 44:11 45:14 46:11 47:11 48:38 49:11 50:36 51:52 52:12 
53:65 54:83 55:10 56:27 57:4 58:52 59:52 60:24 60:35 61:63 62:83 14:52 14:52 63:89 63:83 64:5 65:43 66:52 20:52 20:52 67:81 68:41 69:65 70:41 71:41 72:83 
73:88 74:83 75:24 76:83 77:73 78:10 79:24 80:5 81:52 40:52 40:52 82:52 42:52 42:52 43:52 43:52 51:52 

• model-final.others
  一些参数

alpha=0.5
beta=0.1
ntopics=100
ndocs=100
nwords=4815
liters=999

---

根据结果文件,发现训练集和测试集的结果文件类型数量相同,并且类型也相同.所以没必要把estimate和inference拆开写.
另外inference的测试过程也是一种训练,只不过训练时要参考原有训练集训练的结果,这样会使训练更准确.
测试集new的model叫newModel,如果测试集的数据量比原有测试集还要大,这newModel可以代替原有的训练集.

---

问题:

测试集的训练过程是以自身为主,还是以原有训练集为主.要弄清楚这一点.

---

下一步计划:

1.下面的工作是继续分析测试集的内容,并且重新分析这个代码如何完成.
2.给源代码添加文件,让各种结果可视化.
3.中文lda.