|
|
|
@ -1,549 +0,0 @@
|
|
|
|
|
相机导购专家系统
|
|
|
|
|
====================================
|
|
|
|
|
|
|
|
|
|
如今单反相机的高度发达为摄影爱好者带来了无限的可能,让每个人都可以在花不是那么多钱的情况下创造艺术或是记录生活。我认为,这是科技给人带来的福音。
|
|
|
|
|
|
|
|
|
|
然而,正是因为单反市场太过发达,初入单反坑的玩家往往被搞得晕头转向。常见的单反厂商有尼康、索尼、佳能。高端市场有哈苏、徕卡,低端市场有理光、富士等。每个品牌又有很多个产品线,各有不同,设计给不同需求的用户使用。
|
|
|
|
|
|
|
|
|
|
如果对单反相机的参数没有深入的研究,仅仅是听商家的吹捧,那很难在最高的性价比上买到合适自己的那款相机。往往会花冤枉钱,或发现相机的主打功能自己根本就用不上。
|
|
|
|
|
|
|
|
|
|
本专家系统就为了解决这个问题而创建,担任一个细心公正的相机导购师,通过询问用户更在意哪方面的内容、平时使用情景、预算、对重量的承受能力等,为用户推荐最合适的相机,让用户少花冤枉钱,用最少的钱买到最有用的相机。
|
|
|
|
|
|
|
|
|
|
# 数据库的采集
|
|
|
|
|
|
|
|
|
|
对专家系统而言,不仅要有规则数据库,相机本身各项指标的数据库也非常重要。本专家系统采集权威平台:DxOMark的数据库,来获得详细的相机各项指标数据。
|
|
|
|
|
|
|
|
|
|
由于没有现成的数据可供下载,我自己使用NodeJS写了采集程序来做这件事情。代码如下:
|
|
|
|
|
|
|
|
|
|
```javascript
|
|
|
|
|
let request = require('request').defaults({ 'proxy': "http://127.0.0.1:1080" });
|
|
|
|
|
let Queue = require('promise-queue');
|
|
|
|
|
let fs = require('fs');
|
|
|
|
|
let path = require('path');
|
|
|
|
|
let queue = new Queue(10); //最多同时10线程采集
|
|
|
|
|
|
|
|
|
|
request('https://www.dxomark.com/daksensor/ajax/jsontested', //获得相机列表
|
|
|
|
|
(error, response, body) => {
|
|
|
|
|
let cameraList = JSON.parse(body).data;
|
|
|
|
|
let finishedCount = 0;
|
|
|
|
|
cameraList.forEach(cameraMeta => {
|
|
|
|
|
let camera = Object.assign({}, cameraMeta);
|
|
|
|
|
let link = `https://www.dxomark.com${camera.link}---Specifications`; //拼合对应的规格网址
|
|
|
|
|
queue.add(() => new Promise(res => { //把request放入队列,以保证同时http请求不超过10个
|
|
|
|
|
let doRequest = () => {
|
|
|
|
|
request(link, (error, response, body) => {
|
|
|
|
|
if (error) { //如果失败则重试
|
|
|
|
|
console.log("retrying.." + camera.name);
|
|
|
|
|
doRequest();
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
let specMatcherRegexp = /descriptifgauche.+?>([\s\S]+?)<\/td>[\s\S]+?descriptif_data.+?>([\s\S]+?)<\/td>/img;
|
|
|
|
|
let match = specMatcherRegexp.exec(body);
|
|
|
|
|
while (match) { //用正则表达式匹配table里的每一个项目,并添加至采集结果中
|
|
|
|
|
camera[match[1]] = match[2];
|
|
|
|
|
match = specMatcherRegexp.exec(body);
|
|
|
|
|
}
|
|
|
|
|
fs.writeFileSync(path.join('./scraped', camera.name + '.txt'), JSON.stringify(camera, null, 4), { encoding: "UTF8" });
|
|
|
|
|
finishedCount++;
|
|
|
|
|
console.log(`Finished ${finishedCount}/${cameraList.length}: ${camera.name}`);
|
|
|
|
|
res();
|
|
|
|
|
})
|
|
|
|
|
};
|
|
|
|
|
doRequest();
|
|
|
|
|
}));
|
|
|
|
|
});
|
|
|
|
|
});
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
采集结果如图所示,总共采集到357款单反产品:
|
|
|
|
|
![scrape](imgs/scrape.gif)
|
|
|
|
|
|
|
|
|
|
# 数据清洗
|
|
|
|
|
|
|
|
|
|
数据采集完了,但有些数据比较脏(如含有 ),有些数据是我们不需要的,有些数据用起来不方便(如分辨率是 1234 x 1234形式的字符串),属性名中有空格和大写字符,也不美观。因此,额外添加一步数据清洗。
|
|
|
|
|
|
|
|
|
|
数据清洗之后,希望留下以下属性:
|
|
|
|
|
|
|
|
|
|
> 相机名称、相机图片、价格、像素数量、每秒连拍数量、重量、对焦系统质量、最大感光度、模型出厂日期、触屏存在、可录视频、有闪光灯、有蓝牙、有GPS、防水、机身材质质量
|
|
|
|
|
|
|
|
|
|
数据清洗的代码实现为:
|
|
|
|
|
```javascript
|
|
|
|
|
let fs = require("fs");
|
|
|
|
|
let path = require("path");
|
|
|
|
|
|
|
|
|
|
function parseNumberFunctionFactory(keyMatcher, valueMatcher = /(\d+\.?\d*)/im,
|
|
|
|
|
returnValueDecider = match => +match[1]) {
|
|
|
|
|
return (data) => {
|
|
|
|
|
let key = Object.keys(data).filter(_ => keyMatcher.test(_.trim()))[0];
|
|
|
|
|
if (!key) return null;
|
|
|
|
|
let match = data[key].toString().match(valueMatcher);
|
|
|
|
|
if (match)
|
|
|
|
|
return returnValueDecider(match);
|
|
|
|
|
else
|
|
|
|
|
return null;
|
|
|
|
|
};
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
let parseResolution = parseNumberFunctionFactory(/^resolution$/im, /(\d+)\s*x\s*(\d+)/im, match => [+match[1], +match[2]]);
|
|
|
|
|
let parseFrameRate = parseNumberFunctionFactory(/frame rate/im);
|
|
|
|
|
let parseWeight = parseNumberFunctionFactory(/weight/im);
|
|
|
|
|
let parseAutoFocus = parseNumberFunctionFactory(/number of autofocus points/im);
|
|
|
|
|
let parseISO = parseNumberFunctionFactory(/ISO latitude/im, /(\d+)\s*-\s*(\d+)/im, match => [+match[1], +match[2]]);
|
|
|
|
|
let parseLaunchDate = parseNumberFunctionFactory(/launchDateGraph/im, /(\d+)-(\d+)-(\d+)/im, match => new Date(match[1], match[2] - 1, match[3]));
|
|
|
|
|
let parseTouchScreen = parseNumberFunctionFactory(/Touch screen/im, /yes/im, match => !!match);
|
|
|
|
|
let parseVideo = parseNumberFunctionFactory(/^Video$/m, /yes/im, match => !!match);
|
|
|
|
|
let parseFlash = parseNumberFunctionFactory(/^flash$/im, /yes/im, match => !!match);
|
|
|
|
|
let parseWaterproof = parseNumberFunctionFactory(/^waterproof$/im, /yes/im, match => !!match);
|
|
|
|
|
let parseBluetooth = parseNumberFunctionFactory(/^Bluetooth$/im, /yes/im, match => !!match);
|
|
|
|
|
let parseGps = parseNumberFunctionFactory(/^GPS$/im, /yes/im, match => !!match);
|
|
|
|
|
let parseIsMetal = parseNumberFunctionFactory(/^camera material$/im, /metal/im, match => !!match);
|
|
|
|
|
|
|
|
|
|
let files = fs.readdirSync("./scraped");
|
|
|
|
|
files.forEach(file => {
|
|
|
|
|
let data = JSON.parse(fs.readFileSync(path.join('./scraped', file), { encoding: "utf8" }));
|
|
|
|
|
let resolution = parseResolution(data);
|
|
|
|
|
//机身材质质量
|
|
|
|
|
let frameRate = parseFrameRate(data);
|
|
|
|
|
let weight = parseWeight(data);
|
|
|
|
|
let autoFocus = parseAutoFocus(data);
|
|
|
|
|
let iso = parseISO(data);
|
|
|
|
|
let launchDate = parseLaunchDate(data);
|
|
|
|
|
let touchScreen = parseTouchScreen(data);
|
|
|
|
|
let video = parseVideo(data);
|
|
|
|
|
let flash = parseFlash(data);
|
|
|
|
|
let waterproof = parseWaterproof(data);
|
|
|
|
|
let bluetooth = parseBluetooth(data);
|
|
|
|
|
let gps = parseGps(data);
|
|
|
|
|
let isMetal = parseIsMetal(data);
|
|
|
|
|
let cleanedData = {
|
|
|
|
|
name: data.name,
|
|
|
|
|
image: data.image,
|
|
|
|
|
brand: data.brand,
|
|
|
|
|
price: data.price,
|
|
|
|
|
pixelDepth: data.pixelDepth,
|
|
|
|
|
pixels: resolution ? (resolution[0] * resolution[1]) : 0,
|
|
|
|
|
ISO: iso,
|
|
|
|
|
maxISO: iso ? iso[1] : 0,
|
|
|
|
|
launchDate: +launchDate,
|
|
|
|
|
touchScreen,
|
|
|
|
|
video,
|
|
|
|
|
flash,
|
|
|
|
|
waterproof,
|
|
|
|
|
bluetooth,
|
|
|
|
|
gps,
|
|
|
|
|
isMetal,
|
|
|
|
|
frameRate,
|
|
|
|
|
resolution,
|
|
|
|
|
weight,
|
|
|
|
|
autoFocus,
|
|
|
|
|
};
|
|
|
|
|
fs.writeFileSync(path.join('./cleaned', file), JSON.stringify(cleanedData, null, 4), { encoding: "utf8" });
|
|
|
|
|
});
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
清洗结果演示:
|
|
|
|
|
```json
|
|
|
|
|
{
|
|
|
|
|
"name": "Nikon D5",
|
|
|
|
|
"image": "//cdn.dxomark.com/dakdata/xml/D5/vignette3.png",
|
|
|
|
|
"brand": "Nikon",
|
|
|
|
|
"price": 6500,
|
|
|
|
|
"pixelDepth": 20.8,
|
|
|
|
|
"pixels": 20817152,
|
|
|
|
|
"ISO": [
|
|
|
|
|
50,
|
|
|
|
|
3280000
|
|
|
|
|
],
|
|
|
|
|
"maxISO": 3280000,
|
|
|
|
|
"launchDate": 1452009600000,
|
|
|
|
|
"touchScreen": true,
|
|
|
|
|
"video": true,
|
|
|
|
|
"flash": null,
|
|
|
|
|
"waterproof": null,
|
|
|
|
|
"bluetooth": null,
|
|
|
|
|
"gps": true,
|
|
|
|
|
"isMetal": true,
|
|
|
|
|
"frameRate": 14,
|
|
|
|
|
"resolution": [
|
|
|
|
|
5584,
|
|
|
|
|
3728
|
|
|
|
|
],
|
|
|
|
|
"weight": 1225,
|
|
|
|
|
"autoFocus": 153
|
|
|
|
|
}
|
|
|
|
|
```
|
|
|
|
|
看起来好多了。
|
|
|
|
|
|
|
|
|
|
# 问题的设计
|
|
|
|
|
斟酌一番后,我决定将询问用户的问题定为:
|
|
|
|
|
|
|
|
|
|
1. 您将如何使用本相机(多选)
|
|
|
|
|
* 记录旅行
|
|
|
|
|
* 拍摄学校或公司活动
|
|
|
|
|
* 拍摄体育比赛
|
|
|
|
|
* 拍摄自然风景
|
|
|
|
|
* 拍摄人像
|
|
|
|
|
* 拍摄天文
|
|
|
|
|
2. 您看中哪些额外功能吗(多选)
|
|
|
|
|
* 内置闪光灯
|
|
|
|
|
* 可录制视频
|
|
|
|
|
* 可蓝牙传输照片
|
|
|
|
|
* 可触屏
|
|
|
|
|
* 内置GPS
|
|
|
|
|
* 防水
|
|
|
|
|
3. 您是否愿意承受单反的重量
|
|
|
|
|
* 没问题,3公斤的机器都扛得住
|
|
|
|
|
* 在能避免负重的情况下尽可能避免负重
|
|
|
|
|
* 不愿意接受重的单反,必须较为轻便
|
|
|
|
|
4. 您愿意在单反上投入的经济
|
|
|
|
|
* 很多,一步到位买高端设备
|
|
|
|
|
* 普通,好用实用的设备
|
|
|
|
|
* 经济,请推荐入门基本款
|
|
|
|
|
5. 您有什么别的要求吗(多选)
|
|
|
|
|
* 尽量购买新的型号
|
|
|
|
|
* 机身材质要好
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 模糊专家系统的设计
|
|
|
|
|
|
|
|
|
|
一个显著的问题是:模糊专家系统一般只能用于数值的计算(如评估房价),但我这里做的,却是根据用户的输入推荐产品。用模糊专家系统如何做产品推荐呢?
|
|
|
|
|
|
|
|
|
|
绕一个弯,不难解决这个问题:使用模糊专家系统对每一款相机产品计算出一个“推荐度”,根据推荐度排序,给用户推荐排名前三的产品。
|
|
|
|
|
|
|
|
|
|
具体的,我使用专家系统,计算出以下几个指标:
|
|
|
|
|
|
|
|
|
|
> 适合拍摄旅行,适合拍摄活动,适合拍摄体育,适合拍摄风景,适合拍摄人像,适合拍摄天文,型号新,机身材质好,重量轻,价格低
|
|
|
|
|
|
|
|
|
|
用户的要求可以和这几个指标对号入座,来计算要求满足度。
|
|
|
|
|
|
|
|
|
|
对于另外一些如“有内置闪光灯”等指标,用不着用模糊专家系统,直接布尔判断来算满足度。
|
|
|
|
|
|
|
|
|
|
满足度根据规则加权平均,就是产品的总推荐度,排序后,前几名就是推荐给用户的相机。
|
|
|
|
|
|
|
|
|
|
这样做的好处,除了推荐相机之外,还可以列出相机的Pros & Cons (如:+ 适合拍摄天文 + 可录制视频 - 重量较重),只要去取满足度最高/最低的几名就可以了。
|
|
|
|
|
|
|
|
|
|
为实现模糊专家系统,我使用jFuzzyLogic框架,它是Java下最完整的模糊逻辑框架,支持fuzzy language (FCL),IEC 61131-7标准。
|
|
|
|
|
|
|
|
|
|
# 规则的建立
|
|
|
|
|
|
|
|
|
|
模糊专家系统在相机推荐中,最重要的工作就是:建立相机参数和相机适合拍摄的照片种类之间的联系,比如:
|
|
|
|
|
|
|
|
|
|
> 适合拍摄体育类型相片的相机,需要有**很高**的连拍速度、**较好**的对焦系统。
|
|
|
|
|
|
|
|
|
|
在这个实验性的导购系统中,由我自己担任“领域专家”的角色。
|
|
|
|
|
|
|
|
|
|
首先,要定义术语,如:
|
|
|
|
|
|
|
|
|
|
* 价格(美金)便宜: \$0-1000 中等: \$600-1700 贵: \$1500-
|
|
|
|
|
|
|
|
|
|
翻译成FCL之后:
|
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
FUZZIFY price
|
|
|
|
|
TERM low := (0, 1) (1000, 0) ;
|
|
|
|
|
TERM medium := (600, 0) (800,1) (1500,1) (1700,0);
|
|
|
|
|
TERM high := (1500, 0) (1700, 1);
|
|
|
|
|
END_FUZZIFY
|
|
|
|
|
|
|
|
|
|
FUZZIFY weight
|
|
|
|
|
TERM light := (0, 1) (500, 0) ;
|
|
|
|
|
TERM medium := (300, 0) (500,1) (700,1) (1000,0);
|
|
|
|
|
TERM heavy := (800, 0) (1000, 1);
|
|
|
|
|
END_FUZZIFY
|
|
|
|
|
|
|
|
|
|
...
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
每个术语的定义如图所示:
|
|
|
|
|
|
|
|
|
|
![fuzzify](imgs/fuzzify.gif)
|
|
|
|
|
|
|
|
|
|
然后,要定义defuzzify规则,这里方便起见,均只定义“veryGood”、“good”、"average"、"bad"、"veryBad"。
|
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
DEFUZZIFY travel
|
|
|
|
|
TERM veryBad := (0,1) (0.2,0);
|
|
|
|
|
TERM bad := (0,0) (0.1,1) (0.5,0);
|
|
|
|
|
TERM average:= (0,0) (0.5,1) (1,0);
|
|
|
|
|
TERM good:= (0.5,0) (0.9,1) (1,0);
|
|
|
|
|
TERM veryGood:= (0.8,0) (1,1);
|
|
|
|
|
METHOD : COG;
|
|
|
|
|
DEFAULT := 0.5;
|
|
|
|
|
END_DEFUZZIFY
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
最后,要定义规则,举例如下:
|
|
|
|
|
|
|
|
|
|
* 如果 感光度范围 很高 而且 像素 高 那么 适合拍天文 高
|
|
|
|
|
|
|
|
|
|
* 如果 连拍速度 很高 而且 对焦系统 好 那么 适合拍体育 高
|
|
|
|
|
|
|
|
|
|
* 如果 重量 轻 且 有GPS 那么 适合记录旅行 高
|
|
|
|
|
|
|
|
|
|
* ……
|
|
|
|
|
|
|
|
|
|
翻译成FCL之后:
|
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
RULEBLOCK travel
|
|
|
|
|
AND : MIN;
|
|
|
|
|
RULE 1 : IF weight IS light THEN travel IS good;
|
|
|
|
|
RULE 2 : IF video IS yes THEN travel IS good;
|
|
|
|
|
RULE 3 : IF gps IS yes THEN travel IS good;
|
|
|
|
|
RULE 4 : IF flash IS no THEN travel IS bad;
|
|
|
|
|
RULE 5 : IF weight IS heavy THEN travel IS veryBad;
|
|
|
|
|
END_RULEBLOCK
|
|
|
|
|
|
|
|
|
|
RULEBLOCK sports
|
|
|
|
|
AND : MIN;
|
|
|
|
|
RULE 1 : IF frameRate IS high THEN sports IS veryGood;
|
|
|
|
|
RULE 2 : IF autoFocus IS high THEN sports IS veryGood;
|
|
|
|
|
RULE 3 : IF pixels IS high THEN sports IS good;
|
|
|
|
|
RULE 4 : IF frameRate IS low THEN sports IS veryBad;
|
|
|
|
|
RULE 5 : IF autoFocus IS low THEN sports IS veryBad;
|
|
|
|
|
END_RULEBLOCK
|
|
|
|
|
|
|
|
|
|
RULEBLOCK astronomy
|
|
|
|
|
AND : MIN;
|
|
|
|
|
RULE 1 : IF pixels IS high THEN astronomy IS good;
|
|
|
|
|
RULE 2 : IF pixelDepth IS high THEN astronomy IS good;
|
|
|
|
|
RULE 3 : IF maxISO IS high THEN astronomy IS good;
|
|
|
|
|
RULE 4 : IF maxISO IS low THEN astronomy IS veryBad;
|
|
|
|
|
RULE 5 : IF pixels IS low THEN astronomy IS veryBad;
|
|
|
|
|
END_RULEBLOCK
|
|
|
|
|
|
|
|
|
|
...
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
# 打分的具体实现
|
|
|
|
|
数据和规则都准备就绪后,就可以开始进行模糊推理。
|
|
|
|
|
|
|
|
|
|
Java程序框架如下:
|
|
|
|
|
|
|
|
|
|
```java
|
|
|
|
|
package ai.fuzzy;
|
|
|
|
|
|
|
|
|
|
import com.alibaba.fastjson.JSON;
|
|
|
|
|
import net.sourceforge.jFuzzyLogic.FIS;
|
|
|
|
|
import net.sourceforge.jFuzzyLogic.FunctionBlock;
|
|
|
|
|
import net.sourceforge.jFuzzyLogic.plot.JFuzzyChart;
|
|
|
|
|
import net.sourceforge.jFuzzyLogic.rule.Variable;
|
|
|
|
|
|
|
|
|
|
import java.io.*;
|
|
|
|
|
import java.nio.file.Paths;
|
|
|
|
|
|
|
|
|
|
class CameraData {
|
|
|
|
|
public CameraAssessment assessment;
|
|
|
|
|
public String name;
|
|
|
|
|
public String image;
|
|
|
|
|
public String brand;
|
|
|
|
|
public Integer price;
|
|
|
|
|
public Integer pixelDepth;
|
|
|
|
|
public Integer pixels;
|
|
|
|
|
public Integer maxISO;
|
|
|
|
|
public Integer weight;
|
|
|
|
|
public Integer autoFocus;
|
|
|
|
|
public Long launchDate;
|
|
|
|
|
public Float frameRate;
|
|
|
|
|
public Integer[] resolution;
|
|
|
|
|
public Integer[] ISO;
|
|
|
|
|
public boolean touchScreen;
|
|
|
|
|
public boolean video;
|
|
|
|
|
public boolean flash;
|
|
|
|
|
public boolean waterproof;
|
|
|
|
|
public boolean bluetooth;
|
|
|
|
|
public boolean gps;
|
|
|
|
|
public boolean isMetal;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
class CameraAssessment {
|
|
|
|
|
public double travel;
|
|
|
|
|
public double event;
|
|
|
|
|
public double sports;
|
|
|
|
|
public double scenery;
|
|
|
|
|
public double portrait;
|
|
|
|
|
public double astronomy;
|
|
|
|
|
public double newModel;
|
|
|
|
|
public double durableBuild;
|
|
|
|
|
public double lightBuild;
|
|
|
|
|
public double lowPrice;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
public class Main {
|
|
|
|
|
public static void main(String[] args) throws IOException {
|
|
|
|
|
File rootFolder = new File("input");
|
|
|
|
|
for (final File fileEntry : rootFolder.listFiles()) {
|
|
|
|
|
if (fileEntry.isFile()) {
|
|
|
|
|
CameraData camera = JSON.parseObject(readFileContents(fileEntry), CameraData.class);
|
|
|
|
|
camera.assessment = assess(camera);
|
|
|
|
|
writeFile(fileEntry, JSON.toJSONString(camera, true));
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private static void writeFile(File fileEntry, String jsonString) throws FileNotFoundException {
|
|
|
|
|
...
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private static String readFileContents(File fileEntry) throws IOException {
|
|
|
|
|
...
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static CameraAssessment assess(CameraData cameraData) {
|
|
|
|
|
...
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
```assess```函数如下:
|
|
|
|
|
|
|
|
|
|
```java
|
|
|
|
|
|
|
|
|
|
static CameraAssessment assess(CameraData cameraData) {
|
|
|
|
|
CameraAssessment cameraAssessment = new CameraAssessment();
|
|
|
|
|
String fileName = "fcl/camera.fcl";
|
|
|
|
|
FIS fis = FIS.load(fileName, true);
|
|
|
|
|
|
|
|
|
|
// Set inputs
|
|
|
|
|
fis.setVariable("price", cameraData.price);
|
|
|
|
|
fis.setVariable("pixelDepth", cameraData.pixelDepth);
|
|
|
|
|
fis.setVariable("pixels", cameraData.pixels);
|
|
|
|
|
fis.setVariable("maxISO", cameraData.maxISO);
|
|
|
|
|
fis.setVariable("weight", cameraData.weight);
|
|
|
|
|
fis.setVariable("autoFocus", cameraData.autoFocus);
|
|
|
|
|
fis.setVariable("launchDate", cameraData.launchDate);
|
|
|
|
|
fis.setVariable("frameRate", cameraData.frameRate);
|
|
|
|
|
fis.setVariable("touchScreen", cameraData.touchScreen ? 1 : 0);
|
|
|
|
|
fis.setVariable("video", cameraData.video ? 1 : 0);
|
|
|
|
|
fis.setVariable("flash", cameraData.flash ? 1 : 0);
|
|
|
|
|
fis.setVariable("waterproof", cameraData.waterproof ? 1 : 0);
|
|
|
|
|
fis.setVariable("bluetooth", cameraData.bluetooth ? 1 : 0);
|
|
|
|
|
fis.setVariable("gps", cameraData.gps ? 1 : 0);
|
|
|
|
|
fis.setVariable("isMetal", cameraData.isMetal ? 1 : 0);
|
|
|
|
|
|
|
|
|
|
// Evaluate
|
|
|
|
|
fis.evaluate();
|
|
|
|
|
|
|
|
|
|
// Save results to cameraAssessment
|
|
|
|
|
cameraAssessment.travel = fis.getVariable("travel").defuzzify();
|
|
|
|
|
cameraAssessment.event = fis.getVariable("event").defuzzify();
|
|
|
|
|
cameraAssessment.sports = fis.getVariable("sports").defuzzify();
|
|
|
|
|
cameraAssessment.scenery = fis.getVariable("scenery").defuzzify();
|
|
|
|
|
cameraAssessment.portrait = fis.getVariable("portrait").defuzzify();
|
|
|
|
|
cameraAssessment.astronomy = fis.getVariable("astronomy").defuzzify();
|
|
|
|
|
cameraAssessment.newModel = fis.getVariable("newModel").defuzzify();
|
|
|
|
|
cameraAssessment.durableBuild = fis.getVariable("durableBuild").defuzzify();
|
|
|
|
|
cameraAssessment.lightBuild = fis.getVariable("lightBuild").defuzzify();
|
|
|
|
|
cameraAssessment.lowPrice = fis.getVariable("lowPrice").defuzzify();
|
|
|
|
|
|
|
|
|
|
return cameraAssessment;
|
|
|
|
|
}
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
调用后,模糊专家系统就会给每个相机做评测,并作出如下输出:
|
|
|
|
|
|
|
|
|
|
```json
|
|
|
|
|
"assessment": {
|
|
|
|
|
"astronomy": 0.07206054514320881,
|
|
|
|
|
"durableBuild": 0.1999999999999999,
|
|
|
|
|
"event": 0.1999999999999999,
|
|
|
|
|
"lightBuild": 0.1999999999999999,
|
|
|
|
|
"lowPrice": 0.1999999999999999,
|
|
|
|
|
"newModel": 0.20107756449438624,
|
|
|
|
|
"portrait": 0.2280106572609703,
|
|
|
|
|
"scenery": 0.2280106572609703,
|
|
|
|
|
"sports": 0.07875190169689873,
|
|
|
|
|
"travel": 0.16167332382310975,
|
|
|
|
|
},
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
分别是相机对于每一项情景的适合度,在0-1之间。
|
|
|
|
|
|
|
|
|
|
接下来,需要把用户输入和适合度整合,得出总评分。
|
|
|
|
|
|
|
|
|
|
这里,总评分直接通过用户输入和适合度做内积的方法获得,评分的同时,也记录分数变动记录,这样给用户推荐就可以说明推荐原因。
|
|
|
|
|
|
|
|
|
|
```javascript
|
|
|
|
|
function evaluate(item, tags) {
|
|
|
|
|
let score = 0, changes = [];
|
|
|
|
|
tags.forEach(tag => {
|
|
|
|
|
let reverse = false;
|
|
|
|
|
let normalizedTag = tag;
|
|
|
|
|
if (tag.startsWith("!")) {
|
|
|
|
|
//允许使用!开头,表示相反。如:!lowPrice时,lowPrice原本加分现在变成减分
|
|
|
|
|
reverse = true;
|
|
|
|
|
normalizedTag = tag.substr(1);
|
|
|
|
|
}
|
|
|
|
|
let scoreChange = 0;
|
|
|
|
|
if (item.assessment[normalizedTag]) { // 如果是专家系统assess出来的结果,一个占20分
|
|
|
|
|
scoreChange = (reverse ? -1 : 1) * (item.assessment[normalizedTag] * 20) * (weight[normalizedTag] || 1);
|
|
|
|
|
} else { // 如果不是,那么是“防水”等基本要求,一个占3分
|
|
|
|
|
scoreChange = (reverse ? -1 : 1) * (item[normalizedTag] ? 1 : -1) * 3 * (weight[normalizedTag] || 1);
|
|
|
|
|
}
|
|
|
|
|
if (scoreChange) {
|
|
|
|
|
score += scoreChange;
|
|
|
|
|
changes.push([scoreChange, tag]); // 记录评分变化,之后好出pros & cons
|
|
|
|
|
}
|
|
|
|
|
});
|
|
|
|
|
return { score, changes };
|
|
|
|
|
}
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
系统核心算法大致就完成了。
|
|
|
|
|
|
|
|
|
|
# 参数调整
|
|
|
|
|
参数调整方面,花了不少心思。
|
|
|
|
|
|
|
|
|
|
原来的专家系统规则中,用了比较多的OR,导致多款相机某个指标打分非常相近,没有区分度。因此,我后来重写了FCL,尽量避免使用OR。为了拉开差距,我还把原来的“good”、"average"、"bad"改成了“veryGood”、“good”、"average"、"bad"、"veryBad",以说明某些规格的重要性大于其他规格。
|
|
|
|
|
|
|
|
|
|
由于RULES的good、bad不平衡,会导致一些assessment普遍偏低,另一些普遍偏高,因此,在后续处理中,我把这些assessment进行了normalization,让他们的平均值归一到0。
|
|
|
|
|
|
|
|
|
|
又出现了新的问题:推荐的相机,基本都是“价格低”、“型号新”,即这两个特性给相机总体加分太多,掩盖了别的优点,因此,我人工指定了指标的权重,以削弱价格和型号新旧对总体评分的影响,彰显“适合拍的类型”在推荐中的占比。
|
|
|
|
|
|
|
|
|
|
但还有很多问题,其中最主要的是:数据区分度不明显,高端相机的数据都差不多;指标内部有较强的关联性(合适拍天文的,一般也合适拍风景)。数据内部关联性不好解决。
|
|
|
|
|
|
|
|
|
|
经过一番调整,系统可以产生尚可的结果。
|
|
|
|
|
|
|
|
|
|
# 前端与运行效果
|
|
|
|
|
|
|
|
|
|
为了获得更好的效果,我为专家系统做了一个用户友好的前端界面,可以引导用户输入信息、以用户友好的方式给出推荐结果。前端是使用React + antd做的,基于Web。代码省略,界面长这样:
|
|
|
|
|
|
|
|
|
|
![ui](imgs/ui.png)
|
|
|
|
|
|
|
|
|
|
以上用户选择会被转换为tags:
|
|
|
|
|
|
|
|
|
|
```json
|
|
|
|
|
["travel", "scenery", "video", "gps", "!lowPrice", "durableBuild"]
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
然后,这些tags会被送到之前设计好的系统中进行运算(内积rank),获得结果。
|
|
|
|
|
|
|
|
|
|
![ui2](imgs/ui2.png)
|
|
|
|
|
|
|
|
|
|
可以看出,本系统可以根据用户的需要推荐相机(Pros和Cons都与用户第一步选择的“需求”有关)。
|
|
|
|
|
|
|
|
|
|
并能以一种易于理解的方式说清楚每一项“需求”是如何影响最终打分的。
|
|
|
|
|
|
|
|
|
|
# 结论
|
|
|
|
|
|
|
|
|
|
本导购系统只是一个试水,效果还过得去。
|
|
|
|
|
|
|
|
|
|
在进行导购系统制作的过程中,我充分体验了制作一个专家系统需要的每一个步骤。从采集数据,到模糊集定义,到模糊集推理,到整合结果并输出给用户看。
|
|
|
|
|
|
|
|
|
|
这个专家系统与普通的估测房价/工资/小费的模糊专家系统不同,必须对模糊专家系统defuzzify之后的结果进行进一步的整合,并给用户做出推荐。这个整合算法,应该也能算是广义专家系统的一部分吧。
|
|
|
|
|
|
|
|
|
|
效果比预期的稍微差一些,我认为原因主要在:
|
|
|
|
|
|
|
|
|
|
1. 原始数据不够完整(DxOMark上仍然很多型号缺很多数据)
|
|
|
|
|
2. 数据区分度不高、内部关联性太强(高端相机的指标都差不多;合适拍天文的,一般也合适拍风景)
|
|
|
|
|
|
|
|
|
|
进一步完善数据库、调整参数、或换一种性价比的模型之后,有望提升整体效果。
|
|
|
|
|
|
|
|
|
|
但无论如何,本专家系统至少做到了往正确的方向响应用户的输入,有理有据地为用户推荐符合要求的相机。
|
|
|
|
|
|
|
|
|
|
这种专家系统是有价值的,在进一步的优化后,说不定能产生一些商业效益。
|
|
|
|
|
|
|
|
|
|
现在我仅仅对机身做了推荐,如果加上镜头的组合,那么规则更加复杂、要考虑的因素更加多。
|
|
|
|
|
|
|
|
|
|
但按照规则解决这些复杂的问题,或许就是专家系统真正意义所在。真的解决之后,专家系统就可以和一个专业的人类导购师一样,耐心服务每一位顾客,让每个人都把钱花在刀刃上。
|