{{title}}
3 |{{summary}}
4 |
152 |
153 |
260 | Problem motivation
154 |Do I need Impulse-ML: Recommender, the Recommender System?
155 |If you are a PHP developer who maintains any PHP social application and you 156 | want to predict the "rating" or "preference" that a user would give to an item 157 | the Impulse-ML: Recommender is library that you might consider to use! 158 | In further readings I will show you how to use Impulse-ML: Recommender 159 | and give you a hint on how to choose parameters which makes the predictions 160 | more accurate.
161 |Problem definition
162 |Consider the following data:
163 |+---------------------------------+------+---------+---------+---------+
164 | | Movie \ User | Anna | Barbara | Charlie | Dave |
165 | +---------------------------------+------+---------+---------+---------+
166 | | The Dark Knight | 0 | 0 | 5 | 5 |
167 | +---------------------------------+------+---------+---------+---------+
168 | | Guardians of the Galaxy | 0 | ? | ? | 5 |
169 | +---------------------------------+------+---------+---------+---------+
170 | | Logan | ? | 0 | 4 | ? |
171 | +---------------------------------+------+---------+---------+---------+
172 | | Forrest Gump | 4 | 5 | 0 | 0 |
173 | +---------------------------------+------+---------+---------+---------+
174 | | The Kid | 5 | 5 | 0 | 0 |
175 | +---------------------------------+------+---------+---------+---------+
176 | In this particular example we can notice:
179 |-
180 |
- we have 5 items - 5 movies 181 |
- we have 4 categories - 4 users 182 |
- we can notice 2 types of items: action movie and comedy movie 183 |
- it seems that Anna and Barbara hate the action movies but love the comedy movies 184 |
- it seems that Charlie and Dave love the action movies but hate the comedy movies 185 |
- the table is incomplete because every user has not rated at least one movie 186 |
Using this data you might want to:
188 |-
189 |
- predict user rating of movie that is unrated by user i.e. to send user the movie which he would like but he does not rated that movie yet 190 |
- get movies similar to given movie 191 |
- get the prediction of the movie for user that has no rated any movie and use this data 192 |
Using Impulse-ML: Recommender you might end up with such predictions:
194 |+---------------------------------+------+---------+---------+---------+
195 | | Movie \ User | Anna | Barbara | Charlie | Dave |
196 | +---------------------------------+------+---------+---------+---------+
197 | | The Dark Knight | - | - | - | - |
198 | +---------------------------------+------+---------+---------+---------+
199 | | Guardians of the Galaxy | - | 0 | 5 | - |
200 | +---------------------------------+------+---------+---------+---------+
201 | | Logan | 0 | - | - | 4 |
202 | +---------------------------------+------+---------+---------+---------+
203 | | Forrest Gump | - | - | - | - |
204 | +---------------------------------+------+---------+---------+---------+
205 | | The Kid | - | - | - | - |
206 | +---------------------------------+------+---------+---------+---------+
207 | We might notice:
210 |-
211 |
- Anna hates action movies so the prediction of "Logan" will be 0 212 |
- Barbara also hates action movies so the prediction of "Guardians of the Galaxy" will be 0 213 |
- Charlie loves the action movies so the prediction of "Guardians of the Galaxy" will be 5 214 |
- Dave also loves the action movies so prediction of "Logan" will be 4 (not 5 since the maximum rating of this movie is equal 4) 215 |
That's how Collaborative Filtering works.
217 |Training and training parameters
218 |As each machine learning problem after filling with data in order to get correct prediction the training 219 | (based on the dataset) is required.
220 |There is only one parameter for a Learning Model created from a dataset:
221 |-
222 |
- number of features. 223 |
Understand 225 | it like 226 | type or real category of the item. It's value can be set equals number of item types in your 227 | application. You don't need to name them, you have to know number of them.
228 |There are two training parameters:
229 |-
230 |
- learning rate 231 |
- number of iterations 232 |
The learning rate is parameter which describes how much gradient descent 234 | (which minimizes the error) will perform. You might to consider to increase or decrease 235 | this parameter and it has strong 236 | correlation with number of iterations.
237 |The number of iterations is parameter which describes how much steps gradient descent minimize function 238 | will be applied. It's highly correlated with learning rate.
239 |The results of prediction may vary from desired by setting this parameters less accurate.
240 |However, there are some rules of setting these parameters more accurate in order to get 241 | better prediction:
242 |-
243 |
- if you set small learning rate then you might consider increase number of iterations 244 |
- if you set large learning rate then you might consider decrease number of iterations 245 |
- you might expect very low error - in this example a reasonable error would be less than 0.0001 246 |
- setting too high learning rate may cause algorithm get computation error and the predictions become 247 | useless 248 |
For this particular example i have set:
250 |-
251 |
- learning rate === 0.01 252 |
- number of iterations === 20000 253 |
- number of features === 2 (since i noticed two types of movies or two user preferences) 254 |
The key to get well trained model is to choose the right ratio of learning rate and number of iterations.
256 |You might consider try different number of features according to your Application so the dataset also.
257 |Above example was fully implemented in examples/1_train_and_predict.php.
258 | 259 |