At that time, it was considered that the logic of the initial analysis was confused and the model was single, so it was improved from these two angles to continue the analysis.

The idea originated from the following background: in the current rental market, the renting party is often in a weak position, which requires not only intermediary fees but also additional risks, one of which comes from some black intermediaries who do not fulfill their obligations to protect information in order to save costs and even provide false information for their own profits. therefore, it is particularly dazzling to see the subjective recommendation marked "must look good house" in a series of objective data screened by Lianjia. In fact, renters do not have the ability to summarize and screen a lot of objective data, so they rely heavily on the recommendation of "must look good at housing". The combination of these two points leads to the ultimate goal of the question, whether the label "must take good care of the house" is true and reliable, that is, whether what was said in the previous report is really good?

In fact, there are obviously two points that are very confusing. first, the scarcity that has not been discussed before. when analyzing "must-see housing" with a single index, it is found that it is not so-called scarcity, or even on the contrary, such as in the price index, especially high or low prices of several pieces of housing information, is not marked as "must-see housing". In the location index, Chongming, which has a small number of housing, also appears the phenomenon that none of the housing is marked as "must see good housing". Another point is that of the 20 messages presented on the first page of rental information through comprehensive ranking, only 5 are marked as "must-see", accounting for 25%, which is basically the same as "must-see" accounting for 24.87% of all housing sources, indicating that "must-see" does not have an advantage in comprehensive ranking, putting a question mark again for its higher performance-to-price ratio.

The difficulty of the above methods is that we are not people in the industry, and there is no evaluation system for the ratio of performance to price. In fact, if Lianjia's colleagues have a similar suitable method for constructing the index of performance-to-price ratio, then we can directly use the above methods to judge whether Lianjia's "must-see good house" has a higher performance-to-price ratio.

*Since there is no suitable evaluation index of performance-to-price ratio, it is considered that starting from the definition of performance-to-price ratio, when all kinds of factors of the house are the same, compare the price level of "must-see good house" and non-"must-see good house". But the problem is that for a "must-see house", it is impossible to find a non-"must-see house" with exactly the same factors as it except the price, and there is no way to compare the price. Therefore, I hope to use the non-"must-see house" data to train a price prediction model and take the price as the target label. Substituting the other characteristic values of the "must-see house" is equivalent to generating a non-must-see house with all kinds of factors other than the price, and we also predict that in addition to its price, compare this price with the price of the "must-see house" to compare the performance-to-price ratio, so as to verify the conclusion.

The above is the previous practice, in fact, it is not necessary to choose the method of price prediction. I take the area as the target of the forecast to build a prediction model to compare the area of "must-see" and non-"must-see" houses with the same other factors except area. Of course, it is also possible for me to change the prediction target to a classification value, for example, orientation. It is generally believed that it is more cost-effective to face south when other conditions are the same, so a classification model is needed at this time, but I think this is not good. The reason is that the explanation of the classification model is not strong here, such as the comparison between the north and the south is less convincing, and there is no concept of how much the performance-to-price ratio is as bad as in the previous regression.*

Back to the first step of the train of thought, try to build a cost-effective index, at this time similar to the two groups of samples of hypothesis testing may be used to verify the conclusion.

The unexpected reason for the analysis effect may be that only the factors considered to have a great influence on the performance-to-price index were adopted in the construction of the performance-to-price index, which magnified the influence.*

After verifying the conclusion through different ways, the initial goal has been basically achieved, but we still hope to go back to the previous price forecasting model to explore whether the forecasting model can continue to improve the accuracy.

The problem encountered in feature engineering is similar to that in previous hypothesis testing, that is, there is no experience in the construction of cost-effective indicators, so we still start with reducing the possible multicollinearity between irrelevant indicators and control.

Because we need to explore the meaningful deviation value in order to verify the conclusion, and there is no regularization, the verification of the conclusion has been completed at this time. From the point of view of the evaluation model, we retrain after regularization. The goal is to explore the important features of the rental price forecasting model.

In the regression model trained after regularization, the importance of feature is the absolute value of its characteristic parameters. The characteristics of the top ten of importance and the absolute values of their parameters are as follows. It is obvious that except for area and floor, all the other classification indicators are regional classification indicators. The top ten classification indicators of area and area are also in line with the phenomenon that the cost-effective index designed by ourselves can well verify the conclusion in the hypothesis test, but unexpectedly, the number of floors is also one of the important characteristics. Indeed, its influence should not be ignored in feature engineering.