Clojure 数据科学与数值计算：tablecloth / tech.ml 实战

(require '[tablecloth.api :as tc])

;; 从 CSV 加载成数据框（ds = dataset）
(def df
  (tc/dataset "orders.csv"
              {:parser-fn {:order_id :integer
                           :amount    :double
                           :paid_at   :datetime}}))

(tc/head df)  ; 查看前几行
;; => #tech.ml.dataset/table [...]（类似 pandas 的 df.head()）

(def clean-df
  (-> df
      ;; 按列处理：缺失值填充、类型转换
      (tc/replace-missing :amount {:value 0})
      (tc/convert-types :amount :double)
      ;; 筛选：只保留近一年的有效订单
      (tc/select-rows (fn [row] (and (:paid_at row)
                                     (> (:amount row) 0))))
      ;; 新增列：月与金额档位
      (tc/add-column :month #(month (:paid_at %)))
      (tc/add-column :amount-tier
                     #(cond (< (:amount %) 100) :low
                            (< (:amount %) 500) :mid
                            :else :high))
      ;; 排序 + 去重
      (tc/order-by [:paid_at :desc])
      (tc/distinct-by :order_id)))

;; 按月份 + 金额档位做汇总：group-by + aggregate
(def monthly
  (-> clean-df
      (tc/group-by [:month :amount-tier])
      (tc/aggregate {:order-count #(count %)
                     :total-amount #(reduce + (map :amount %))})))

;; 与 pandas 的 groupby().agg() 完全对标

(require '[tech.v3.datatype.functions :as fns])
(require '[tech.ml.dataset.column :as col])

;; 对一列做描述性统计
(def amounts (col/->array (:amount clean-df)))

{:mean (fns/mean amounts)
 :std  (fns/std amounts)
 :min  (fns/min amounts)
 :p90  (fns/quantile amounts 0.9)}

;; 用 tech.ml.dataset 计算两列相关性
(defn corr [ds a b]
  (let [xs (col/->array (a ds))
        ys (col/->array (b ds))]
    (fns/correlation xs ys)))

(corr clean-df :amount :order-count)

(require '[hanami.core :as h])
(require '[tablecloth.api :as tc])

;; 柱状图：每月订单量
(def bar-spec
  (h/xform :bar-chart
           {:data  (-> monthly (tc/rows))
            :x     :month
            :y     :order-count}))

;; hanami 生成 Vega-Lite 规范（JSON），可在浏览器渲染
(println bar-spec)

(require '[tech.ml :as ml])

;; 把 dataset 拆成特征与标签
(def split-data
  (ml/feature-dataset->row-major-feature-dataset
    clean-df
    {:feature-cols [:amount :order-count :month]
     :target-cols  [:amount-tier]}))

;; 训练 / 测试集划分
(def [train test] (ml/split-dataset split-data {:seed 42 :split-percent 0.8}))

;; 训练一个随机森林分类器
(def trained
  (ml/train {:model-type :classification/random-forest
             :dataset    train}))

;; 在测试集上评估
(def report
  (ml/evaluate {:trained-model trained
                :dataset       test
                :metric-map    {:accuracy :classification/accuracy
                                :f1       :classification/f1}}))

(report :accuracy)
;; => 0.87

;; 保存 / 加载模型，用于服务端推理
(ml/save-model trained "models/order-tier.model")
(def reloaded (ml/load-model "models/order-tier.model"))

(require '[libpython-clj2.require :refer [require-python]])
(require '[libpython-clj2.python :as py])

(require-python '[numpy :as np])
(require-python '[sklearn.linear_model :as lm])

(def X (py/py-tuple->list (np/array (map #(vector (:amount %) (:order-count %)) (tc/rows clean-df)))))
(def model (lm/LinearRegression.))
(py/call-attr model :fit X (py/py-tuple->list (map :month (tc/rows clean-df))))